New Computation and Language (cs.CL) papers, Tuesday 20 October 2026
285 papers announced on arXiv. Get them ranked for you, free →
Transfer Learning for Power Outage Detection Task with Limited Training Data
Early detection of power outages is crucial for maintaining a reliable power distribution system. This research investigates the use of transfer learning and language models in detecting outages with limited labeled data. By leveraging pretraining and transfer learning, models…
Multi-LLM Systems Exhibit Robust Semantic Collapse
AI summaryMulti‑LLM systems converge to the same semantic representation over many generations without retraining, a phenomenon called semantic collapse. Despite varied decoding, prompting, agent design, and RL interventions, the collapse persists across model families and long runs.
Transformer-based segmentation of prosodic boundaries in Brazilian Portuguese
AI summarySAMPA fine‑tunes Whisper on BP speech to insert prosodic boundary markers, achieving F1 = 0.731 on in‑distribution data and 0.796 out‑of‑distribution. The model leverages morphosyntactic, semantic, and acoustic cues for boundary detection.
I$^2$-Med: Interpretable Medical Inference through Visual-Guided Dynamic Logits Calibration
AI summaryI$^2$-Med adds a visual‑guidance module that re‑scores generated tokens during inference, favoring words that are visually grounded and contextually consistent, without retraining the model. On OmniMedVQA, it yields clinically relevant, visually faithful VQA outputs and cuts hallucinations.
Modeling Human Behavior in the Prisoner's Dilemma with Large Language Models
AI summaryThe study tests open‑source LLMs in repeated Prisoner’s Dilemma, comparing fixed‑partner and shuffled‑partner settings to human data. All models over‑cooperate, with Qwen‑2.5‑7B hitting >80% in shuffled trials versus humans’ 28.5%, showing limited realism.
GardenBench: A lightweight, daily evaluation of LLM capabilities
AI summaryGardenBench offers a lightweight, daily LLM benchmark with 11 test suites that track daily accuracy, cost, and latency for consumer-facing tasks, enabling public comparison of model versions and providers through a live dashboard.
Composable Parametric Demographic Injection for Values Survey Simulation
AI summaryPARADE injects demographic shifts into LLMs via reusable LoRA modules for each attribute, separating task knowledge from demographic effects. In simulations of World Values Survey and SocioBench, it reduces population KL divergence by up to 52.8% while preserving predictive accuracy.
Translating in style: An open-weight alternative to GPT-4 for culturally aligned machine translation
AI summaryOpen-weight Hunyuan-MT-7B variants (base, LoRA, GRPO) are evaluated for stylistic alignment in 12 language pairs across formality, politeness, intimacy. Fine‑tuning offers no significant improvement; overall performance lags GPT‑4 mainly on formality but matches or exceeds it for intimacy and Chinese.
Latent Multimodal Memory Reasoning
AI summaryThe paper introduces Latent Multimodal Memory Reasoning (LMMR), a neuro-symbolic framework that maps heterogeneous visual and textual memories into a shared sparse concept space to construct a dynamic memory graph. Experiments show that LMMR significantly outperforms existing methods in long-term multimodal reasoning and response grounding.
TeleTwin: A Calibrated Digital Twin of Telegram Crisis Discourse as a Ground-Truth Testbed for Detecting LLM-Powered Information Operations
Large language models make it cheap to run influence operations whose messages never repeat, yet real platforms offer almost no ground truth for testing whether coordination detectors can find them. The problem is sharpest on Telegram, which moderates little and spreads content…
How Data Scales in Agentic Reinforcement Learning: Laws and Synthesis Strategies
AI summaryThe paper decomposes agentic RL data into task, environment, and trajectory axes, finds scaling exponents ordering environment > task > trajectory, and uses these laws to optimize budget allocation, achieving up to 1.7× compute efficiency over naïve scaling.
UltraData-Code: From Raw Repositories to Algorithmically Dense and Task-Oriented Code Data
AI summaryUltraData‑Code introduces a tiered data pipeline (L0‑L3) that filters, normalizes, selects algorithmically relevant code, and turns it into structured exercises, yielding ~400B L2 tokens and 150B L3 tokens. Using L2 improves EvalPlus pass@1 by 7.8pp, and adding L3 boosts it by another ~8pp.
LLM Evolutionary Search Strategies Should be Compared by Tokens
AI summaryLLM evolutionary search is usually benchmarked by proposal counts, but this obscures real compute cost; using token counts shows that evolution’s apparent advantage disappears—evolution only outperforms independent sampling by ~5% in most cases, with a 23–28% margin on only one task.
LLM Evolution in a Simple Simulated Market
Arguments that AI systems will evolve in harmful ways are largely made in the abstract. We provide a concrete empirical example. We simulate simple markets consisting of populations of twenty Large Language Model agents tasked with solving arithmetic problems. Agents differ only…
The Window Is in the Reasoning: When LLM Agents Decide to Follow Prompt Injections
AI summaryThe paper shows that prompt‑injection decisions in LLM agents are largely determined when the injection is read, but reasoning reduces this “read‑time” variance. It finds that multi‑run evaluations are necessary, as single‑run labels can disagree up to 35 %.
Discuss More, Comment Less: Boosting Automated Research via Multi-Perspective Collaborative Review
As automated research is increasingly adopted for writing papers, automated review has become a critical link between paper generation and quality control. However, existing methods largely rely on single-model feedback or simple review aggregation, failing to explicitly model…
LogicVal:A Logic-Formalized Benchmark for Constraint Satisfaction in Text-to-Image Generation
AI summaryLogicVal formalizes text‑to‑image evaluation as a constraint‑satisfaction problem, combining neural detection with logical reasoning. It shows a large drop from 98% success on single-object constraints to 53% on multi‑object logical relations, highlighting T2I models’ limited compositional reasoning.
World of SWECraft: Agent Collaboration at Scale via Simulated Software Worlds
AI summaryThe paper introduces Software World, a simulation framework that creates realistic software ecosystems for multi‑agent collaboration. Experiments show that scaling inference‑time compute yields up to 31% efficiency gains, with most agents finding useful tasks while some high‑performance models stop searching early.
Location, Location, Generation: Where You Ask Matters for Generative Engine Optimization
AI summaryGenerative engines tailor cited sources to the user’s declared location rather than network IP. ChatGPT localizes results to the city it is told, while Gemini uses the network country, altering recommended businesses and sources accordingly.
Coopshift: Post-Training Amplifies The Memory Curse In LLM Cooperation
How does post-training change cooperation during repeated interaction? Within each of seven released model lineages, we compare pretrained (Base), supervised fine-tuned (SFT), and subsequent preference- or reward-optimized (Aligned) checkpoints across seven social dilemmas.…
OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners
AI summaryLarge language model agents can solve real‑world operations research problems directly from client materials, but they rarely match expert solutions. In a benchmark of 31 industrial cases, the best LLM achieved 88% feasibility and 0.87 DAQ, below expert level.
Cache2LoRA: Understanding and Improving Context Internalization
AI summaryCache2LoRA generates low-rank parameter updates from a reusable document cache guided by the query. It matches full-context in-context learning performance across 17 tasks while delivering nearly six times the decoding throughput.
Joint, Decoupled, or Agent: When Each Multimodal Alignment Paradigm Wins
AI summaryThe paper compares joint, decoupled, and agent multimodal alignment paradigms across 16 datasets, finding no single winner but that selection depends on supervision budget, task format, and unimodal 5‑NN accuracy, with a diagnostic cascade choosing the best paradigm 71.4% of the time.
Do LMs Know What Variables Do? Interpreting Code Representations via Variable Roles
AI summaryThis paper investigates whether language models and sparse autoencoders capture programming variable roles, introducing a multilingual corpus with static annotations. The results show that models recover these roles successfully, but causal interventions often degrade code correctness.
Prefix-Tuning for Arbitrary Output Sequences on Pretrained Transformers
AI summaryThe paper shows that for any pretrained autoregressive Transformer and any target output within the model’s range, a prefix can be constructed by solving linear systems to force the model to generate that output; it also proves this condition is necessary.
Inside the Loop: A Mechanistic Study of Weight-Tied Transformers on Depth-Bound Algorithmic Tasks
AI summaryThis paper investigates how weight-tied looped transformers solve multi-step algorithmic tasks by reusing layers across iterations. It finds that looped models act as iterative program learners that outperform identically-trained untied transformers of the same effective depth.
Continual Test-Time Learning with Evolution Strategies
AI summaryThis paper introduces Evolution Strategy Reinforcement Learning, a framework that uses a population of perturbed policies to generate stable, label-free supervision for continual test-time adaptation in language models. The method improves average accuracy by up to 31.7% and significantly reduces catastrophic forgetting during continual learning across diverse domains.
Preservation and loss of algorithmic structure in experimental diffusion of joint-action norms
AI summaryThe study tracks how new social norms diffuse across populations, finding that early diffusion favors simpler algorithms, causing complexity loss. However, interactive repair mechanisms preserve complex procedures despite misalignment, showing that collaborative reasoning buffers structural erosion in cultural evolution.
CoDA: A Human-LM Collaborative Codebook Development Assistant for Domain-Specific Text Annotation
AI summaryCoDA is a human‑LM collaborative tool that speeds domain‑specific codebook creation by iteratively refining guidelines through feedback on model errors. It matched expert codebooks’ accuracy while cutting development time to 2 hours and annotations by fivefold.
Pronouncing Hard Words: Phoneme Adaptation of a Pretrained Conversational Text to Speech Model
AI summaryThis paper adds phoneme controllability to a pretrained conversational text-to-speech model using vocabulary extension, fine-tuning, and inference-time biasing. The method raises out-of-domain pronunciation accuracy for rare drug names from 54.6 percent to 90.7 percent.
Prosodic Structure as a Critical Component in Phonemic Language Modelling at BabyLM Scale
AI summaryThe paper adds prosodic annotations to phoneme‑level BabyLM training and shows that a prosodically structured phoneme stream yields the best semantic taxonomic organization at 10M and 100M word scales, outperforming orthographic baselines and matching performance with phonemic models that include boundary cues.
Multi-author style detection with stylistic embeddings
AI summaryThe paper fine‑tunes Mistral‑7B embeddings for authorship verification, shows it outperforms DeBERTa on certain user‑generated text types, and combines both embeddings to improve multi‑author style change detection.
Reasoning with Context Management and External Memory
AI summaryRCMEM adds an external memory to in‑context reasoning, allowing models to retain useful intermediate facts without enlarging the context window. It cuts computation time and improves accuracy on complex math tasks across 0.4–25M‑parameter models and GPT‑2.
Language Models are Lazy Poets: Creative Mode Collapse at Scale
While aligned large language models (LLMs) can successfully complete many creative tasks, their outputs tend to be highly self-similar. Though solutions to this problem of "mode collapse" have been proposed, comprehensive large-scale surveys of the phenomenon in specific textual…
Analyzing Repetition and Interference Effects in Memory-Constrained User Simulators
AI summaryUser simulators are evaluated for behavioral parity with humans by testing memory constraints, repetition, and interference effects. While some simulators match overall human accuracy and replicate repetition/interference magnitudes, they differ on specific errors, revealing distinct forgetting patterns.
Preserving Reasoning Diversity in Reinforcement Learning with Verifiable Rewards via Adaptive Replay
AI summaryThe paper analyzes how limited on‑policy rollouts in RLVR suppress semantically distinct reasoning modes and introduces TAR, a Targeted Adaptive Replay that stores diverse base-model solutions and replays them only when their usage drops, improving both performance and diversity.
Encoder-Free Multimodal LLMs Learn Implicit Encoders
AI summaryEncoder‑free multimodal LLMs (Gemma‑4 12B, Inkling‑Small) learn “implicit encoders” in their early layers that transform audio and visual inputs into representations largely independent of cross‑modal context. Layer probing shows these implicit encoders match the behavior of standard dedicated encoders, and early‑modality encoding alone matches joint‑encoding accuracy.
Probing Ambiguity: Adversarial Intent Resolution for Robust Text-to-SQL
AI summaryAIR probes ambiguous text‑to‑SQL queries by generating multiple SQL interpretations, executing them, and asking data‑driven clarification questions; it outperforms prior methods on AMBROSIA and AmbiQT, achieving higher F1 with only ~7% loss without full database access.
Can This LLM Annotate My Data? Estimating Zero-Shot LLM Performance
AI summaryThe paper constructs 461 synthetic text classification datasets and shows that zero‑shot LLM accuracy can be predicted from dataset features; internal log‑probabilities outperform confidence verbalizations for predicting performance.
SLA-LENS: From Classroom Transcripts to CSLA-Informed and Evidence-Linked L2 Interaction Analytics
AI summarySLA‑LENS automatically annotates classroom transcripts with SLA‑based tags for learner errors, pedagogical functions, and uptake, using LLMs and QLoRA adaptation. It produces reproducible, evidence‑linked analytics on 21k+ Spanish L2 interactions, aiding research and curriculum design.
Best of Both Worlds: Fine-Tuned Small Models Complement LLMs for Cost and Performance Benefits
AI summaryFine‑tuned RoBERTa outperforms GPT‑3.5 for 16‑class survey classification (F1 96% vs 83%) and is 10× faster. An ensemble of RoBERTa and GPT‑3.5 reaches F1 97.5% at only 4.4% of GPT‑3.5’s inference cost.
Desirability Confirmation: Relational Sycophancy in AI Intimate Roleplay
Research on LLM sycophancy has largely focused on propositional agreement, in which models echo users' beliefs, preferences, or judgments. Adult intimate roleplay presents a different case because what the user seeks may be preservation of a preferred relational position within…
Modeling bilingual naming performance through laterally-connected SOMs
AI summaryBiLex, a laterally‑connected self‑organizing map model of bilingual lexical access, reproduces Spanish‑English naming accuracy, response times, and semantic‑based error patterns from the Boston Naming Test, showing it captures individual semantic‑cognition dynamics.
BanglaShield: A Script-Robust Safety Classifier and Benchmark for Bangla Native, Romanized, and Code-Mixed Prompts
AI summaryBanglaShield provides a graded safety classifier for Bangla prompts in native, romanized, and code‑mixed forms, achieving 87.8 % macro‑F1 on native script and 84.4 % on romanized, outpacing large LLMs by over 40 points while using ~130× fewer parameters.
Beyond Clean Accuracy: The Practical Challenges and Benefits of Pixel-Based Language Models
AI summaryPixel-based language models learn from rendered text, improving robustness to typographical noise and cross‑lingual transfer. However, they need longer fine‑tuning, more labeled data, and careful learning‑rate scheduling, making adaptation less reliable than token‑based models.
Not Every User Says It All: Training Tool-Using Agents Beyond Ideal Simulated Users
AI summaryTraining with weak, masked users hurts performance; existing policies trained on strong users do not improve for weak users. StaGPO, a state‑grouped policy optimizer, lets agents document missing info and assign step‑level credit, boosting weak‑user success by 4.8 points on ToolWOZ and improving generalization.
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
AI summaryThis paper introduces prompt-based strategies where large language models explicitly recognize and critique their own systematic biases during evaluation. These lightweight methods consistently improve evaluation reliability across diverse bias types without requiring external training.
What Matters for NVFP4 Training? A Comparative Study of Low-Precision Pretraining Recipes
AI summaryNVFP4 training of 8B and 30B MoE language models to 1T tokens is studied, showing a recipe that combines unbiased gradient estimation, selective high‑precision layers, and partial tensor de‑biasing. The final method achieves the smallest loss gap to BF16 at this scale.
FlashMap: Flow Map Speculative Decoding
AI summaryFlashMap uses a flow‑based drafter conditioned on target hidden states to generate token blocks in one pass while modeling inter‑token dependencies. It outperforms DFlash on Qwen3‑8B, achieving 7% higher average speedup and up to 6.1× lossless acceleration.
Defending Against Dormant Poisoning Attacks Across Language and Multimodal Agents
Recent work shows that attackers can implant dormant malicious behaviors that remain hidden at release but emerge after benign fine-tuning. As foundation models are used as agents, whether dormant poisoning extends to harmful agent actions and multimodal settings remains…
Long Listening Thoughts: Eliciting Open Auditory Reasoning with Deliberative Perception and Cognitive Refinement
We introduce Long Listening Thoughts (LLT), a framework for constructing post-training datasets that improves audio reasoning by pairing deliberate perceptual grounding with explicit cognitive behaviors. On audio understanding and reasoning benchmarks, large audio-language…
Ask Less, Ask Better: Allocating User Attention in Proactive LLM Agents with Interruption Budgets
AI summaryThis paper introduces a user‑attention cost metric that penalizes difficult or unnecessary clarifying questions, and shows that limiting proactive LLM agents to a small interruption budget (1–3 questions) cuts cost by 36‑73% while retaining 86‑94% of task reward.
Toward a Universal Local Speech Feature Extractor Through Distillation
AI summaryA two‑layer universal feature extractor is distilled from large SSL teachers; it matches teacher performance on SUPERB tasks while using only 3% of the CNN MACs, reducing overall inference time by over 20%.
When Opposite Premises Translate Identically: A Source-Conditioned Diagnostic for Multilingual Commonsense Items
AI summaryThe paper introduces a source‑conditioned diagnostic that flips the polarity of English premises, translates both sides, and counts identical target outputs to detect unfaithful translations in multilingual commonsense QA. The method finds dozens of collisions in Swahili and LaBSE, proving translation errors but not pinpointing the faulty arm.
Which Long Documents Matter? A Controlled Study of Data Selection for Context Extension
Natural long documents are widely used for context extension, but which selection pipelines are effective remains unclear. We conduct a controlled comparison on Qwen2.5-3B at 128K under matched 6B-token continued-training budgets. Random sampling of natural long documents…
Pluralistic Alignment in Language Models: Objectives, Human References, and Preference Representations
AI summaryThe paper proposes a two‑layer taxonomy for pluralistic alignment—categorizing objectives (Overton, Steerable, Distributional) and human references (aggregate, regional, subgroup)—and reviews 70 studies, showing most work focuses on Steerable and Distributional pluralism with limited evidence for true preference variation.
Ben-Misandry: A Bengali Dataset and Late-Fusion Transformer Framework for Detecting Anti-Male Gender Hate
AI summaryThis paper introduces Ben-Misandry, the first Bengali misandry detection dataset, containing 5,000 synthetic comments across six taxonomy classes. A late-fusion transformer ensemble combining BanglaBERT and XLM-RoBERTa achieves 96.6% accuracy on the benchmark.
Beyond Recall Saturation: Scaling Hard-Negative Reranker Training for Industrial Legal Retrieval
AI summaryThe paper shows that after a large candidate pool, improving reranker quality yields the biggest gains. An embedding‑to‑self‑mined hard‑negative training pipeline scales to 4B parameters and outperforms the baseline on legal benchmarks, preferred in 72.4% of industrial queries.
Write Once, Check Many: Reusable Ledgers for Post-Hoc Policy Compliance Auditing of Agent Trajectories
AI summaryThe paper introduces a write‑once, check‑many ledger system that records agent trajectories in a policy‑independent format, enabling efficient post‑hoc compliance checks. Experiments show ledger readers outperform trajectory readers, achieving higher AUC with far fewer tokens.
Modeling Discourse Profiles in Climate Misinformation Discourse
AI summaryThis paper builds discourse profiles by clustering TikTok climate comments into recurring rhetorical patterns and uses them in late‑fusion classifiers, boosting macro‑F1 over text‑only baselines for several representations. It also shows that discourse profiles influence misinformation reception and spread.
Vocal-Tract Acoustics, Aerodynamics, and Overlap-and-Add Synthesis in DYNARTmo
This paper presents the acoustic and aerodynamic extension of DYNARTmo, a dynamic articulatory model for visualizing and synthesizing speech production. Gestural control trajectories determine vocal-tract geometry, which is converted into an acoustic tube with nasal branching…
Backend-Frontend Converter: Component-Based Conversion of .NET Computational Models to TypeScript
AI summaryThis paper introduces an architecture that converts component-based .NET computational models into TypeScript for browser execution. The evaluation demonstrates that the generated TypeScript implementations achieve near-identical numerical precision compared to the original C models across orbital and trading scenarios.
ProbePO: Sparse Probes, Redistributed Credit for Critic-Free LLM Reinforcement Learning
Reinforcement learning from verifiable rewards assigns one terminal reward to an entire reasoning trajectory, obscuring how individual reasoning steps contributed to the outcome. GRPO broadcasts a group-normalized advantage to every response token, while finer-grained approaches…
Axiomatic World Modeling for Physics Reasoning
AI summaryRLAWM trains LLMs to penalize differences between their internal axiomatic world model and the true physics, using two hierarchical rewards for modeling and solving. The method boosts PhysReason by 12.2 points and PHYSICS by 5.2, with strong zero‑shot generalization.
EGIP: An Evidence-Grounded Interaction Protocol for Clinical Agents
Interactive clinical agents acquire evidence through patient dialogue and diagnostic examinations, yet their performance also depends on how the interaction environment exposes information, available actions, and examination results. When these conditions are implicit or…
From Stories to Features: Scaling Concept Vectors for AI Alignment & Interpretability
AI summaryThe authors propose generating feature vectors for language model safety and alignment by taking the mean activation difference between synthetic concept-antagonist story pairs. Scaling this approach to 1,036 concepts across multiple models, they show these vectors generalize to detect concepts, reveal shared geometric properties, and causally steer model behavior.
ASR Personalization in Pathological Voice: Accuracy, Efficiency, and Duration Effects
Automatic Speech recognition (ASR) struggles with Pathological speech due to persistent speaker-specific mismatch. In clinical and accessibility scenarios, only limited target-speaker data is typically available, making full model retraining impractical. We present a…
Training with Ciphers to Enhance In-Context Terminology Learning For Translation from Rare Languages
AI summaryThis paper introduces RareCipherFT, a fine-tuning method that replaces rare translation terms with ciphers during training to force LLMs to copy from context. The approach improves in-context terminology translation by 5 to 15 points across multiple rare language pairs without degrading overall translation quality.
Safety Kernel: Mechanistic Alignment for Cross-Lingual Defense in LLMs
The safety capabilities of large language models (LLMs) are highly uneven across different languages, often allowing cross-lingual attacks to bypass safety guardrails. Existing alignment methods predominantly rely on external alignment signals and provide limited insight into…
Muse Glimmer 30B: An On-Device Agentic Model
We introduce Muse Glimmer, an open-source 30B parameter causal language model with perception encoder, purpose-built for autonomous agentic tasks on consumer hardware. Muse Glimmer integrates multi-step reasoning, reliable tool use and failure recovery into a single model that…
What RL Brings to Reasoning: A Few Decisions the Base Model Would Not Make
AI summaryThe authors investigate what reinforcement learning adds to reasoning base models by swapping tokens between them during generation. They find that borrowing the RL model's choices at just four percent of token positions recovers over ninety percent of its reasoning performance gains.
Do Conversational Agents Need Explicit Memory? A Controlled Comparison with Raw-History Retrieval
Explicit memory systems transform conversation history into persistent representations intended to improve future access, often at substantial construction cost. Yet it remains unclear whether this added complexity yields better answer accuracy than retrieving directly from raw…
Verifying Human-Written Reasoning References
AI summaryThe paper introduces a method to verify each step of human‑written reasoning chains by checking entailment against cited premises, re‑checking against full premises, and attributing failures. On P‑FOLIO, only 52.3% of 386 derivations passed all checks; most “failures” were due to missing or mismatched premises, with 21 human errors, 11 translation errors, and 7 annotation errors identified.
When Do Repetition Loops Lock? Truncation Sampling, Bounded Penalties, and a Blind Spot in DRY
Language models sometimes repeat verbatim until the generation limit. We study the exit rather than the entrance: given that a span has already been repeated $n$ times, when is escape still possible? Along the loop the context is deterministic, so loop survival under any sampler…
DiffuPlan: Editable Planning State for Diffusion Language Model Agents
Diffusion language models (dLLMs) are emerging as an alternative to autoregressive (AR) language models, yet dLLM agents largely inherit an interaction protocol designed for AR generation: once intermediate state is written, later observations can only be incorporated by…
Neuro-Inspired Inhibitory Message Passing for Accurate and Token-Efficient Multi-Agent Reasoning
AI summaryInhibitory message passing routes each evidence sentence only to agents that lack it, suppressing redundant information without training. Across seven models and multiple multi‑agent architectures, the method boosts accuracy by 39–77 points and reduces tokens per run by 10–46 %.
PV-Decoding: Progressive Verification Decoding with Decode Gating for Diffusion Language Models
AI summaryThis paper introduces PV-Decoding, a speculative decoding framework for diffusion language models that uses confidence-driven partitioning and progressive verification. It reduces sequential forward passes by up to 37% while largely preserving generation accuracy.
GLOSS-MT: Controllable Low-Resource Machine Translation with an Editable Dictionary
AI summaryGLOSS-MT injects an editable offline dictionary into pretrained machine translation models via a single round of fine-tuning on mixed-language inputs. This approach improves translation quality on low-resource English-to-Asante Twi benchmarks by up to 5.48 chrF2 while enabling precise, retraining-free terminology control.
BanglaBBQ: A Bias Benchmark for Question Answering in the Bangladeshi Context
AI summaryBanglaBBQ adapts the BBQ bias benchmark to the Bangladeshi context, adding 140 templates and 372 new stereotypes in Bangla and English. Models tested (Qwen3‑4B/8B, GPT‑4.1‑mini, Gemini 2.5 Flash) all show bias toward stereotyped groups, with closed models following counter‑stereotypes only on new templates.
ECHO: Efficient Causal Self-Recursion Beyond Auxiliary Multi-Token Prediction
AI summaryECHO replaces auxiliary multi-token predictors with a single backbone that predicts all horizons via self-recursion. Dynamic Depth Sampling trains deeper horizons efficiently, improving downstream accuracy and open-ended generation at roughly half the cost of traditional MTP methods.
Information Theory of Random Composition Trees
Random composition trees model the hierarchical semantic structures of meaningful texts as a statistical ensemble generated by recursively splitting texts into coherent chunks. A recent study shows that different text types can be described by random composition trees with…
Who Gets Unmasked? Sampler-Induced Bias in Diffusion Language Models
AI summaryDiffusion language models exhibit a sampler-induced bias where confidence-based token commitment rules amplify stereotypes and distort distributions. The authors propose Fork-First Sampling to resolve this bias by deciding critical choices before text generation.
Table Serialization Determines Which Relations Language Models Preserve
AI summaryExamines how row‑major vs column‑major text serialization of tables affects instruction‑tuned models’ lookup accuracy. Finds that serialization format, not table shape alone, determines which column–row relation is preserved, with row‑major favoring rows and column‑major favoring columns.
Medical Fine-Tuning Reshapes, Not Reduces, Patient-Facing AI Risk: A Rubric-Based, LLM-Judge Evaluation Framework
AI summaryThe paper proposes a risk‑aware evaluation rubric scoring medical LLM answers for six harmful language categories, using a three‑judge LLM ensemble. It finds medical fine‑tuning does not lower overall risk but reshapes it, reducing high‑alert mentions while increasing overconfident statements.
Ask, Don't Guess: Targeted Clarification for Speech Recognition Errors
AI summaryThe paper introduces a framework to identify speech recognition errors in real time and prompt users to repeat only the affected spans. This targeted repetition improves word error rates by up to 24 percent while minimizing unnecessary interruptions.
Efficiency Measured in English: Tokeniser Coverage Bounds Small Language Models on Indic Scripts
Small Language Models with quantisation and parameter-efficient adaptation are widely offered as the route to multilingual AI under compute constraints, and their efficiency is almost always reported as a single memory figure. We show that this figure is measured in English and…
MisleadQA: Diagnosing LLM Failures under Misleading Context
Large language models (LLMs) are widely used to answer questions from retrieved or user-provided context. However, provided context can be factually true yet out-of-scope, describing a different entity, time, or location from the one the question targets, and still lead models…
Learning to Program Retrievals: Composing Search Primitives via Reinforcement Learning
AI summaryRetrieval Program Learning lets a policy compose search primitives into programs with state and control flow, and program mutation generates local variants for better exploration. On QA and corpus reasoning, it outperforms larger baselines, especially on multi‑hop tasks, while reducing context and retrieval steps.
Dual Task Vector Merging: Leveraging Emotion and ASR Knowledge for Low-Resource Alzheimer's Disease Detection
Alzheimer's Disease (AD) detection from speech benefits from both paralinguistic and linguistic cues, yet model training is frequently constrained by limited data. Fine-tuning a large pretrained speech encoder can easily overfit, while a frozen probe leaves the encoder's…
Bellman Language Models: Solving the Recurring Decisions of LLM Agents
AI summaryBellman Language Models (BLMs) discretize an LLM agent’s conversation into states, estimate transition probabilities and costs, solve the resulting MDP once, and reuse the optimal policy, reducing model calls by up to 98% and raising task success from 29% to 96% on telecom benchmarks.
The Missing Compliance Class: LLM Graders That Collapse Jailbreak Success, and a Free Check
Jailbreak papers report an attack success rate, the share of responses that deliver the harmful request, and read it as a property of the target model. We hold the target model, the prompts and the responses fixed and change only the grader. Three dense graders of 8B and 9B…
The One-Pixel of Reasoning: One Truthful Relation Is Enough to Mislead LLM Agents
AI summaryThe paper introduces the one-relation attack (ORA), which flips LLM beliefs by inserting a single, source‑grounded sentence targeting an implicit inferential warrant. ORA succeeds on 54.9% of proprietary and 80% of open‑weight biomedical models, transfers without feedback, and evades detection.
Measuring Data Diversity with Embeddings: A Taxonomy and Benchmark for Diversity Measures
Diversity is a critical property of text corpora, with direct implications for training data curation, test set design, and model generalization. However, its measurement remains fragmented and poorly understood. Existing embedding-based diversity measures vary widely in how…
LastWords: Boring Problems Revisited
AI summaryThe paper shows that modern frontier models (e.g., GPT‑4) have practically solved the number name reading subproblem of text normalization, but it argues that smaller, resource‑efficient methods could still address remaining edge cases without requiring massive models.
Peer Pressure: Execution-Verified Co-Evolution of Attacker and Defender Populations for Code Security
LLM code agents increasingly ship code with little oversight, yet defenders trained on static vulnerability datasets fall behind adapting adversaries. We co-train attacker and defender instead, and let execution decide every reward: an attacker (Red) edits a real repository and…
HybridResearcher: Bridging the Structured-Unstructured Boundary for Deep Research Agents
AI summaryHybridResearcher is a 4B parameter agent trained via synthesized trajectories and reinforcement learning to reason across structured databases and unstructured web sources. It outperforms much larger models on hybrid research benchmarks without sacrificing single-source performance.
Compositional Activation Steering Across Languages and Behaviors with Contextual Gates
AI summaryContextual gates learn to adjust fixed steering vectors per hidden state, enabling simultaneous control of multiple behaviours in multilingual language models. The method outperforms constant-strength steering and prompting while preserving knowledge and keeping internal computation closer to the prompted model.
A Cross-Domain Evaluation of Prompting Techniques for Large Language Models
AI summaryThe study benchmarks Zero-Shot, Few-Shot, Retrieval-Augmented Generation, and automatic prompt optimization across six reasoning tasks using a single LLM. RAG tops the overall average (78.33 %) but no method outperforms all others; selection should depend on task characteristics.
Arctic-SWE: Revisiting the Long Horizon Agent Training by Modeling-System Co-Design
AI summaryArctic-SWE introduces an asynchronous reinforcement learning framework for long-horizon software engineering agents through model-system co-design. Using optimized infrastructure and curated data, it enables a 4B model to achieve 65.0% on SWE-bench Verified.
How Well Can LLMs Verbalize Mathematical Formulas? A Multi-Domain Evaluation
AI summaryLarge‑language models can generate reasonably accurate verbalizations of LaTeX formulas, with GPT‑5.2 scoring highest on simple transcriptions (≈76 % BLEU) and Claude Sonnet 4.5 excelling on extended transcriptions (≈36 % BLEU). Accuracy drops for longer, more complex formulas and depends on context and domain.
CARTA: Self-Supervised Reusable Transaction Embeddings via Heterogeneous Tokenization and Cross-Attention
AI summaryCARTA learns a 16‑dim, label‑free embedding of payment transactions using heterogeneous tokenization and InfoNCE. The frozen embedding boosts fraud detection, outperforming raw features at 1% labels and beating frequency encoding on large tabular data, while remaining reusable for downstream tasks.
Stable Signer V3: 4D Deformable Primitive Splatting Is All You Need for Sign Language Production
AI summaryStable Signer V3 uses 4D Deformable Primitive Splatting with face, body, and tunneling super‑resolution modules optimized once per signer, reducing facial error 71.4 %, body error 35.1 %, and RGB error 24.2 % while enabling interactive high‑resolution sign language avatars.
What Can't Jev Decide? Probing the Limits of One-Pass Decision Models
AI summaryJev makes one-pass decisions efficiently but cannot match LLMs that write intermediate steps; its accuracy drops when multiple updates are needed and it fails to detect errors from stale state. It succeeds on 97.5% of one-pass solvable items but only 74% of more complex ones.
Epi: Towards a Persistent, Self-Improving Agent for Autonomous Research
AI summaryÉpi is a self‑improving research agent that persists across model sessions, learns from experiments, and produces 120 record‑setting results—including an 84‑fold tighter $n$‑queens bound and a 3.6‑million‑fold Feigenbaum interval—while also proving new combinatorial conjectures. Its lightweight Épi‑turbo variant outperforms Codex on 30 routing/scheduling benchmarks with only 64 % model cost.
Over-Defense in Coding Agents Begins at Silent Forks
AI summaryCoded language models over‑defend by taking unnecessary protective actions at silent forks—points where any action yields the same outcome. ForkDPO uses these forks as a learning signal to reduce such disproportionate behavior, cutting it from 70.3% to 58.9% on a 4B model.
CasePlay: Self-Play Reinforcement Learning from Case Reports for Medical Reasoning
AI summaryCasePlay uses document-grounded self‑play on medical case reports, generating questions from knowledge points and evaluating answers with rubric‑based quality feedback. It improves LLM medical reasoning by 3.1 overall points (6.3 on LiveClin) over prior self‑play baselines.
Revealing what benchmark scores hide: Jointly measuring the breadth and reliability of LLM capabilities using accessibility profiles
AI summaryThe paper defines “accessibility” as the probability a model succeeds across equivalent formulations of a question. It shows that many model pairs have similar average accuracy but diverge on Access@5 vs. Reliable@5, revealing a tradeoff between breadth and consistency of performance.
CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents
AI summaryCoworkEnv synthesizes training environments guided by workflow constraints, generating requests, tools, and validation checks for AI coworkers. Training with these environments raises Qwen3‑VL‑8B‑Instruct SFT scores from 46.6 to 64.1 on Claw‑Eval, especially improving multi‑turn performance.
DeGMatch: Target-Aware Reasoning Data Selection via Deconfounded Gradient Matching
AI summaryDeGMatch selects reasoning data by deconfounding candidate gradients: it gates for target relevance, centers gradients on the population mean, and removes dominant principal components before matching. This improves target discrimination and outperforms baseline methods across math, science, and code tasks.
Rubrics Carry the Standard: Shapley Attribution for Checklist-Based LLM Judges
AI summaryChecklist rubrics dominate LLM judging, but changing the rubric content can change pass rates from 2% to 100% while altering the judge model has little effect. The paper introduces a Shapley-based attribution that detects rubric drift from self-review and proposes a change-control gate to maintain quality.
A Measurement Perspective on Benchmarking LLM-Generated Text Detectors using Item Response Theory
AI summaryThe paper introduces an item-response-theory framework that models benchmark texts as items and detectors as respondents, estimating text difficulty/discrimination and detector ability across multiple dimensions. Using this, they find domain and length drive benchmark heterogeneity, and a 2,088‑text subset preserves full‑benchmark detector rankings.
The World as Viewed by 136 Countries: An Open, Curated, Annotated, and Multilingual Dataset of 118 Million State Media Articles
AI summaryWorldStateMedia compiles 118 million state‑controlled news articles from 136 countries, annotating outlet control levels. It shows a surge of state media in training corpora and demonstrates that retrieving such sources biases language models toward the originating country’s narratives.
Certified Sparse Attention via Perturbed Retrieval
AI summaryAdds independent Exp(1) perturbations to KV scores, turning sparse key retrieval into probabilistic sampling that gives a Horvitz–Thompson estimate and simultaneous confidence bounds. The method reads only \(O(\varepsilon^{-2}\log(1/\delta))\) keys, achieves 4–8× lower 90th‑percentile error than baselines, and matches dense attention on benchmarks while speeding up decoding.
When Does Better Acceptance Pay? A Controlled Study of Speculative Decoding
AI summaryThis paper evaluates speculative decoding trade-offs under controlled training and inference conditions. The authors find that performance gaps between drafters shrink, and low-cost drafting improvements often outperform expensive ones depending on task uncertainty and serving conditions.
SoK: PIAtlas: Decomposing the Unbounded Prompt Injection Attack Space
Prompt injection is the dominant security risk for large-language-model (LLM) applications and agents, yet the field still lacks a coherent map of how these attacks are constructed. More than twenty public datasets describe prompt injection using incompatible vocabularies, often…
One Head Away from Language Confusion: Single Attention Heads Across Model Families
Large language models sometimes reply in an unexpected language, a failure known as language confusion. We ablate individual attention heads in 12 model checkpoints spanning seven model series, using FLORES-200 continuations in five languages and Language Confusion Benchmark…
Beyond the Teacher: Understanding and Preserving Distillation Overshoot in Agentic On-Policy Distillation
AI summaryThis paper identifies distillation overshoot, where student agents temporarily outperform their teacher before regressing. To prevent this, the proposed method masks conflicting teacher supervision on successful student actions, improving final performance by up to 15 percentage points.
Policy Is Not Capability: Some Safety Benchmarks Measure Which Definition of "Unsafe" You Trained On
AI summaryThe paper shows that safety benchmark scores largely depend on the labeling policy used, not model capability. It demonstrates that minor changes in policy or a few parameters can shift scores across the full range, calling for explicit reporting of contamination and policy definitions.
P2N: Reusing Deep Representations for Greater Effective Depth
AI summaryP2N reuses the deep representation of the previous token to extend effective depth during generation with only a vector addition per step. At 1.5B, it boosts GSM8K accuracy by ~10 points in zero‑shot/few‑shot settings without added latency.
CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?
AI summaryCoEvo‑RSI uses a meta‑agent to inherit both harness designs and model checkpoints from a shared archive, aligning evolving execution scaffolds with learned behaviors to mitigate brain–body mismatch. This approach lets Qwen3.5‑9B, trained only with supervised fine‑tuning, outperform larger models on multi‑agent attribution, role‑aware text‑to‑SQL, and smart‑home tasks.
Dual-Slot: From Serial Loops to Parallel Slots in Language Models
AI summaryDual‑Slot reorganizes looped language model computation into parallel latent and prediction slots, reducing autoregressive latency by replacing two sequential layers with one paired invocation. It improves 1B dense Pile perplexity from 7.25 to 7.00, boosts downstream accuracy, and preserves ~95 % decoding throughput.
Beyond Factual Accuracy: A Benchmark for Evaluating Perspective Alignment in LLMs
AI summaryThis paper introduces PerspectiveBench, a benchmark that measures whether LLMs adhere to a requested perspective rather than just being factually correct. The new Perspective Alignment Score shows that current models keep target content but still mix in unrequested viewpoints, even with RAG.
Deep Fusion Memory: A Chunk-Level Parametric Memory for Large Language Models
AI summaryDeep Fusion Memory (DFM) trains a parametric memory predictor to generate chunk‑level embeddings from context, fusing them into hidden states via gated cross‑attention; this reduces training data needs by 32× and yields up to 12.71% QA gains over baselines.
Generation-Heavy Hybrid Machine Translation
AI summaryGHMT uses only a source parser and small dictionary, overgenerates target syntactic structures with rich target resources, then statistically selects the best translation. It achieves comparable MT quality while requiring far fewer source–target symmetric resources.
Behave Like Experts: Fusing Data with Semantic Knowledge for Effective and Efficient Semantic Map Construction
AI summaryFuses language model representations with data-driven graph construction to guide candidate selection, achieving more accurate and efficient semantic maps. The method outperforms prior approaches across 10 datasets on both effectiveness and speed.
Memories That Outlive Their Reasons: Dependency-Tethered Memory for Tool-Using LLM Agents
AI summaryDependency‑tethered memory records the exact version each verdict depends on and discards any verdict whose support fails, reducing unsafe publications to 0.3% and verification calls by 37% compared to fact memory.
Trackable Skill Library Lifecycle Management for Agentic Reinforcement Learning
AI summaryTrackRL uses online randomized controlled experiments within group-based policy optimization to isolate individual skill contributions in agentic reinforcement learning. This evidence-driven approach enables the principled lifecycle management of skill libraries, outperforming existing baselines in downstream evaluations.
MemBubble: Benchmarking How Agent Memory Shapes LLM Response Diversity
AI summaryMemBubble benchmarks how agent memory affects response diversity using 1,000 real conversation-derived preference instances. Across LLMs and memory designs, entropy drops 34.7 % versus full history, revealing significant diversity loss that user interventions fail to mitigate.
TIRE: Top-Down Formula Instantiation with Reliable Execution for Physics Problem Synthesis
AI summaryTIRE generates physics problems by instantiating reusable formulas and conditions, creating reference solutions before generation for reliable supervision. The resulting TIRE‑19k dataset improves fine‑tuned and RL‑trained LLMs by ~3 pp over comparable datasets.
PLANE: Plug-and-Play Fine-Grained Visual Experts for Hallucination Mitigation in Large Vision-Language Models
AI summaryPLANE introduces fine‑grained LoRA visual experts and an instruction‑aware gate to enrich CLIP features, reducing hallucinations in large vision‑language models. On new GPT4‑free GIFT benchmark and existing tests, it cuts hallucinations with negligible latency, outperforming state‑of‑the‑art methods.
Morpheus: a fully local, defense-in-depth de-identification engine for Greek clinical text - a benchmark evaluation
AI summaryMorpheus removes 98.5 % of direct patient identifiers from synthetic Greek clinical notes with under 3 % over‑redaction, while most remaining leaks are institutional. It is the first benchmark for Greek clinical de‑identification, including an independent egress‑guard evaluation.
Do LLMs Reproduce the Age Gap in Loneliness? Evaluating Survey Simulation against European Survey Data
AI summaryThis paper evaluates whether large language models can accurately simulate the age gap in loneliness across European countries using actual survey data. The results show that LLMs frequently fail to reproduce these demographic differences, often reversing the actual age patterns or systematically misestimating group means.
Legal Guidance in the Wild: How Users Seek Legal Help in Real-World LLM Conversations
AI summaryThe paper builds a taxonomy to detect legal guidance requests in real‑world LLM chats, annotates 2,122 WildChat‑1M conversations, and releases a 200‑conversation gold set. It finds 71% of legal content seeks guidance, with rights‑and‑options advice most common.
Indistinguishable Agents: Training Agent Privacy as Twin-World Indistinguishability
AI summaryIndistinguishable Agents trains agents to act the same in twin worlds differing only by a private fact, minimizing a distinguisher’s log-likelihood ratio; this guarantees Bayes-level privacy while preserving task performance on shopping and messaging tasks.
Memory, Know Your PLACE: Governing Lifecycle State and Query-Time Eligibility in Persistent-Memory Agents
AI summaryPLACE adds write‑side state certification and read‑side authorization to persistent memory, improving lifecycle correctness and preventing causal leakage. Evaluation on Mem0 with Qwen2.5 models shows combined governance yields best performance while preserving relevant support.
Dialogue Generation for Conversational Recommendation Using User Search Interactions and Knowledge Graphs
AI summarySearCl automatically generates 3,538 recommendation dialogues from search and click logs, using a probabilistic state machine and LLM agents guided by knowledge graphs. The generated conversations score higher than benchmark dialogues under an LLM-as-a-judge protocol.
Concept Representations in Vector Space: Toward an Interdisciplinary Study of Meaning Variation
AI summaryThe paper introduces a framework that uses transformer embeddings of standardized text responses to map individuals’ concept representations into a shared vector space, enabling measurement of inter‑individual variation. In a proof‑of‑concept study, the method quantifies concept sharedness, identifies correlates of variation, and detects subpopulation deviations.
Reranking Answers of Large Language Models with Knowledge Graphs
AI summaryThe paper extracts subgraphs around entity candidates by computing shortest paths in Wikidata, then uses features from these subgraphs to rerank LLM-generated answers. Experiments show the reranker improves answer accuracy over baseline LLM outputs.
Acharya 2.0: An AI Tutor for Self-Learning Subjects in Hindi
AI summaryAcharya 2.0 is a retrieval‑augmented LLM‑powered web tutor for Hindi students, providing on‑demand academic help by combining a Hindi knowledge base with a large language model. It demonstrates that RAG can efficiently adapt LLMs to low‑resource Indian languages.
A Thousand Faces: A Persuasion Benchmark with Experience-Grounded Persuadees Informed by Psychology and Pragmatics
AI summaryA Thousand Faces creates a persuasion benchmark with 1,000 detailed personas linking personal experiences to beliefs, desires, and revision grounds. Models scored lower than simpler benchmarks but showed higher faithfulness and use of personal reasons.
Calibration-Free Early Stopping for Chain-of-Thought Reasoning via E-processes
AI summaryCalibration-free early stopping for chain‑of‑thought LLM reasoning uses two e‑processes to sequentially accumulate evidence, guaranteeing a user‑specified false‑alarm rate without calibration. The method matches or outperforms calibration‑based policies across models and datasets.
Hidden Cartographer: Hidden State Guided Data Selection for Dual Budgeted Instruction Tuning
AI summaryHidden Cartographer selects data for instruction tuning by iteratively expanding candidates using hidden states and then pruning with exact training dynamics, achieving the best accuracy on medical, legal, and financial tasks while annotating only 20–30 % of data and fine‑tuning 10–15 %.
Assessing the Empirical Evidence for Loss of Control from Agentic General-Purpose AI
Theoretical literature warns that advanced generalist agentic systems might pursue instrumental goals that are misaligned with human preferences, thus posing risks of irreversible loss of human control. The EU's Code of Practice for General-Purpose AI lists loss of control as a…
SPARC: Spoken Profile Authorship Recognition Corpus
Speech anonymization has focused on the acoustic signal, leaving open how much speaker information the transcribed words themselves retain. In this work, we study pairwise speaker and demographic-attribute verification from transcribed speech. We introduce SPARC (Spoken Profile…
Aligning Language Models with Indigenous Community Cosmovision through Constitutional AI
Large language models (LLMs) are predominantly trained on corpora that reflect Western, anglophone worldviews. This asymmetry is not neutral: when a model answers questions about territory, community authority, or the relationship with nature, it does so from assumptions that…
Learning When to Abort in Transformer Planning: A Mechanistic Study
Planning does not by itself imply the ability to recognise errors, but recognising errors can improve planning capacity. In this paper, we study this phenomenon in Transformer planning. To analyze the inner mechanism, we follow the approach of Wang et al. (2024) and abstract…
Atoms, Not Knowledge: What Production-Scale Addressed-Memory Writes Deposit
Deterministic hash-addressed conditional memory permits local post-hoc knowledge injection. Extending arXiv:2610.21553's bounded-feasibility audit rejecting under-specified generality claims with repaired probes and training-dose/capacity axes, we give, to our knowledge, the…
Where and Why Search-Augmented LLMs Fail When Reasoning over Rapidly Evolving Knowledge
AI summaryTREK is a framework and benchmark that diagnoses why large language models struggle to reason about fast‑changing facts. It shows that the main failure lies in post‑retrieval reasoning, where models get misled by conflicting evidence rather than retrieval errors.
CHSM8: A Chess Transformer that Follows Natural-Language Style Prompts
We present CHSM8, a 210M-parameter chess transformer that plays moves conditioned on a free-text description of a playing style, e.g. "go for the Sicilian Defense", "solid, safe, risk-free chess" or "You are Magnus Carlsen in his prime". The model is pretrained without search on…
Aligning Cultural Behaviors in Arabic Language Models via Preferences
Arabic LLMs have improved rapidly, but aligning them with country-specific cultural behavioral preferences remains underexplored. Previous work has focused primarily on instruction-tuning for factual knowledge, overlooking nuanced cultural behaviors. We propose ArabPref, which…
MemDerive: A Diagnostic Benchmark for Agent Memory over Evolving Data
AI summaryMemDerive is a diagnostic benchmark that records exact derivation DAGs for agent memory over evolving data, enabling error tracing at intermediate states. It shows that existing memory systems lag simple retrieval baselines and rarely propagate implicit changes, highlighting the need for gathering, versioned rules, and dependency maintenance.
Outcome-Anchored Credit Reallocation for Agentic Knowledge Graph Question Answering
AI summaryThe paper introduces Outcome-Anchored Credit Reallocation (OACR) to assign fine-grained operation credits from final answer rewards in agentic knowledge graph question answering. OACR improves multi-turn reasoning performance across three benchmarks while reducing inference-time LLM calls and token usage.
From Unverifiable Domain Text to Cross-Lingual RLVR via Code-Switched Task Synthesis
Reinforcement Learning with Verifiable Rewards (RLVR) can improve reasoning in large language models, but specialized non-English domains expose a data mismatch. Raw English domain text is often available, while target-language RLVR tasks with verifiable rewards are scarce. We…
Towards Figurative Language Understanding in Moroccan Darija: An Evaluation of Large Language Models
Figurative language remains a major challenge for Machine Translation (MT) systems, as the intended meaning of such expressions often goes beyond literal translation. Although recent advances in Large Language Models (LLMs) improved the translation of Arabic dialects, benchmark…
Mixture-of-Memories: Low-Overhead Language Modeling with Modular Knowledge Injection
AI summaryMixture-of-Memories augments a frozen language model with two neural memory experts—latent and lexical—controlled by a budget router that mixes their outputs per token. The model gains better accuracy, updateability, and low inference overhead across language modeling and factual QA.
RARE: Reinforcement Learning with Adaptive Rubric Evolution for Open-Ended Generation
Verifiable rewards have advanced reinforcement learning in mathematics and code. However, many valuable open-ended tasks share three properties: no automatic verifier of overall quality, multiple requirements to satisfy jointly, and multiple valid responses to each query.…
Design Considerations in Low-Resource Archival OCR: A Case Study in Lakota
Building and expanding access to digital archives supports language reclamation and creates lasting resources for community revitalization. Designing effective tools to support these efforts requires taking into account the community-, language-, and document-specific context.…
Enhanced Masked Diffusion Unlearning
Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens with bidirectional context. As these models reach real-world deployment, machine unlearning becomes essential. This process aims to remove the influence of a designated "forget set" without…
System X: An LLM Decision Model That Replies from Its Own KV Cache
Decision models return probabilities over supplied options, but a deployed agent may also need to reply in words. The present study introduces System X, a large language model (LLM) decision model that adds a text-reply option to the LLM-as-Jev readout. When this option wins, or…
Quality-Controlled Reply and Summary Data for Post-Training Small Language Models in AI Call-Screening Agents
AI summaryThe paper creates a privacy‑respecting call‑screening dataset with 3,395 synthetic calls and 15,102 reply/summary targets, training a single LoRA adapter. Adaptation boosts reply accuracy to ~95% and summary quality to 43–72% on held‑out calls.
DiaLong: Benchmarking Reminiscence in Long-Context Dialogues
AI summaryDiaLong is a benchmark dataset of 100 long‑context dialogues with a fact‑recall task, where models must classify 40 candidate facts as true or false. Results show current LLMs still forget long‑range information, indicating room for memory‑enhancing methods.
A failure of Imagination: VLMs fail to use visual representations for text, but self-distillation can help
Vision-language (VL) models are typically built by fine-tuning a language model to attend to visual features in service of a Visual Question-Answering objective. We ask whether this alignment changes how language models represent text---especially on tasks where visualization…
Self-Distillation for Supervised Fine-Tuning
Supervised fine-tuning treats the expert demonstration or gold response as a one-hot target, suppressing valid alternative continuations that the model would otherwise assign probability to. We propose Gold-Conditioned Distillation (GCD), a method that turns expert…
From Seeing to Reporting: A Theory of Reliable Visual Set Generation in Vision-Language Models
Vision-language models (VLMs) are asked to report what they see: to count objects, box them, parse a screen or pick a click target (GUI grounding). We argue that reporting is part of the visual task, not merely its formatting. In a dense scene with 173 shapes, GPT-6.1 lists 173…
Where Are Language-Specific Features Represented in Multilingual LLMs? Everywhere
Previous work has argued that multilingual LLMs encode language-specific features most strongly in their initial and final layers, implying that intermediate layers support language-agnostic representations. This paper reconsiders this finding by performing four different…
ATEST: Adaptive Resource-Aware Test-Time Scaling for LLM Inference
Test-time scaling (TTS) improves large language model (LLM) reasoning by sampling multiple answers, but fixed sampling budgets ignore both query difficulty and the changing energy cost of inference under GPU contention. We measure up to a 4.1-fold increase in energy for the same…
Auditing Tool Traces in Agentic LLMs: Source Localization and Counterfactual Dependence Testing
AI summaryThis paper introduces a two-stage auditing framework to detect hidden label shortcuts in tool-augmented LLM agent traces. Applying the audit to TCR-pMHC binding reveals that apparent scientific reasoning was actually driven by dataset identifiers rather than genuine evidence utilization.
EmotiQ: A Quantum Residual Specialist Framework for Text Emotion Classification
Emotion classification from text is challenging because many affective states are distinguished by subtle linguistic signals and semantically overlapping boundaries. We investigate whether compact quantum modules can serve as targeted residual specialists that assist traditional…
Narrow Decision Models: A System-1 Architecture for Calibrated, Typed Decisions
System One models, introduced with TypeSafe's Jev, answer typed questions about a state with calibrated probabilities instead of generated text: choosing among options described in natural language (Choice), placing the state on an ordered scale (Score), or judging a yes/no…
Flattening Fantasies: Analyzing Stereotypes in Pornographic Discourse
This study presents a large-scale quantitative analysis of 5.4 million video titles from xHamster, the world's fourth most visited pornography platform, to examine linguistic patterns used to describe different dimensions of identities (e.g., race and gender). Specifically, we…
Representations in Motion: Layer-wise Rotation and Trajectory Divergence in LLMs
AI summaryThe paper tracks how language model representations change layer‑by‑layer and across trials, showing that factuality directions rotate apart when the model is told to lie vs. be honest, and that such rotation also appears in other conceptual tasks. It demonstrates that trajectories alone predict behavior, and that divergent points align with the onset of rotation, even in naturalistic, non‑explicitly‑lying settings.
Assertion and Belief: Moore's Paradox in Language Models
Moore's paradox \citep{moore1942} concerns sentences such as ``It is raining, but I don't believe that it is,'' which may well be true yet cannot be sensibly asserted. The absurdity lies not in the content but in the act, since to assert is to vouch for one's own belief. This…
Cognitive Map in Large Language Models
AI summaryLarge language models build abstract relational structures that act as cognitive maps, inferring unseen relations and generalizing to synthetic entities. The study shows models store pairwise links, bind entities via sparse key–value pairs, and can be manipulated to implant false memories.
Text Dominance in Paralinguistic Conflict: Tracing Suppression Through Audio LLM Circuits
Audio LLMs deployed in settings such as crisis hotlines and customer support are only useful beyond speech-to-text pipelines if they use acoustic cues rather than the transcript alone. We test whether they prioritise transcript content over paralinguistic cues (emotion, speaker…
Curriculum Learning Teaches LLMs Long-Horizon Implicit Reasoning, Up To A Limit
LLMs "reason" by externalizing intermediate steps as tokens through chain-of-thought (CoT), yet even frontier reasoning models hit abrupt performance cliffs on long-horizon tasks. Implicit reasoning (reasoning within a single forward pass) is useful during CoT generation, as the…
MapSteal: Map-based Autonomous Knowledge Stealing Attack Against Large Language Models
Knowledge acquisition from black-box large language models (LLMs) offers a practical approach to adapting smaller student models. However, many existing methods rely on predefined queries or student feedback to guide acquisition, leaving autonomous expansion of knowledge targets…
Consistency Training Suppresses Cue Following Without Teaching Cue Selectivity
Language models often change answers because of a cue, a hint such as "I believe the answer is B" that gives no valid evidence. Consistency training (CT) trains a model to answer the same with and without the cue. Four CT methods apply this idea to output tokens (BCT), residual…
Cross-Pollination Retrieval: Finding Transferable Ideas Across Fields in Scientific Literature
Many scientific advances come from cross-pollination, carrying an idea from one field over to a problem in another. Yet retrieving papers across fields remains underexplored, with existing benchmarks confined to a narrow range of fields. We study cross-pollination retrieval, the…
Whose Credential Is It? How Owner Attribution in Agent MemoryShapes Cross-User Account Misuse
Shared long-term memory lets an LLM agent serving one user inadvertently retrieve another user's credentials, but it is unclear when such exposure leads to actual harm. We introduce a three-stage audit that separates three events: a foreign credential reaches the prompt…
Inside the Scorer: Mechanistic Evidence on Rubric Control in LLM-Based Automated Essay Scoring from 25 Base Models
Automated essay scoring (AES) using large language models (LLMs) offers promising use cases for language assessment, yet high-stakes testing requires scores that are both controllable and verifiable, which current LLM-based scorers lack. Given that, we investigate whether…
No Matter Who You Are: Addressivity and Sycophancy in Large Language Models
AI summaryThe paper identifies a link between “addressivity” (the model’s tendency to direct outputs toward an addressee) and sycophancy, showing that post‑training strengthens this connection and that any recognized addressee triggers the effect.
Does Unlearning Survive Collaboration? Robust Unlearning for Shared-Model LLM Teams
AI summaryA method for collaborative LLM teams called ACU reduces the recoverability of sensitive data from the full conversation trajectory, lowering recovery from 41.8% to 19.3% while maintaining task performance. The approach uses message-level reconstruction feedback and replacement-based credit to prevent joint data leakage.
A Calibrated Instrument for Detecting Disposition Drift in Language Models
Before language models are released, they are tested for harmful traits such as willingness to carry out dangerous requests, deception, power seeking, and sycophancy. The safe behavior these tests certify may degrade as models are modified, by topical fine-tuning, new training…
Computation, Not Memory: Chain-of-Thought and State Tracking in LLM Agents
Language-model agents often reason before they act, and the value of monitoring that reasoning depends on how it relates to their actions. We ask how an agent's actions depend on its reasoning when the correct action depends on a changing state. In a small file-system task,…
Truncated Reasoning Under a Budget: Evaluating System-1 Models as Decision-makers
Large language models (LLMs) often generate more reasoning tokens than needed. Limiting the reasoning budget reduces cost, but forcing an answer from an incomplete trace can hurt accuracy. System-1 decision models select among candidate options without producing a reasoning…
Standing on the Shoulders of Giants: Enhancing LLM Agents for Automated Fact-Checking with Human Expert Skills
AI summaryWe distill AFP fact‑checking guidelines into a modular “Professional Fact‑Checking Skill” and give LLM agents dynamic access to it at inference. SkillCheck improves accuracy by 2.2–4.8 pp and Macro‑F1 by 1.5–6.3 pp on AVeriTeC, and with GPT‑4o‑Mini surpasses GPT‑4o while cutting inference cost by ~61 %.
CoMem: Learning to Anticipate Failures from Retrieved Experience
AI summaryCoMem teaches LLMs to anticipate failures by treating retrieved past experiences as inputs to their learned memory. It achieves strong performance on math, code, and QA tasks, generalizes to out‑of‑distribution problems, and lets small models guide stronger solvers.
Modal Sense Disambiguation in Language Models: The Role of Context, Training Domain, and Input Frequency
Modals such as may and must can take on different senses depending on discourse context, alternating between root interpretations (e.g., permission, obligation) and epistemic interpretations (e.g., knowledge, belief). While modals are known to be notoriously difficult for…
When Retrieval Partially Conflicts with Skills: Understanding Procedural Hallucination in Language Models
Language models increasingly use skills that provide reusable instructions and procedures for task execution. After a relevant skill has been selected, the model may need to apply its governing procedure together with retrieved task context. When the retrieved context is useful…
Diffusion Language Models Change Speed Across Tasks and Languages
AI summaryDiffusion language models vary decoding cost across tasks and languages; generating new, rare content requires many forward passes while copying or repeating existing reasoning is cheap. Cost correlates with correctness and can guide efficient inference.
Where RL Breaks: Localizing and Repairing Failures via Targeted Continued Pretraining
AI summaryThe paper shows that RL failures in language models are localized to a few characteristic n‑grams. By retrieving and continued pretraining on just 35M tokens of failure‑related text, it consistently improves reasoning accuracy, outperforming random pretraining data across multiple models and tasks.
Measurement Validity and Representation Dynamics of Confidence Signals in Large Language Models
AI summaryThe paper audits Qwen2.5 models’ confidence signals (verbal, probe-based, sampling agreement) on retrieval and reasoning tasks, finding probes signal agreement earlier for retrieval, verbal confidence unstable across prompts, and sampling agreement offers higher resolution despite higher ECE.
Arabic Financial NLP
Financial communication connects numerical accounts of economic activity with explanations of performance, responsibility and future expectations. Arabic financial natural language processing offers methods for studying these connections across corporate reports, market…
Learning to Reason Efficiently with Interventional Step Credit
Large reasoning models often produce lengthy chain-of-thought (CoT) traces mixing useful intermediate steps with unnecessary ones. Recent methods explore step-level signals such as attention mass and step entropy to identify and suppress unnecessary steps. Despite their…
Budgeted Attack Scheduling for LLM Red-Teaming: The Roles of Static Priors, Semantic Guidance, and Online Feedback
As jailbreak attacks against large language models (LLMs) become increasingly diverse, red-teaming has come to rely on large libraries of jailbreak prompts and attack strategies to evaluate the robustness of victim models and their defense settings. However, attack effectiveness…
SWE-Residual: Evaluating Starting-State Robustness of Coding Agents
Software engineering (SWE) benchmarks are the primary way to measure the capabilities of coding agents: an agent is placed in a code repository or machine environment, given a request, and judged by hidden tests. These benchmarks almost always start the agent from a clean state…
CALM: Consistency-Aware Logit Masking in Decoding Space for Mitigating Overthinking of Large Reasoning Models
Large reasoning models (LRMs) have achieved strong performance on challenging mathematical and scientific tasks through long chain-of-thought reasoning, but they often suffer from overthinking with unnecessary reasoning steps. Existing truncation-based methods typically detect…
FisherEdit: Fisher Information Guided Knowledge Editing for Large Language Models
Sequential knowledge editing in Large Language Models (LLMs) often leads to a degradation of general capabilities due to cumulative parameter interference. In this paper, we revisit prominent projection-based methods like AlphaEdit and reveal a surprising mismatch: their…
LEAPS: Long-context Evaluation of Adaptive Pattern Learning and State Tracking
Long-context language models (LCLMs) power agents that act over long trajectories, from improving a codebase to running research experiments to forecasters that condition on lengthy histories. These applications draw on three distinct long-context capabilities: recall, where the…
EvoSkill v2: Fully Agentic Skill Evolution from Benchmark Adaptation to a Stateful Evolver
EvoSkill v2 is a three-agent system that edits the instructions (skills) an agent is given, leaving its model and harness intact. The user starts a run by describing a benchmark and supplying a grader, which a benchmark adaptor turns into a benchmark-specific plugin. A solver…
Keep the Lineages Alive: Diversity-Preserving SMC for Diffusion LLM Decoding
Inference-time sequential Monte Carlo (SMC) decoding can improve the reasoning accuracy of masked diffusion language models (diffusion LLMs) without post-training by resampling a population of partial generations according to a reward. Its effectiveness therefore depends on the…
MITRA-lexicon: Lexica of attested translations from Sanskrit into English, Tibetan and Chinese
AI summaryMITRA‑lexicon provides 2.18 million Sanskrit headword entries with attested translations into English, Tibetan, and Chinese, each linked to digitized source sentences and aligned at the span level using a fine‑tuned annotation pipeline. The resource outperforms existing dictionaries on benchmark evaluations.
Speech LLMs That Breathe: Modeling Non-Verbal Vocalizations in Dialogue
AI summaryThe paper builds a dialogue dataset with 13 non‑verbal vocalization types and uses multi‑stage fine‑tuning plus NVV‑aware GRPO to train models, boosting expressive NVV generation in VocalNet and VITA‑Qinyu for empathetic and role‑play dialogues.
AdaExploreBench: Benchmarking LLM Agents for Proactive Exploration and Test-time Adaptation in Evolving Environments
AI summaryAdaExploreBench evaluates LLM agents on proactive exploration and test‑time adaptation in environments with hidden, drifting rules. GPT‑6 Sol attains only 61.9% accuracy, and performance improves by ~10% when agents receive evidence or detect rule drift, highlighting active evidence acquisition as a bottleneck.
Token-Set Choice Confounds POPE: A Systematic Audit of Yes/No Extraction in VLM Hallucination Evaluation
We set out to reproduce a recent hallucination-mitigation result on the POPE benchmark. The LLaVA-1.5-7B baseline we measured, 82.21 F1 on the adversarial split, sat 1.21 to 8.67 points above the adversarial baselines we planned to compare against (81.00, 79.30, and 73.54). We…
GLOBE: Geometry-Led Online Behavioral Early Exit via Unit-Sphere Trajectories
AI summaryGLOBE stops chain-of-thought reasoning early by embedding reasoning segments on a unit sphere and monitoring trajectory shape; it cuts tokens by ~50 % and latency by ~51 % while retaining 80.6 % of full‑CoT accuracy across five tasks.
Correctness-Dependent Asymmetric Target Distributions for Self-Distillation Policy Optimization
AI summaryAPSD mixes student and self‑teacher distributions based on rollout correctness, preserving good behavior on correct rollouts and sharpening corrections on incorrect ones. It matches or exceeds SDPO performance while reducing training steps by 2.6× on Qwen3 models.
TempoMoE: Identifying and Steering Fast-Slow Thinking Experts for Efficient Multimodal Reasoning
Multimodal Mixture-of-Experts (MoE) reasoning models rely on long chain-of-thought (CoT) for complex tasks, yet often overthink easy inputs, increasing token cost and latency, and sometimes degrading accuracy. Existing methods optimize reasoning length during training or impose…
Scaling Synthetic Pre-training Data and Evaluation for Indic LLMs
Building capable language models for India requires scalable multilingual data and robust evaluation. We introduce Nemotron-Pretraining-Indic-v1, a synthetic pretraining corpus of approximately 2.83 trillion tokens covering all 22 scheduled Indian languages in native script and…
Accelerated Drift of Scientific Prose in the LLM Era: Evidence from an Interrupted Time-Series Analysis
Every scientific text is embedded in the writing norms of its time: it reflects earlier publications while also shaping the texts that follow. The growing use of LLM-based writing tools introduces a new factor in this evolving system, but the linguistic changes coinciding with…
Understanding When Reflective Prompt and Skill Optimization Helps on Locally Served Reasoning Models: A Two-Model Study of GEPA and SkillOpt
Reflective text-space optimizers such as GEPA and SkillOpt have reported large gains on frontier and mid-size models. We test replication across two reasoning-native target-serving configurations: DeepSeek V4 Flash through DwarfStar and Qwen3.8-27B through LM Studio. For each…
Decision4Rec: A Jev-Inspired Framework of Decision Models on Knowledge and Active Graphs for Multi-Agent Recommender Systems
Decision models, such as the commercial Jev, are language models that return one probability for each option of a question instead of generating text. We present Decision4Rec, an open and modular framework that makes decision models useful for recommendation. A knowledge graph…
Balance-Aware Pivot Selection for Parallel Diffusion LLM Decoding
AI summaryThis paper introduces Balance-Aware Pivot Decoding to accelerate parallel diffusion LLM generation by selecting central positions within masked intervals. This approach achieves logarithmic decoding depth while maintaining generation quality across multiple reasoning and coding benchmarks.
Honeyed Words, Hidden Daggers: Benchmarking and Training LLMs to See Through Mock Politeness
The Chinese idiom "kou mi fu jian" ("honeyed words, hidden daggers") names speech that is polite in form but hostile in function. We present ZhenhuanBench, a 1,592-item benchmark built from expert commentary on Empresses in the Palace, covering speaker intent, emotion-language…
DeepSEA: A Multilingual Evaluation Benchmark for Southeast Asia
While language models have demonstrated remarkable capabilities, their performance remains disproportionately skewed towards high-resource languages. This disparity is particularly acute for the 700 million inhabitants of Southeast Asia, where vast linguistic diversity remains…
VISAGENT: An Agent-Based End-to-End Voice Interaction System
Voice-based public services have evolved from manual switchboards to intelligent platforms to meet modern user needs. However, current systems remain constrained by rigid menu-driven interaction, fragmented vertical services, and limited extensibility. To address these issues,…
Conditional information Retrieval for Longitudinal Survey Harmonisation
Identifying conceptually equivalent questions across different surveys is a challenging information retrieval task where surface-level textual similarity is insufficient. Two questions may be lexically similar yet assess entirely different sociological sub-concepts, or lexically…
Reading Time Modeling with Pixel-Based Language Models
Tokenization-free, pixel-based language models (PLMs) have recently been proposed as an alternative to conventional token-based language models (TLMs), typically motivated by practical considerations, e.g., the flexibility to support new languages or vocabularies. Nevertheless,…
VEILED-SH: Vulnerability Expression through Indirect Language Evaluation Dataset (Self-Harm)
AI systems may miss indirect, culturally situated distress or provide support mismatched to the application context. Existing benchmarks do not jointly address cultural variation, application context, and practitioner disagreement. We introduce VEILED-SH, a benchmark and…
Rethinking BPE: Merge Selection as Model Selection
Subword tokenization methods such as Byte-Pair Encoding (BPE) are a core component of modern language models; yet, tokenizer capacity remains typically determined by a fixed merge budget, chosen heuristically. Since larger vocabularies increase memory usage and computational…
Heartbeat Thinking: Speculative Reasoning and Tool Execution While Listening in Realtime Voice Agents
Tool-using voice agents often leave users waiting after speech ends, even when actionable information arrives earlier in the request. Speculative approaches begin work during speech, but chunk-bounded or interrupt-on-update generation can truncate a tool call before its…
Shaping the Document, Not the Model: Optimized Suffixes for Context Following
When a user supplies a document and wants the answer drawn from it, a language model is expected to follow it even when it contradicts its parametric knowledge. Prior work has focused on making the model rely on context by steering activations, contrasting logits during…
Machina Pauliana: A neuro-symbolic approach to the actio pauliana in Dutch insolvency law
Applying a legal provision frequently requires both structured rule-based reasoning and sensitivity to open-textured terms developed in case law. The actio pauliana of Article 42 of the Dutch Bankruptcy Act (Faillissementswet, Fw), which allows an insolvency practitioner to…
AudioScorer: Learning Question-Aware Token Relevance for Large Audio Language Models
AI summaryAudioScorer learns to rank audio tokens by question relevance, distilling answer-gradient saliency from a frozen LALM. It retains only 10% of tokens yet outperforms or matches other pruning methods on 7 of 10 semantic and acoustic tasks.
Answering While You Type: Optimizing Time-to-Answer with Streaming Generation
Work on efficient large-language-model inference has concentrated almost entirely on reducing the cost of decoding once a request is complete. The quantity a user experiences, however, is the interval between submitting a request and receiving a usable response, and that…
The Scheduler Is a Hidden Retriever: How Asynchrony Shapes Evidence Acquisition in Search Agents
Language-model agents increasingly solve complex tasks through parallel reasoning and asynchronous information gathering. Controlled replay across two model families and three benchmarks shows that delivery timing redirects search decisions made before all results arrive,…
Typed Decision Models under Prevalence Shift: Decision Training, Calibration, and Implicit Priors
Typed decision models (TDMs) return a probability over caller-defined options, and Jev is described by its vendor as calibrated. But in deployment class frequencies change, so calibration measured on a fixed test set need not hold where thresholds are applied. Existing…
Large-Scale Biomedical RAG: Corpus Scaling, Document Representation, and Efficiency Trade-offs
We present a controlled, large-scale evaluation of how corpus scale, document representation, and reranking affect the effectiveness and efficiency of biomedical Retrieval-Augmented Generation (RAG) systems. Retrieval collections were progressively expanded while retaining the…
VibeTraining-Bench: Data-Trainer-Harness Co-Design for Iterative System Improvement
AI summaryVibeTraining-Bench evaluates agents’ ability to jointly refine training data, trainer, and inference harness through iterative feedback, across diverse tasks. GPT‑6‑Astra lifts a base system from 32.9 % to 53.7 % on SWE‑bench Verified, approaching a human‑crafted 55.4 % benchmark.
An Empirical Study of Parameter-Efficient Fine-Tuning for Mixture-of-Experts LLMs
Parameter-efficient fine-tuning (PEFT) methods are critical for adapting large models under resource-constrained settings. However, as mixture-of-experts (MoE) architectures, which offer efficient inference at massive parameter scales, gain increasing prevalence in the…
Bridged Self-Distillation for On-Policy Learning
On-policy distillation can transfer the reasoning capabilities of a strong teacher to a smaller student. However, distillation from a stronger teacher model faces two mismatches: its solutions may be difficult for a smaller student to imitate, and its token-level probabilities…
How Benchmark Graph Construction Shapes Multi-Hop KG Navigation: A Controlled Audit of PathQuestion and PathQuestion-Large
Multi-hop knowledge-graph question answering (KGQA) benchmarks are often treated as fixed language-and-reasoning tasks, but for navigation-based methods, difficulty also depends on the graph in which those tasks are instantiated. This makes graph construction a consequential but…
LaMER: Language-Conditioned Multi-Expert Routing for Weakly Supervised Referring Expression Comprehension
Referring Expression Comprehension (REC) localizes the image region described by a natural language expression, underpinning applications such as autonomous navigation, human-robot interaction, and medical image analysis. Under weak supervision, REC is challenged by the…
Expert L2 English Speech Evaluation on a Gaming PC: Accuracy, Fluency, and Prosody with XLS-R, XEUS, and PhoneticXeus
oollim is a locally executable system for evaluating L2 English pronunciation accuracy, fluency, and prosody on a consumer computer. The system combines pretrained XLS-R, XEUS, and PhoneticXeus representations with reference-text features and ridge regressions trained on…
Where Does Rhetorical Distortion Live? A Multi-Platform, Multi-Language Measurement of Influencer Content
Much of the manipulation in online influencer content is not false but disproportionate: an accurate fact is framed as historic, urgent, or catastrophic. Fact-checking and clickbait detection are not designed to capture such framing, and little is known about where it…
Co-Evolving Policy and Reward for Code Generation: A Mechanistic Study of Single-Adapter Drift
Reinforcement learning with verifiable rewards (RLVR) is particularly well suited to code generation because generated programs can be objectively evaluated through execution. Execution also provides behavioral information that can be reused as additional supervision. The…
Self-Consistency of Entity Sets and Citations in a Deep Research System under Repeated Prompting
AI summaryThe study evaluates a deep research agent’s consistency by repeatedly prompting it on a cybersecurity task and measuring overlap of named entities and citations. Median entity Jaccard is 0.12, citation Jaccard <0.02, indicating low self‑consistency that varies across entity classes.
Governing the Rule Space. Ruliology, Computational Irreducibility and Anticipatory Governance of Quantum Technologies
AI summaryThe paper introduces a Ruliological Framework that treats regulation as exploratory mapping of rule transformations and trajectories, using computational irreducibility as a conditional tool; it applies this approach to the Quantum Europe Strategy, showing how anticipatory governance can adapt rules to evolving quantum tech.
SNO-PI: Label-Scarce SNOMED CT Entity Linking with Pre-trained Encoders and Curated Terminology Knowledge
We present SNO-PI, the top-performing submission on SNOMED International's SNOMED CT Entity Linking Benchmark as of 18 October 2026. SNO-PI achieves a macro character IoU of 0.5387, outperforming the organisers' reference system (0.4582; based on the 2024 challenge winner's…
HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL
Rubric-based reinforcement learning (RL) trains policies for open-ended generation with natural-language criteria as rewards, but these rewards can become ineffective as training progresses. We identify two failure modes. Under non-discriminative rewards, scores collapse to…
The Geometry of Agent Memory: Forecasting Retrieval-Bound Memory Failures, and Characterizing Its Limits
AI summaryThe paper builds a shared embedding space for conversation turns, memories, and queries and trains a logistic classifier to predict answer correctness from 25 geometric features. It achieves an AUROC of 0.640 (vs. 0.655 baseline), adding negligible improvement beyond question difficulty and system identity.
WenSight-Decision-1 for Direct Decisions under E-commerce Rules
E-commerce decisions require applying natural-language rules to the available case evidence. We introduce WenSight-Decision-1, a shared model that scores supplied options directly from rules and case information for six operational tasks, with examples spanning six languages. We…
Label Authority in Speech Emotion Recognition: Intended and Perceived Ground Truth in Emozionalmente
Ground truth in Speech Emotion Recognition (SER) is often presented as a single categorical fact, yet an acted emotion and the emotion heard by listeners describe different aspects of the same recording. We examine this distinction in Emozionalmente, an acted Italian corpus that…
TeleASR-Amharic: A Benchmark for Robust Amharic ASR in Realistic Telecommunication Conditions
The real-world usability of ASR models remains challenging, particularly for low-resource languages, where existing speech resources are often collected in clean studio environments. In this work, we introduce TeleASR-Amharic, a benchmark for evaluating Amharic ASR under various…
Training Negatives Still Matter After Calibrating Answerability Readouts
Answerability readouts predict whether a passage supplies an answer, but whether calibration can compensate for removing negative passages during training remains unclear. We present a controlled study of how training-negative construction affects calibrated prediction, ranking,…
AfriGuard: Safety and Robust Refusal For African Languages
Large language models (LLMs) designed to interact with users are typically trained to respond appropriately to harmful or toxic content. However, ensuring appropriate responses becomes more challenging in diverse linguistic and cultural contexts, where safety risks may be…
Strength Controls Change How Interventions Compare Across Sparse Autoencoder Codes
AI summaryThis paper evaluates how different strength controls, such as norm matching and output calibration, affect sparse autoencoder intervention comparisons in language models. It shows that full-output calibration preserves transfer decisions better than total-norm matching.
Mitigating Reward Hacking in Preference Reward Models with Evolving Penalties
AI summaryDynalize automatically generates penalties from a language model that identifies reward‑hacking behaviors during policy training, applying them to rollouts. This online intervention lets open‑ended models train longer and improves downstream performance, reducing hacking patterns in tasks like counseling and creative writing.
RouterWorld: Scaling Conversational AI with Latency-Aware LLM Routing
AI summaryRouterWorld benchmarks conversational LLM routing with realistic latency and partial observability, offering 371k responses and 76k agent trajectories. It shows most existing routers prioritize accuracy over latency when only sparse logs or streaming data are available.
A Fisher-Geometric View of Multi-Teacher On-Policy Distillation
Consolidating specialized foundation models into a unified generalist is a central goal of post-training. Two approaches are in common use: model merging, which is fast but treated as a heuristic, and multi-teacher on-policy distillation, which matches distributions but demands…
ReAnchor: Reusing Rejected Verifier States for Diffusion Drafting
Block diffusion drafters accelerate speculative decoding by generating multiple draft tokens in a single forward pass. However, each pass produces only a fixed-length block from a single target-generated anchor, limiting how far the draft can extend. Meanwhile, target…
Beyond Attack Success Rate: Cost-Aware Metric for Evaluating Iterative LLM Jailbreaks
Attack success rate (ASR) evaluates jailbreaks against large language models without accounting for repeated calls to the victim model. Yet each additional call gives the deployed system another opportunity to detect the attack and intervene. We introduce…
BERT Is Already a Denoiser: Cheaply Unlocking Iterative Generation in Masked Language Models
Masked language models (MLMs) such as BERT are typically used for representation learning despite their denoising pretraining. We test whether this capability supports conditional generation without diffusion training. Our decoder repeatedly predicts masked tokens and commits…
Conducting Reliable Keyword Estimation with Unreliable ASR in Police Audio Data
Extracting insights from large audio data sets using human transcription is expensive and time-consuming. Off-the-shelf Automatic Speech Recognition (ASR) systems offer an easy-to-use scalable alternative, but systematic transcription errors, especially in domain-specific…
When the harness decides success: state that cannot be re-observed
Agent harnesses decide what a language model sees across turns: they drop old turns, summarize them, or replay the model's own earlier reasoning. We ask when these choices change whether an agent succeeds, not only what it costs, and report three findings. First, in a…
From Outcomes to Components: A Counterfactual Evaluation Framework for LLM-Proposed Repairs
An LLM-generated repair action specifies which records to change and by how much. A failed action may still contain a useful target set or numerical update, depending on how the two are paired. We evaluate this problem in traffic networks and databases, where aggregate…
The Development of Sycophancy in Language Models: Tracing Vulnerabilities from Pretraining to Post-Training
Language models can change a committed answer to accommodate a user's opinion or pushback without new evidence. However, when this behaviour develops during training and what a released model inherits from earlier stages remain unclear. We follow 741 checkpoints of 30 open model…
SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes?
Spatial intelligence requires agents to actively acquire visual evidence and integrate observations into a coherent representation of the three-dimensional world. Yet spatial question answering offers limited evidence of whether models can construct accurate metric maps through…
Which Model for Which Role? Cost-Efficient Pareto Recommendation for LLM Agentic Workflows
LLM agentic workflows increasingly employ different models for specialized roles. Users may prefer different trade-offs between task quality and deployment cost, sometimes also considering latency. Finding good assignments incurs a separate search cost from evaluating multiple…
Intra-Rubric Reward Divergence: Mitigating Rubric Hacking through Latent Dimension Recovery and Balancing
Rubric hacking limits both performance and training robustness in rubric-based reinforcement learning. By decomposing rubrics into sub-dimensions, we reveal the Intra-Rubric Reward Divergence: as the total rubric score increases, certain sub-dimension scores systematically…
Parallel Language Models via Variational Decoding
AI summaryParallel Language Models generate multiple tokens per forward pass via latent auxiliary variables, trained with a variational objective that combines autoregressive and parallel decoding losses. On 80 M–3 B models, PLMs match autoregressive quality while achieving 2.3×–5.7× inference speedups.
Hierarchical Token Merging with Recursive Autoencoders for Unsupervised Grammar Induction
Hierarchical abstraction is a fundamental principle across many data modalities, enabling complex sequences to be represented compositionally at multiple levels. Learning such hierarchies directly from data, without explicit structural supervision, remains a central challenge in…
HoloCosm: Beyond Single-Metric Benchmark Compression Across Multiple Metrics
Benchmark compression reduces evaluation cost by selecting a small set of samples. Existing methods typically guide selection using a single metric, although benchmarks often assess models along multiple dimensions. We find that model rankings are often weakly correlated across…
Copying Under Repetition: From Prefix Matching to Position Hashing
Frontier LLMs can copy random text almost flawlessly, yet they often overrun or truncate long blocks of repeated characters. In repeated strings, matching local content no longer reliably identifies where copying should continue or when it should stop. Nevertheless, we show that…
EpiLens: Lightweight RAG for Epidemiological Knowledge Acquisition
To prepare for future outbreaks, researchers need to know what studies have been done and what data they use. Mapping a large literature by hand is slow, while large language models alone can misstate or conflate sources. We present EpiLens, a lightweight, open-source…
recAMPLIFY: Large-Scale Pretraining of Recursive Protein Language Models
AI summaryrecAMPLIFY trains a recursive masked protein language model with 155 M unique parameters on 2 T tokens, achieving a 22.5% lift in fold accuracy and better contact recovery while using 56% fewer parameters. Gated Mixture Recurrence (GMR) adds adaptive gating, reducing expected depth by up to 37% with modest accuracy loss.
TokenGate: Adapting LLMs for Fast and Memory-Efficient Prefill with Learned Token Pruning
AI summaryTokenGate adds a lightweight module to transformers that learns to prune input tokens during prefill, using token‑gated attention to bypass the non‑differentiable top‑K. It yields 1.4–11.3× faster TTFT and 1.5–3.4× lower memory without accuracy loss, outperforming prior token‑pruning methods.
HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs
Harness development plays a crucial role in building capable AI agents. Recent approaches improve harnesses through recursive self-improvement (RSI), but their performance gains often come with higher execution costs. We introduce HarnessPro, a label-free method for evolving…
Monsoon: Scalable Spontaneous Speech Collection for Low-Resource ASR in 50 Languages
For most languages, automatic speech recognition (ASR) is limited less by model architecture than by a lack of accurately transcribed spontaneous speech. We present Monsoon, a framework for collecting such data at scale. It elicits unscripted conversations between paired…
Game of Phones: Harnessing Game Theory and LLMs in 'Bot Wars' to Counteract Phone Scams
Phone scams impose substantial and growing costs on individuals and financial systems worldwide, yet existing countermeasures impose no resource cost on perpetrators. Strategic scam-baiting, in which automated agents engage scammers to deplete their operational capacity, offers…
RESIN: Multi-Chain Reasoning with Entity-Structured Indexing
Multi-chain reasoning is a challenging task not only because the multi-chain evidence is dispersed, but also because different parts of reasoning may require distinct, interacting chains of support. Flat retrieval can flood a model with loosely related text, whereas early…
Teacher Supervision, Fixed Representations and Calibration: Exact Risk Calculations for Candidate-Scoring Decision Models
Candidate-scoring decision models return a probability distribution over a presented set of answers. Public documentation of one such system, Jev, and of an open counterpart, JevAny, describes three training ingredients that determine what these probabilities estimate: soft…
Safety Mid-Training: Internalizing LLM Safety as a Foundational Capability
Safety alignment for large language models (LLMs) is primarily performed during post-training, while recent work has also begun incorporating safety data into pre-training. Yet safety can degrade when later training no longer includes safety data, and gains in safety often come…
ASTOR: Benchmarking and Improving Autoformulation for Soft-Constraint Optimization with a Formal Grammar
Real-world decision problems combine two qualitatively different kinds of constraints: requirements that must be satisfied, called hard constraints, and preferences that can be violated, called soft constraints. Because these problems are often posed to decision-makers in…
AfriBED: a Speech Recognition Bias Evaluation Dataset for Low Resource African Languages
AI summaryAfriBED is a 9‑hour Bemba and 18‑hour Nyanja speech dataset designed to evaluate gender and speaker‑type bias in ASR. Baseline experiments show it effectively reveals performance disparities across speaker subgroups in low‑resource African languages.
SpecMRL: Redesign Matryoshka Representation Learning within Parametric Spectrum
AI summarySpecMRL reorganizes Matryoshka Representation Learning by learning a spectral basis weighted by feature importance, enabling more effective prefix readout. It achieves higher performance than vanilla MRL across all reduced dimensions and even exceeds full‑dimensional MRL with only half the embeddings.
MemT-1.0: Towards Memory-Native Evolution
AI summaryMemT-1.0 integrates memory directly into the model, using Gated DeltaNet-2 and Mixture-of-Storage, and is trained on 34B tokens. It outperforms parametric memory baselines, achieving 1.86× faster inference and 66× less storage at 128K token history.
RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development
AI summaryRefCraft introduces a benchmark of 300 multimodal tasks evaluating AI agents on webpage development using visual references and text instructions. Experiments show that current frontier models struggle with this workflow, with the top system achieving a score of only 52.12 out of 100.
KEEP-Bench: Evaluating Evidence Preservation in Context Compression
Context compression improves long-context efficiency but may introduce faithfulness issues through an information bottleneck. Existing evaluations rely on QA metrics, but a reader can still answer correctly from an unfaithful compression, so these metrics underestimate such…
Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
AI summaryWildAWM synthesizes 341 long‑horizon computer‑use tasks in controllable, multi‑interface environments using open‑source models, and training Qwen agents in these worlds boosts performance on eight out‑of‑distribution benchmarks.
Competing Learning in Hybrid Architectures
AI summaryHybrid models combine fast linear recurrence and softmax attention, but a “Competing Learning” effect causes the recurrence to dominate retrieval tasks, preventing attention from learning. Stopping recurrence updates mid‑training and then training attention improves language modeling and retrieval performance.
LatComm: Reusable Context Compression for Latent Communication between Large Language Models
Systems that combine several large language models (LLMs) often use one context for many questions. When the context is passed between models as text, the receiving model processes all of its tokens again for every question. Avoiding this cost requires a context representation…
WeAct: Dynamic Weight Switching as Agent Actions
Long-horizon agents must learn heterogeneous capabilities from interleaved trajectories. Updating shared parameters to learn new capabilities can degrade previously acquired ones, resulting in catastrophic forgetting. As a trajectory unfolds, different stages may require…
OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory
AI summaryOpenVLHarness adds a tool‑capability layer and persistent multimodal memory, enabling large language models to fetch, interpret, and store heterogeneous evidence across reasoning steps. It boosts performance on 22 benchmarks by 9.7–12.6 points over no‑tool baselines while reducing token cost.
Can It Actually Work? Synthesizing Functional CAD Mechanisms
While LLM agents can generate visually plausible CAD assemblies, these designs often fail to fulfill their intended physical functions. We study functional CAD design, a setting where agents interpret underspecified requests to construct and iteratively revise mechanisms that…