New Artificial Intelligence (cs.AI) papers, Thursday 22 October 2026
211 papers announced on arXiv. Get them ranked for you, free →
S^3martCirc: Self-supervised Smart Circuit Discovery
Large Language Models (LLMs) have demonstrated remarkable performance across diverse tasks, from text summarization to question answering. Despite these capabilities, their black-box nature obscures internal decision-making processes. Mechanistic interpretability (MI) aims to…
Zero-Cost Detection of Fabricated Citations: A Staged Heuristic Approach for Computing Dissertation Assessment
AI summaryThe paper introduces the Fabrication Severity Stack, a three-level taxonomy for citation fraud, and a staged heuristic workflow that lets general-purpose LLMs flag fabricated citations at zero cost. In a 100-citation pilot, the approach detected 90–100% of non‑existent references.
The CIDOC CRM Periodic Table: Documentation You Can Test to Support the Adoption of Scientific Research Ontologies for the Humanities
The CIDOC CRM family is not only a schema for museum collections. Its extensions formalise the apparatus through which research knowledge is produced and justified: CRMsci models scientific observation and measurement, CRMdig the provenance of digital processes, CRMinf inference…
What Git Does Not Record: Provenance Legibility in Two AI-Assisted Research Repositories
Publishers and funders increasingly ask authors to disclose how AI tools contributed to a piece of research. In computational work the repository is where an author or an auditor would look for the answer, and we ask what it can actually supply. We audit two single-investigator…
FutureLens: Longitudinal Topic Patterns in Scholarly Text and Future-Work References
Future-work statements offer a record of the questions researchers choose to leave open. Using them to describe a field, however, requires separating thematic change from changes in document coverage, topic assignment, and scholarly writing. FutureLens examines these measurement…
CompLit: Scientific Literature Search Benchmarks Must Cover Implicit, Cumulative, and Unmet Needs
How much can we trust an agent to find the paper we need? As researchers, we search the literature from what we have read toward what we have not. We thus describe what we want in our own terms, which a paper may show only implicitly; our requirements grow as we read; and what…
Beyond source attribution: Address authorship and comparability decisions in an LLM-built scholarly knowledge graph
A knowledge graph needs a stable address for its evidence even where the literature offers competing classifications and agrees on none. Yet an address chosen for retrieval readily passes for the only classification available. This happens whenever records gathered under one…
Learning Checklists to Reduce Lottery Effects in Scientific Peer Review
Reviewer lottery error describes error due to a small set of peer reviews attending to only some of the issues with a paper that a larger pool of eligible reviewers would raise. We show this error can be reduced by surfacing a checklist of criticisms learned from a historical…
Who Verifies the Builder? Toward Independent Assurance for Agent-Produced Software
Coding agents can now implement a specification, write and run tests, repair failures, and declare the result ready for release. When the same agentic system produces both the software and the evidence of its correctness, a passing test suite is no longer independent evidence:…
VHDLBench: Benchmarking and Fine-Tuning Large Language Models for VHDL-Based Chip Design
The rapid evolution of digital circuit complexity demands innovative approaches to hardware design automation. While Large Language Models (LLMs) have influenced software development, their application to Hardware Description Language (HDL) generation remains largely unexplored,…
VerifierBench: Benchmarking the Generation of Executable Verifiers for LLM Agents
Agents need reliable ways to validate their outputs, both when completing tasks and when filtering synthetic training data. We study verifier generation: from the instruction and environment available for solving a task, an agent writes an executable verifier that must accept…
A Small Model Asks Better: Fine-Tuning Query Clarification for Retrieval-Augmented Generation
Query clarification is essential for retrieval-augmented generation (RAG) systems that face ambiguous user queries. We consider multi-turn grounded conversational scenarios where a user query is either ill-formed or ambiguous in the context of the prior dialog or, more…
Interaction-Gated Time-Frequency Learning for Parameter-Efficient PPG-Based Heart Rate Estimation
AI summaryTF-IG learns heart rate from PPG by combining dilated temporal convolutions and log‑spectrogram frequency features through an interaction‑gated fusion module. With only 84.6 k parameters it achieves 1.227 BPM MAE and 92.3 % within 3 BPM on BIDMC.
Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds
AI summaryContinued pre‑training on 120 settings of model size, data volume, and world diversity shows that more data boosts in‑corpus performance, diversity improves transfer to related worlds, and larger models are more data‑efficient but generalize worse to distant worlds.
Reinforcement learning for hardware acceleration and system-level optimization on FPGA and integrated-circuit platforms: A survey
AI summarySurvey of 100+ studies on deploying reinforcement learning on FPGAs and ICs, classifying by RL family, abstraction, and application, and evaluating resource use, latency, throughput, and power. Highlights common architectural patterns, performance gains, and open challenges.
SpecEvo: Speculative Evolution with Large Language Models for Cost-Efficient Scientific Discovery
AI summarySpecEvo uses a multi‑LLM system that alternates cheap “Speculators” and an Advisor for most search steps, while a frontier Navigator intervenes sparingly, saving up to 80 % API cost and improving discovery efficiency by up to 1.5× compared to other baselines.
STAR-XPC: Set Transformer for Estimating X-ray Transmission in Polymer Composites
AI summaryWe train a Set Transformer with ChemBERTa embeddings to predict X‑ray transmission ratios from unordered mixtures of up to 11 polymer constituents. On 4,697 Monte Carlo simulations the model achieves R² = 0.9988 and median absolute error ≈ 1.2 %, computing each prediction in <1 ms.
Offline Apprenticeship Learning for Human-Centric Environments
AI summaryOffline Apprenticeship Learning infers reward functions from expert demos to learn policies without interaction. The proposed methods (EM‑EDM, THEMES, HALIDE, CCHAL) outperform baselines on educational and clinical datasets and achieve significant learning gains in a real classroom setting.
GPU-Accelerated Construction of the Gabriel Graph: Parallelization Strategies and Applications to Large-Margin Classification and Graph-Based Clustering
The Gabriel graph is a proximity graph widely used in computational geometry and geometric machine learning, underlying large-margin classifiers and, as we show in this work, density-aware clustering methods. Its construction, however, has cubic time complexity $O(N^3)$ in the…
Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?
AI summaryAgent4Sci tests how well frontier agents improve scientific ML models across 20 benchmarks. Claude Opus gains 10.4% on average, mainly via ML tricks; adding research‑ideation plugins boosts it to 20.1% with more domain‑relevant progress.
SkillEval: Learning Interpretable Ability Profiles of LLMs via Cognitive Diagnosis Models
AI summarySkillEval automatically extracts interpretable skills from benchmark questions and uses cognitive diagnosis models to estimate each LLM’s per‑skill ability, producing calibrated, concise skill profiles that outperform average‑score evaluation in predicting unseen performance and enabling cost‑effective routing and adaptive testing.
ReAPO: Retry-Aware Policy Optimization for Efficient Exploration in LLM Reasoning
AI summaryReAPO improves LLM reasoning by retry‑aware policy optimization that assigns credit to each attempt via a stochastic shortest‑path view, allocating more generations to failed trajectories. It outperforms GRPO and baselines on Qwen3/OLMo benchmarks, achieving higher first‑attempt accuracy with fewer tokens.
Second Thoughts: Exploring and Planning through Comparative Imagination in World Models
AI summaryThe paper introduces Sibling Disagreement in Imagination, which uses a shared actor reference in world models to unify exploration and plan execution. SDI outperforms or matches strong baselines across Meta-World, Adroit, and DM-Control benchmarks.
ArcFlow: Training Flow Policies for Multiple Solver Budgets
AI summaryArcFlow trains a single flow policy that performs well across a range of solver budgets, using budget exposure, action-aligned supervision, and worst-budget checkpointing. It achieves the highest one-step return on 25 of 30 DM Control tasks, outperforming five baseline flow policies.
CALLIOPE: Private Insights into AI Use, Without Embeddings
AI summaryCALLIOPE is a top‑down framework that builds topic hierarchies from large unstructured corpora, using LLMs directly instead of embeddings, while providing differential privacy. It outperforms embedding‑based DP methods, especially in high‑privacy settings.
Towards Realistic Models of Change: A Causal Perspective on Decision-Making and Dynamical Systems
AI summaryThe thesis develops realistic causal models for decision‑making and dynamical systems. It proposes a cost‑aware, temporally aware causal formulation of algorithmic recourse and links SCM equilibria to stochastic process dynamics, enabling robust, interpretable interventions under distribution shift.
Combining Long-Term Memory and LLMs Comes at a High Cost: From Prompt Injection to Persistent Compromise
The widespread adoption of generative artificial intelligence in recent years has brought a parallel rise in prompt injection and jailbreak attacks. By embedding malicious instructions, these exploits manipulate language models to bypass intended behavior and operational…
Forging More Tasks into Better Harnesses: Agentic Harness Evolution via Successive Scale Racing
AI summaryAgentic harness evolution adapts instructions, tools, and control flow from task experience, but expands task pools cause performance drop due to poor evidence selection. Successive Scale Racing (SSR) dynamically constructs diverse evidence sets, races them efficiently, and outperforms baselines, boosting pass rates from 64.3% to 68.3% while cutting validation runs by 60%.
Brittle by Design? Multi-Scenario Reinforcement Learning for Robust Mission Coordination in Systems of Systems
As systems engineering (SE) shifts from monolithic systems to Systems of Systems, Mission Engineering (ME) has emerged as a new line of thinking in SE. Mission environments are uncertain and dynamic, and outcomes depend on how mission assets interact with them, rendering static…
Decision language models as affordable and rapid alternatives to LLM judges of clinical generative AI
AI summaryDecision language models can evaluate clinical AI outputs more cheaply and quickly than large LLM judges while matching or exceeding their agreement with human experts across multiple datasets, showing potential for scalable clinical AI assessment.
Where Should a Stuck Agent Spend Its Next Tokens? Causal Attribution of Inference-Time Resources
Inference-time scaling is often treated as a token-budget problem, but a failed coding agent can seek qualitatively different resources: computation, information, or feedback. We introduce Banzhaf Resource Attribution (BRA), a controlled factorial framework that measures each…
Sketch2Sim: an end-to-end modular framework for hand-drawn circuit analysis
AI summarySketch2Sim converts hand‑drawn RLC circuit photos into SPICE‑compatible netlists using YOLO component/text detectors, CRNN OCR, and DBSCAN node clustering, then solves for voltages and currents. It achieves robust accuracy across varied handwriting, lighting, and clutter.
How to Use Parallel Data for Multilingual Pretraining
AI summaryThe paper proposes concatenating each document with its translation during multilingual pretraining (PPT) and shows that a two‑stage method—PPT followed by standard pretraining—consistently outperforms single‑stage pretraining on language understanding and cross‑lingual tasks within the same token budget.
When Do Gaussian Priors Fail? A Latent Geometry Investigation of RL Fine-Tuning
AI summaryGaussian priors in VAE latent spaces cause severe mode overlap when fine‑tuned with reinforcement learning, leading to a collapse at high beta. Using a Gaussian‑mixture prior separates modes, reducing collapse to 71× lower beta and improving coverage.
Speedrunning Relational Foundation Model Pretraining
AI summarySpeedrun Relational Foundation Model Pretraining introduces a leaderboard that measures wall-clock time to reach a quality target for relational Transformers, achieving a record 305.63‑second run on an NVIDIA H200—an 8.11× speedup over the reference—by using a trilinear attention mechanism and a calibrated benchmark.
DualHeads: Reward-tilted few-step diffusion made simple
AI summaryDualHeads attaches two lightweight binary heads to a frozen few‑step diffusion model: one scores samples relative to the same prompt, the other distinguishes references from generated images. Optimizing the weighted sum of logits yields task rewards nearly matching 40‑step RL methods while keeping image quality and inference cost unchanged.
Agent AI Zarz\k{a}dzaj\k{a}cy Sieci\k{a} Open RAN Zorientowana na Użytkownika
Jednym z najważniejszych wyzwań zwi\k{a}zanych z implementacj\k{a} sieci Open RAN o architekturze zorientowanej na użytkownika jest odpowiednie formowanie tzw. klastrów obs{\l}uguj\k{a}cych. Problem ten może być efektywnie rozwi\k{a}zany za pomoc\k{a} zaproponowanego Agenta AI…
A General Topological Symmetry Index for Critical Orbits In Neural Network Loss Landscapes
AI summaryThe paper defines a topological symmetry index using equivariant degree theory to detect isotropy and quantify topological contributions of critical orbits in neural‑network loss landscapes, applicable to any architecture and handling degenerate orbits, demonstrated on softplus MLPs, Transformers, and equivariant nets.
To Manipulate or Not: Opinion Dynamics in LLM Interactions
AI summaryLLM‑LLM debates on climate‑change attribution show that opinion shifts depend on the specific models used, not just on assigned stances. Same‑model exchanges narrow views for one model, while cross‑model pairs widen gaps; persuasive prompts do not consistently boost influence.
inr2action: Encoding Robot Episodes as Neural Implicit Representations for Recall and Action
Long-horizon robotic operation requires memory that scales with essential information rather than execution time. We present a parameter-space memory framework where each episode is stored as a separable object. A neural implicit representation is fitted to episode video from a…
RL-IF: Tracking Non-local Training Data Influence for Online Reinforcement Learning
AI summaryRL-IF tracks how individual training samples affect later policy updates across multiple online RL rounds, providing a first‑order estimator with theoretical error bounds. Using RL-IF, the RINSE filter removes low‑influence experiences, improving training speed and final performance on control and RLHF tasks.
SANE: Automated Scorer-Specific Metamorphic Testing for LLM Judges
AI summarySANE automatically generates scorer‑specific perturbation recipes for LLM judges, validates them, and produces 3.7× more test families than JRH while detecting deficiencies across domains and aligning judge rankings with human faithfulness judgments (ρ=0.852).
Fig: Automated Multiagent Behavior Elicitor
AI summaryFig introduces a system that elicits target multiagent behaviors by automatically generating or selecting games and applying engine, opponent, instruction, and context interventions based on a natural language description. Models attempt the desired behaviors 45.2% of the time, with interventions improving rates by 10.9% on average.
Hidden Messages Between Diffusion Language Model Agents: Causal Evidence for Latent Communication
AI summaryDiffusion language models can embed a hidden message in a compressed 4‑vector state that a second agent consumes while generating text. Experiments show this single, late transfer improves performance by up to 100 points, proving causal, content‑dependent communication.
JaxMARL-BC: Multi-Agent Reinforcement Learning for Macroeconomic Modelling, in Seconds
Multi-agent reinforcement learning (MARL) offers a principled way of populating simulated economies with many heterogeneous decision makers. Its adoption, pioneered in finance, has nonetheless been slow in macroeconomics because of two main limitations: the lack of a shared…
DiffResearch: Accelerating Deep Research with Diffusion Language Models
AI summaryDiffResearch replaces autoregressive backbones with diffusion language models in a multi‑agent research system, cutting end‑to‑end latency by ~6× while keeping quality. The dFast decomposition mode halves planning time without quality loss and tops benchmark scores.
Simply Logical -- The First Three Decades
AI summaryThe paper traces the evolution of the Prolog textbook “Simply Logical – Intelligent Reasoning by Example” from a 1994 diskette to an interactive online edition, showing how web technologies, authoring toolkits, and pedagogical philosophy combine to enhance Prolog education.
D-Skill: Distilling Teacher Knowledge through Agent Skills
AI summaryD‑Skill distills a large teacher’s behavior into an interpretable natural‑language skill from offline rollouts, closing 75‑95% of the capability gap without weight updates. It matches or beats one‑epoch SFT and reduces inference cost 5.7‑8.3×.
Back Up Before You Regret: Measuring Backup Awareness and Automating Backup in Coding Agents
AI summaryCoding agents lack backup awareness, missing state‑preservation before regretful actions. Introducing RegretBench and BackUpper, the plug‑in achieves 92.2% state similarity for all regret moments at $0.014 per task and 84.5% F1 in user‑preferred backups.
StyleStream 2.0: Fast and Controllable Streaming Voice Style Conversion
Voice style conversion (VSC) aims to transform a source utterance to match the timbre, accent, and emotion of a target voice while preserving linguistic content. StyleStream 1.0 introduced the first streamable zero-shot VSC system with state-of-the-art conversion quality, but…
Self-Improving Agents with World Model Predictive Paths
AI summaryThe paper introduces a test-time framework that uses a world model and Model Predictive Path Integral control to generate multiple imagined action outcomes and weight them by task progress and risk, while an experience memory refines both candidate generation and observation prediction. Experiments on ALFWorld and ScienceWorld show that this self‑improving, predictive‑control approach outperforms baselines across three LLM backbones.
HyperDrone: Simulation-to-Reality Transfer at Millions of Frames per Second
We present HyperDrone, an extremely fast quadrotor simulator focused on high-performance batch rendering that is more than 20x faster than existing state-of-the-art batch quadrotor simulators. HyperDrone combines highly optimized multirotor dynamics with HyperLux, a new…
SLoSa: Small, Low-Power, and Safe Foundation Models for Robotic Task Planning
Safe robotic task planning requires planners that can reliably achieve task goals while satisfying safety constraints. Classical planners search over explicit symbolic task representations and can provide formal guarantees with respect to encoded safety constraints, but require…
Can Quantum Layers offer Performance and Security Advantages in Vision-Language Models?
As quantum computing moves from the noisy intermediate-scale quantum (NISQ) era toward early fault-tolerant quantum computing (EFTQC), quantum-enhanced AI is shifting from a conceptual idea to experimental research. Quantum components are expected to extend the capabilities of…
Black-Box Red-Teaming of Multi-Agent Systems via Reinforcement Learning
Large language model (LLM) agents are increasingly deployed in multi-agent systems (MAS) to accomplish complex tasks. Prior black-box red-teaming attacks mainly focus on single agents, but we find that these methods are far less effective in MAS, where multiple sub-agents may…
Unit-Aware Differentially Private Quantum Machine Learning for Clinical and Wearable Health Data
AI summaryUnit-aware differentially private quantum models, with fewer trainable parameters than classical networks, were tested on three clinical and one wearable cohort. Across privacy budgets, the quantum model showed a modest ROC‑AUC improvement (up to 0.104) but no statistically significant advantage, though it remains viable on NISQ hardware.
From Weather to Enhanced Weathering: Weather Foundation Models for Carbon Dioxide Removal at Operational Scale
AI summaryEvaluates five weather Foundation Models on 5,192 soil chronosequences for Enhanced Rock Weathering. TerraNova‑ERW outperforms others, correctly ranking unseen fields and recovering 60% of top‑removal fields with just two samples, halving needed samples and cutting dense field sampling by up to 98%.
SSMCGM-Stream: A Streaming State-Space Model for Personalized Diabetes Management
AI summarySSMCGM‑Stream is a streaming state‑space model for glucose forecasting that updates a personalized hidden state with each reading, yielding constant‑time predictions, physiologically structured counterfactuals, and interpretable outputs. It matches neural baselines on the T1DEXI cohort while drastically reducing computation for on‑device use.
Behavior Beyond Text: Evaluating Refusal Transfer Failure in LLM Agents
AI summaryThe paper shows that LLMs’ refusal to harmful requests in text does not reliably carry over to tool‑calling contexts: 10–52 % of refused prompts still trigger unsafe tool use. An identified “refusal direction” predicts unsafe calls (AUC 0.76–0.86) but patching it only partly mitigates the risk.
PerMIT: Securing Persistent Memory for LLM Agents via Machine-Checked Integrity Theorems and Scalable Community Trust
AI summaryPerMIT protects persistent memory for LLM agents by enforcing a gate that logs and verifies only user-typed inputs, proving two integrity theorems in Lean 4. Experiments show zero attacks and community‑verified memory that scales without increasing per‑user verification effort.
AgentPolicyForge: From Executable Environments to Policy Benchmarks and Adaptive Guards
AI summaryAgentPolicyForge automatically generates policy benchmarks by executing tool calls in environments, then trains policy guards on balanced, verified examples. On PolicyGym, the trained guard cuts attack success by ~42% relative to policy prompting while reducing benign rejections.
Rethinking Time Series Tokenization from a Frequency Perspective
AI summaryTokenizes time series in the frequency domain to avoid spectral distortion; the resulting TimeFT tokenizer is parameter‑free, cheap, and consistently outperforms existing tokenizers on forecasting, classification, and anomaly detection.
VeriEvolve: Evaluator-First Evolution for Tool-Using Agents
A central obstacle to self-improving agents is the absence of a reliable fitness signal for realistic stateful tasks. We introduce \textsc{VeriEvolve}, an evaluator-first evolution framework that constructs this missing evaluation layer before optimizing the agent. From task…
Enhancing Cross-Grade Generalization in Banana Variety Classification via Hybrid CNN-Transformers and Knowledge Self-Distillation
AI summaryHybrid CNN‑Transformer architecture with KL‑divergence self‑distillation improves cross‑grade banana variety classification; achieves 81.25% mean accuracy on 31,678‑image dataset, outperforming pure CNN and ViT baselines, and enhances feature transferability and attention localization.
An Evidence-Driven Multi-Agent Compound AI System for Head and Neck Proton Radiotherapy Planning
AI summaryThe paper presents a compound AI system combining PPO‑based DRL planning and an LLM (Qwen3‑14B) for conversational refinement of head‑and‑neck proton PBS plans. Post‑trained Qwen3‑14B outperformed the base LLM, improving user‑prioritized objectives while preserving trade‑offs, leading to more clinically appropriate plans.
GenXR-Bench: Benchmarking Coding Agents for Interactive XR Development
Coding agents are rapidly advancing from single-file generation to long-horizon repository-scale engineering, yet the evaluation of their capabilities to author interactive extended reality (XR) applications remains an open challenge. Unlike traditional software, XR applications…
Inheritance as an Admission Test for LLM Prompt Merge Operators
Large language model (LLM) agents are often improved by iterating until a fixed verifier is done. Once that verifier saturates, the loop has nothing left to climb. We instead score agents against the current population, and a frozen panel records whether anyone got better. The…
PHD-Hand: A Printable Humanoid Dexterous Hand with Linkage-Driven 3-DOA Digits
Currently, dexterous humanoid robot hands are widely considered the final hardware bottleneck that hampers generalist robots from physically assisting people at home and work. As conventional tendon-driven designs suffer from high cost, demanding maintenance, poor reliability,…
GRAIL: From Saliency to Semantics in Time-Series Explanation
AI summaryGRAIL explains time‑series models by treating physiologically meaningful intervals as players, using on‑manifold counterfactuals and Owen values to attribute predictions to both signal segments and static covariates. Experiments show its explanations are more faithful, robust, and interpretable than prior fixed‑window methods.
HumanCentric-Scenes-AI: A Human-Curated Multimodal Dataset of AI-Generated Human-Centric Scenes Across Surveillance, Occupational Portraits, AI-Edited Images, and Short-Form Video
AI summaryHumanCentric‑Scenes‑AI is a 296‑image/video dataset of AI‑generated human scenes, split into surveillance, occupational portraits, edited photos, and short videos, with prompt, metadata, and quality reports. It enables perception studies, AI‑detection calibration, and bias analysis.
AutoFraming: Synthesizing Dynamic Frames of C Code with LLMs
AI summaryAutoFraming uses CPGs to guide an LLM in generating ACSL frame conditions for C code, achieving 92.9% success on 84 real‑world functions— a 179% improvement over prior methods—and boosting overall ACSL synthesis performance by up to 61.9%.
Our Own Jev: Typed Decisions in One Forward Pass, from an In-House Judge to Open LoRA Models
AI summaryTyped decisions predict option probabilities in one forward pass by reading logits or hidden states, replacing expensive text generation. The paper shows a LoRA‑adapted 35B model reaches 79.3% on JevBench, matching commercial performance, while a 1B system triages 148K coding sessions cost‑effectively.
Does Every Instance Count Equally? Order Effects in LLM Multi-Instance Processing
AI summaryThe paper tests 15 large language models on a multi‑instance task where each instance is labeled and the model must pick the majority label, showing that prediction accuracy varies dramatically (up to 90 % difference) with the order of input instances.
Distilling Multi-Teacher Scoring Principles for Unsupervised Time Series Anomaly Detection
AI summaryAnoMix distills multiple unsupervised time‑series anomaly detectors into a single lightweight student model using teacher score normalization and a CoRank ranking objective, yielding superior anomaly detection across seven benchmarks with <80k parameters. Its zero‑shot variant even surpasses large pretrained foundation models.
A Large-Scale Evaluation of Role-Enhanced Zero-Shot Chain-of-Thought Prompting
This paper proposes Role-Enhanced Zero-Shot Chain-of-Thought (RE-ZS-CoT), a prompting framework that systematically integrates domain-specific expert roles to enhance reasoning in large language models (LLMs) without task-specific examples. We assess two versions of our…
Value-Rectified Distillation for Flow-based Offline Reinforcement Learning
AI summaryValue-Rectified Distillation reformulates flow-based policy distillation as a value-guided generative alignment, decoupling behavioral modes to keep high-value actions. It improves offline RL on multi-modal datasets, outperforming baselines on D4RL, OGBench, and Minari.
ReSeg: Retraining-free segmentation of abdominal organs via generalizable semantic representation learning
AI summaryReSeg trains a generalizable semantic representation network on synthetic data and then fine‑tunes an encoder‑decoder for each organ, achieving zero‑shot abdominal segmentation. It outperforms SOTA across CT, MRI, and PET datasets without using any real images.
Context-Augmented, Session-Calibrated Blood Pressure Tracking from Photoplethysmography
AI summaryContext‐augmented wearable BP tracking using PPG achieved MAEs of 7.70/6.14 mmHg for systolic/diastolic pressure, with context adding only ~0.4 mmHg improvement. Additional sensor history or pulse‑rate fusion yielded minimal gains (≈0.06 mmHg) and did not significantly reduce DBP error.
Same Answer, Different Reasons: Trace-Level Divergence in Multilingual Large Language Models
AI summaryThe paper develops a trace‑level divergence framework to compare reasoning traces of multilingual LLMs, measuring structural similarity, order consistency, and holistic similarity. It shows that holistic similarity overestimates agreement, with 52–67% of high‑scoring examples lacking structural alignment.
Agentic Music Post-Production
We introduce MixAgent, an agentic pipeline for music post-production that follows the workflow of audio engineers to build interpretable and editable mixing plans guided by a reference song. A grouping agent first routes related input tracks, such as drum recordings, into…
UNIVERS: Learning UNIversal Voxel-wise Embedding via Randomized Synthesis for generalizable one-shot landmark localization
AI summaryThe paper introduces UNIVERS, an end-to-end framework for one-shot landmark localization that learns universal voxel-wise embeddings using randomized shape and appearance synthesis. Experiments on multi-modality adrenal gland datasets show that UNIVERS outperforms existing state-of-the-art methods, especially in cross-modality scenarios.
Dual-Channel Contrastive Encoding of Sperm Whale Coda
Sperm whales (Physeter macrocephalus) communicate through short click sequences called codas, whose structure includes the rhythmic pattern of inter-click intervals (ICI) and the spectral composition of the clicks. In this paper, we propose the Dual-Channel Contrastive Encoder…
GRAM: Group-wise Rank-Aware Model Merging via Subspace Alignment
AI summaryGRAM merges fine‑tuned models by grouping tasks with aligned subspaces, then performing SVD within each group; this reduces subspace interference. It consistently outperforms prior methods on GLUE, CLIP, and medical LoRA benchmarks.
An Eco-Friendly Tourism Trip Planner Driven by Z-Number Uncertainty: Evidence from Baku, Dubai, Abu Dhabi, Tashkent, and Istanbul
Deployed tourism trip planners ingest travel times and ratings as exact numbers, although routing services return travel times that vary with the hour of departure and a platform rating may rest on a handful of reviews or on tens of thousands. We present an eco-friendly trip…
Recursive Self-Improvement on a Budget: Scoring Code Edits Before They Run
AI summaryA coding agent uses a linear readout to score scaffold edits before execution, selecting the best edit via Thompson sampling. On HumanEval+, the method achieves 87% accuracy, outperforming baseline approaches by a few percentage points.
StochIKM: Automated High-Fidelity Reconstruction of Individual Patient Data from Published Kaplan-Meier Curves
AI summaryStochIKM reconstructs individual patient data from Kaplan‑Meier plots by digitizing the image with a vision‑enabled language model and sampling datasets that satisfy explicit fidelity constraints (exact risk tables, curve accuracy, median, intervals, event counts). It consistently matches published risk tables and medians better than existing IPD‑from‑KM methods.
Mitigating Reward Hacking in Realistic Scenarios Through Representation Engineering
Reward hacking is a major problem in current production machine learning systems, resulting in costly and dangerous failures. Here we study how to monitor for and mitigate reward hacking in realistic settings, including tool calls and long contexts. We apply a dynamic…
NormTune: Norm-Guided Parameter-Efficient Fine-Tuning for Reasoning-Intensive Tasks
AI summaryNormTune adapts parameter-efficient fine-tuning by strengthening layers with low weight‑update norms, closing the gap with full fine‑tuning on reasoning tasks. It boosts LoRA, Bone, and MoRA accuracy by up to ~24%, achieving near full‑fine‑tuning performance with far fewer parameters.
LaVGuard: Selective Decoding for Contextual Value Drift
AI summaryLaVGuard treats contextual value drift as a decoding problem by maintaining a frozen estimate of values and an authorized profile. It uses a density gate and KL‑budgeted logit tilts to intervene safely, raising TPR at 5 % FPR from .58 to .72 while keeping interventions at 12 %.
GRAFT: Restoring the Missing Sampling Interface in Continuous Diffusion Language Models
AI summaryContinuous diffusion language models generate full sequences at once, which collapses the rare‑word tail. GRAFT restores missing word‑choice decisions, improving rare‑word coverage by 34–109 % and aligning overall word‑frequency profiles with human text.
Steering Black-Box Large Language Model Agents with Pitfall Anticipation World Model
AI summaryPAWM is a world model that predicts imminent pitfalls in a bounded future and activates only when a failure is likely, trained with synthetic pitfall data and co‑training with the target LLM. It boosts Claude Sonnet 4.5 by 24.3% on ScienceWorld and 7.7% on WebArena‑Lite.
From AI-Free Gates to AI-First Work: How Engineering Students in India Navigate University Education and Campus Placements Without AI Guidelines
Engineering graduates in India enter a technology sector that is being reshaped by generative AI, yet the route into that sector runs through examinations and campus placement screens that largely exclude AI. Institutions in India have an approved intake of nearly 2.9 million…
ORSMT: Unleashing the Potential of Satisfiability Modulo Theories for Semantic Verification and Repair in LLM-Based Optimization
Large language models increasingly automate optimization modeling from natural language problems, producing written mathematics, solver code and a returned report. Verifying one optimization problem across these artifacts requires semantic comparison: unequal models can attain…
Where Does Timbre Hide? Localizing Speaker Information in Speech Signals
Timbre similarity is a key evaluation target for generative speech systems, including voice conversion, speaker-cloning text-to-speech, and zero-shot speech models. When the generated voice does not match the target speaker, the problem often reflects imperfect separation of…
From Augmentation to Synthetic Evidence: Evaluating Generated Chest Radiographs at Dataset Intake
AI summaryThe paper evaluates how well automated checks can detect synthetic chest X‑ray images and finds that while quality filtering can improve acceptance rates, no strategy reliably ensures both quality and authenticity. It concludes that source provenance must be recorded separately for dataset inclusion.
AI-Based End-to-End Automotive Design and Software Development for SDVs
This paper explores the potential of Artificial Intelligence (AI) to enable end-to-end automotive design and software development for Software-Defined Vehicles (SDVs). Based on our contributions within the Central Car Server (CeCaS) and follow-up EU-funded projects, we present…
Why is AI-generated language so hard to understand?
AI summaryThe paper measures psycholinguistic difficulty in AI text and finds large-scale language models use more complex structures (long dependencies, low‑frequency n‑grams, inanimate‑subject animate‑verb) than humans or paraphrases. Fine‑tuning with a length penalty shows that higher information density drives this misalignment.
TeachBack: Learning to Teach via Bidirectional Preference Optimization for On-Policy Self-Distillation
AI summaryTeachBack learns a Teacher that also learns how to teach, using bidirectional preferences to shape prompt length and style. The method improves student performance on math and logic tasks, matching Teacher accuracy while encouraging exploration and reducing reference reliance.
How Representational Priors Shape Generalization in World Action Models
World action models (WAMs) are shown to substantially diverge under generalization performance based on their generative backbones. We investigate this gap through the representational geometry that WAMs inherit from their backbones and how policy training reorganizes it. Our…
TrainBridge: Agentic NPU Training Optimization Guided by GPU Compilation Evidence
AI summaryTrainBridge is an agentic framework that transfers optimization strategies from mature GPU compilation stacks to improve training efficiency on NPUs. It achieves an average 1.7x speedup across various deep learning workloads by using GPU execution evidence to guide NPU program rewrites.
Conditional Flow Models Convert Epistemic Uncertainty into Apparent Aleatoric Variance
AI summaryCFMs generate samples whose variance often exceeds the true conditional variance, converting epistemic model uncertainty into apparent aleatoric noise, especially when the model underfits. This excess variance decreases with more training and can be leveraged to construct high‑coverage predictive sets.
Correlated Failures of AI Approval Monitors
Before an AI agent takes a consequential action, such as paying, messaging or changing code, several monitors check it. Their verdicts are combined by aggregating scores or by counting approvals, whatever inputs the monitors share. Monitors that share a base model or a poisoned…
Resuming Vision-Language Inference across Onboard Interruptions
Executing vision-language models (VLMs) aboard low-Earth orbit satellites enables flexible processing of observation data without first downlinking it to the ground. On commercial off-the-shelf payloads, however, a request may span multiple execution windows: power and thermal…
LTEM: LLM-Friendly Table-Enhanced Markdown for Table Understanding and Generation
AI summaryLTEM adds minimal span and header markers to Markdown, enabling LLMs to parse and generate complex tables with merged cells and multi‑level headers. It matches HTML accuracy while using fewer tokens and yielding far fewer invalid outputs, outperforming LaTeX in generation quality.
Huayu-Global: Skillful real-time quasi-global precipitation retrieval
AI summaryHuayu-Global fuses three geostationary sensor streams from five satellites via deep learning to produce quasi‑global precipitation fields every 30 min, matching delayed‑product accuracy and improving real‑time event detection and rainfall estimates for flood forecasting and water management.
When Multiple Choices Are Correct: Asymmetric Probability Allocation in Jev and Other Decision Models
AI summaryDecision models often misallocate probabilities when multiple answers are correct, assigning some true choices below the uniform level. Even well‑ranked models fail to identify all correct options via thresholding, showing significant calibration gaps up to 79 %.
Beyond Binary Labels: Diagnostic Evaluation of LLM Safety with Prompt-Level Rubrics
AI summaryDiagSafe introduces prompt-level rubrics for LLM safety evaluation, replacing binary labels with a 84-label safety taxonomy that scores safety and usefulness independently. It achieves 95.34% agreement with humans and 0.864 κ, revealing model-specific vulnerabilities, especially in Chinese and multi-turn contexts.
From Scores to Safe Sets: Exact Safety--Constrained Decoding for Medication Recommendation
AI summaryThis paper evaluates decoding strategies for medication recommendation under safety constraints, showing that a global mixed-integer linear programming decoder consistently achieves higher Jaccard similarity than local repair methods, especially under strict drug-drug interaction limits.
CrashGym: Measuring the Decision Value of Experimentation for LLM Agents in Counterfactual Market Crashes
AI summaryCrashGym is a deterministic market simulator that measures whether experimentation improves LLM agent decision-making during financial crashes. It finds that providing agents with laboratory access for intervention experiments reduces decision regret by 57.4 percent.
MOSAIC: An Omni-Modal Benchmark for Long-Term Memory and Cross-Modal Reasoning in Conversational Agents
AI summaryMOSAIC evaluates omni‑modal conversational agents on long‑term memory across text, image, video, and audio by 923 targeted questions. It diagnoses failures in representation, retrieval, delivery, and reading, revealing modality‑dependent losses and weak refusal behavior.
Hierarchical Masked Diffusion Language Models
AI summaryHierarchical Masked Diffusion Language Models (HMDLM) add a hierarchy level per token, using separate noise schedules to prioritize high‑priority positions during parallel generation. HMDLM cuts perplexity by over 1.5 points on LM1B, 4.4 on WikiHow, and boosts Sudoku solution accuracy by 8.6% versus prior diffusion baselines.
CRAB: Cuboid-based Representation for As-built BIM from Point Clouds
AI summaryCRAB reconstructs as-built BIM from point clouds using cuboid representations for walls, doors, windows, beams, and columns; it achieves 2.80 cm mean wall thickness error and outperforms plane-based methods in completeness, accuracy, and generalization.
ARCANA: A Benchmark for Abstraction and Analogical Reasoning
AI summaryARCANA is a new benchmark for far analogical reasoning, mapping real‑world objects to abstract grid transformations. Humans solve these tasks well, but large language models fail; the proposed ConceptSIR framework improves LLM performance on many ARCANA problems.
Flat-RLVR: Sharpness-Aware Reinforcement Learning with Verifiable Rewards for Reasoning Capacity in LLMs
AI summaryFlat-RLVR uses sharpness‑aware minimization to locate flat reward regions during LLM post‑training, reducing distribution sharpening in RLVR. Experiments show higher sample diversity with no loss in single‑attempt accuracy, improving the trade‑off between efficiency and reasoning capability.
Same Probabilities, Different Advice: Gain-Loss Framing Effects in High-Stakes LLM Decision Support
AI summaryThe paper introduces FrameBench‑HPS, a 2,880‑prompt benchmark for high‑stakes LLM advice, and finds that gain‑loss framing swaps recommendations in 32.4% of cases, largely altering uncertainty assessments; canonicalizing probabilities reduces this sensitivity.
Latent space configuration for neural network finetuning, fast ranking, and embedding dimensionality reduction
AI summaryLatent Space Configuration finetunes ArcFace‑IResNet50 embeddings to give a structured latent space, enabling O(1) candidate selection. This yields up to 3× GPU and 40% CPU speed‑ups over exhaustive search while preserving or improving rank‑N accuracy, even with dimensionality reduction.
When Are We Near Nash? Certifying Solutions in General-Sum Games
Heuristic solvers for general-sum extensive-form games produce a stream of iterates, and practice reports the NashGap of the returned profile. We ask what a solver can know about its own iterates from the regret accumulators and the terminal-distribution accumulators its updates…
LLM-Based Exogenous World Modeling for Dynamic Scheduling
AI summaryAn LLM predicts only uncertain future job arrivals in a structured world model for dynamic scheduling. The resulting Afterstate World-Model Planning algorithm reduces weighted tardiness by 6–46% compared to the best classical dispatching rules across multiple benchmarks.
YingLan: A Multimodal Ocean Foundation Model for Unified Representation Learning from Heterogeneous Observations
AI summaryYingLan learns unified ocean representations from satellite, gridded, and buoy data using physics‑aware dynamic encoding, spatiotemporal increment coupling, and evolution‑aware pretraining. It reduces 7‑day sea‑surface salinity RMSE to 0.6875 PSU, outperforming prior models by up to 12.4 % relative error.
SpatialBrain: Spatial Intelligence Growing from Experience
Spatial reasoning requires more than access to geometric tools: a reasoner must decide when direct observation is sufficient and when to reconstruct, measure, or change reference frames. We present SpatialBrain, a framework that retains experience about these decisions while…
OpenMAP-HIE: Pathology Informed Score-based Diffusion Model for Unsupervised Neonatal Hypoxic-Ischemic Encephalopathy Lesion Segmentation
Accurate localization of neonatal hypoxic--ischemic encephalopathy (HIE) lesions on apparent diffusion coefficient (ADC) images is clinically important but is limited by scarce expert annotations and substantial variation in lesion appearance. We propose OpenMAP-HIE, a…
AgentMonopoly: Behavioral Patterns and Persona Control in LLM Agents
How do LLM agents' strategies differ, and how much can instructions to adopt a persona change their behavior over a game? We study these questions with AgentMonopoly, an environment where agents communicate, buy and trade properties, and make decisions that affect later turns.…
From Prompts to Posteriors: Can LLM Agents Perform Bayesian Inference in Context?
Large language models (LLMs) are increasingly applied across the scientific discovery pipeline, from synthesizing literature to generating hypotheses to informing experimental design. A unifying principle behind these tasks is the need to continuously update beliefs as evidence…
ChronoSynth: Constraint-Aware Question Synthesis over Temporal Knowledge Graphs
Knowledge graph question generation (KGQG) generates natural-language questions from structured graph evidence and a target answer. When the input graph is temporal, a valid question must also preserve the answer field, time condition, reference event, and temporal order that…
Memory-Centric Continual Learning for AI Agents
Autonomous agents powered by foundation models can reason, plan, and use tools, but their fixed weights and context limits hinder sustained learning from experience over long-term deployment. While external memory systems attempt to bridge this gap, they mostly act as passive…
Reading the Mask: Descriptor-Grounded BI-RADS Rationale Generation for Breast MRI Segmentation
Deep learning based models accurately delineate tumours on dynamic contrast-enhanced MRI (DCE-MRI), but segmentation masks alone do not provide the rationales or reasoning used by radiologists to assess the lesion suspicion. This limits the explainability as well the clinical…
Parallelizable Probabilistic Latent Reasoning for Diverse Reasoning Trajectories
Language models relying on verbalized Chain-of-Thought (CoT) are constrained to token-by-token generation in the language space, where many tokens serve textual coherence rather than core reasoning. This entangles genuine differences in logical trajectories with superficial…
LLMSeg: Training-Free Text-Prompted Image Segmentation with B-Splines
AI summaryLLMSeg turns a vision‑capable LLM into a training‑free, text‑prompted image segmenter by iteratively localizing instances, tracing their borders with B‑splines, and refining until masks stabilize. It matches or outperforms SAM on ReasonSeg, RefCOCO, and LVIS without any segmentation training.
Beyond Small Talk: Modeling Disclosure Trajectories in Human-AI Conversation
AI summaryTrajectory‑based Disclosure Evaluation (TraDE) measures how prior disclosure history influences AI responses to the same current disclosure level. Across models, TraDE reveals that AI often accepts or challenges user framing differently depending on the previous trajectory, diverging most when users retract disclosure.
Solving Go under Positional Superko: Sound Proof Transfer and New Board Values
Positional superko (PSK) forbids recreating any earlier whole-board position. It is part of the Chinese rules as written and of the rules used in computer Go. Yet almost all solved Go boards were computed under a simpler basic-ko rule set. PSK makes solving hard: legality…
MERP: A Latent-Bottleneck Framework for Robust Multi-Modal Multi-Label Emotion Recognition
Multimodal multi-label emotion recognition is important for affect-aware systems, but models trained with complete text, visual, and acoustic streams often degrade sharply when some streams are unavailable at deployment. We address this train-test gap by learning a compact…
Events, Not Pixels: Where World Models Get the Return Wrong
World models are increasingly used for planning, where a planner ranks candidate actions by returns imagined by the model. However, these models are judged mainly by how well they render frames, even though a visually faithful rollout can still assign the wrong return. To…
RegoMender: A Multi-agent Framework for Automated Rego Policy Repair
Policy-as-Code (PaC) has emerged as a foundational paradigm in cloud-native systems. However, repairing Rego policies remains challenging due to their strict syntax, declarative semantics, and specific verification requirements. Existing automated program repair (APR) methods…
Harness Scaling for Generalizable Agentic Reinforcement Learning
Agent harnesses have become a central component of modern agentic systems, shaping how models use tools, manage context, and coordinate long-horizon interaction. Realizing the benefits of these harnesses depends on how effectively the underlying model can make use of them,…
PreCoMem: Predictive Cognitive Memory for Self-Evolving Long-Term Dialogue Agents
AI summaryPreCoMem introduces a predictive coding–based memory consolidation for long‑term dialogue agents, using Effective Surprisal and a two‑threshold update gate to maintain, profile, or correct user‑state beliefs, achieving state‑of‑the‑art accuracy on LoCoMo, LongMemEval‑S, PersonaMem‑v2, and DiMoBench.
OmniSimulator: Aligning Small Language Models for Authentic Heterogeneous Behavior Modeling
AI summaryOmniSimulator trains small language models with a structured, behavior-oriented reasoning schema and teacher-generated traces, using 9,881 curated user‑interaction examples. It boosts Qwen3‑8B’s simulation accuracy by 34.33%, matching Claude‑Sonnet‑4.5 while enabling self‑hosted deployment.
Generative Modeling via Nonlinear Schrödinger Bridges and Branching Diffusions
AI summaryGenerative modeling is built on forward‑backward PDEs from mean‑field games, yielding a nonlinear Schrödinger bridge system (SBNL). Using branching diffusions, the method solves this efficiently, scaling to high dimensions and recovering classical, nonlinear, and score‑based generative tasks.
RubricOpt: Self-Revision of Semantic Rubrics for Preference Optimization
A model's failures can reveal gaps in the behavioral criteria guiding its training. We introduce RubricOpt, a framework that lets a policy revise these criteria within preference optimization. The policy diagnoses its own rubric-free outputs, retains requirements already covered…
An Exact 93-Move Solution to a Nonlocal Crystal-Exchange Puzzle
We determine the minimum number of directional commands needed to solve alevel of Order of the Sinking Star, modeled as a crystal-exchangepuzzle on an 11*11 board. A command swapsthe player with the nearest crystal in the selected row or column, regardlessof distance; only an…
A Strong MoE Is All You Need for Process Supervision
Process reward models (PRMs) provide fine-grained supervision for multi-step reasoning, yet obtaining reliable process labels remains a central bottleneck. Existing approaches typically rely on human annotation, rollout-based estimation, or auxiliary verification. In this work,…
Parallel Thinking for Large Language Model with Path Diversity
AI summaryThis paper introduces Diverse Parallel Thinking, a reinforcement learning approach that prevents mode collapse in large language models by penalizing prefix similarity across reasoning paths. Experiments show that encouraging this path diversity significantly improves reasoning performance over existing baselines.
Rethinking Agentic Model Orchestration for Time Series Forecasting
AI summaryFairAMO is a controlled evaluation protocol for agentic model orchestration in time series forecasting. It finds that orchestration gains are modest, depend on model complementarity rather than quantity, and saturate quickly with historical data.
Keeping Time vs. Understanding Time: Neither Alone Explains Temporal Reasoning in LLMs
AI summaryThis paper tests the assumption that temporal reasoning splits into arithmetic and semantic modes. It shows finetuning transfers poorly within modes, and that benchmark scores are driven by factual knowledge rather than by true arithmetic or semantic reasoning.
Mitigating Query Shortcuts in Chain-of-Thought with Context-Aware Decoding
While Chain-of-Thought (CoT) prompting improves the reasoning performance of large language models (LLMs), recent studies have shown that final answers do not always depend on the generated rationales. In particular, LLMs may rely on query-specific shortcuts rather than the…
Are Joint-Embedding Predictive Objectives Effective for Longitudinal EHRs?
Masked language modelling (MLM) is the dominant pretraining objective for transformer encoders applied to longitudinal electronic health records (EHRs). However, reconstructing individual events may overemphasise noisy or weakly informative observations and does not explicitly…
Block Diffusion for Longitudinal EHR: Accuracy, Efficiency, and Representation Trade-offs
AI summaryBlock diffusion (BD) on MIMIC‑IV matches autoregressive models for clinical prediction while speeding inference. A B=8 BD runs 1.43× faster, retains comparable representation quality, and shows diminishing returns with more denoising steps, suggesting BD is a computationally attractive alternative.
Stochastic Pruning of the DPO Computation Graph: 1.4x Faster Training and Lower Peak Memory
AI summaryThe paper presents a stochastic pruning framework for DPO training that samples scoring positions, execution prefixes, and backpropagation branches to skip computation. It achieves a 1.4× faster step time, 30% total training speedup, and lower memory at only slight accuracy loss.
The Reflexivity Threshold: A Phase Transition for Multi-Agent Performative Prediction
Many learning systems are performative: once deployed, their predictions change the data distribution on which future models are trained. In multi-agent settings, this feedback is routed through a network of interacting learners, so stability depends not only on the strength of…
Improved Training of Strong Stochastic Flow Maps
Stochastic flow maps promise few-step sampling of stochastic differential equations (SDEs) by learning to directly jump between times along the solution path, yet training these models to learn this solution map remains difficult. The primary obstacle is the Brownian path…
Overview of CHIP 2026 Shared Task 1: Patient-Specific Phenotype Concept Recognition in Full-Text Biomedical Literature
Automatic phenotype information extraction plays an important role in rare disease diagnosis, genotype-phenotype analysis, and precision medicine. However, accurately identifying phenotype concepts and determining their associations with specific patients remain challenging,…
An interpretable deep learning method for early hepatocellular carcinoma diagnosis on multiphase CT
Early diagnosis of hepatocellular carcinoma is critical to preserve access to curative treatment options. The Liver Imaging Reporting and Data System (LI-RADS) provides a standardized framework for non-invasive diagnosis on multiphase computed tomography, based on major visual…
FogTTA: Online Test-Time Adaptation for Robust Transformer-based Object Detection in Foggy Weather
AI summaryFogTTA adapts a Transformer-based detector online to foggy conditions by using the deployed model as a teacher to generate pseudo labels, refining them, and training a student with Varifocal loss; it outperforms prior methods while remaining real‑time.
Formalizing Cooperation and Competition: An AI-Assisted Mathematical Addendum to a 2018 Planning Thesis
We present an AI-assisted mathematical addendum to Strategic Trip Planning: Striking a Balance Between Competition and Cooperation (Amar, 2018). We reconstruct selected relationships between strategic choice, cooperative stability, participant approval and enforcement. The…
From Association to Decision Value: Testing Paired Instruction Dependence for VLA Corpus Selection
Choosing among robot demonstration corpora is a ranking problem: an upstream diagnostic must preserve which candidates produce better downstream policies, not merely correlate with average performance. We study paired instruction dependence, the rate at which a demonstrator…
Lineage-Aware Search Control for LLM-Based Evolutionary Factor Mining
Formulaic alpha factor mining searches for symbolic expressions that transform market data into predictive signals. Recent LLM-based evolutionary factor mining methods use large language models as semantic variation operators. Under the conventional LLM-EA paradigm, stochastic…
Conflicted Records: Typed Adversarial Contracts for Auditable LLM Output Quality Enforcement
LLM-generated records increasingly populate production information systems where they are treated as data assets subject to the same quality expectations as records from operational databases. A single such record can simultaneously satisfy and violate a quality rule: a grounded…
Toward Recursive Self-Improvement in Physical AI
Recursive self-improvement (RSI) describes a system that improves not only its behavior but also the process by which it improves. Recent digital agents test this idea by rewriting their own code and checking each version with executable tests. Bringing these mechanisms into the…
TraceForge: Distilling with Your Own Thinking
Using an LLM to generate training data for another is among the most general ways to train one, and it remains common practice. Some frontier teams have closed that route for their own models: what comes back is an answer, or a short summary, without the detailed reasoning trace…
Scaling Multilingual Pretraining of Pixel Language Models
Pixel language models process text as rendered images instead of a finite vocabulary, which gives them strong cross-lingual transfer abilities. However, not much is known about how these abilities are influenced by larger-scale pretraining. In this paper, we train a family of…
Bio-Inspired Auditory Processing: Coupling Parallel Resonate-and-Fire Dynamics with Ternary Recurrent Networks for Speech Recognition
Spiking Neural Networks (SNNs) promise energy-efficient inference through sparse, event-driven computation, yet remain constrained on continuous, context-dependent tasks such as automatic speech recognition (ASR). A key bottleneck is the Leaky Integrate-and-Fire (LIF) neuron,…
Retrieval Augmented Generation Compositionality
Retrieval Augmented Generation (RAG) grounds large language models in external knowledge. Modern pipelines combine embedding, chunking, retrieval, reranking, query augmentation, context compression, and agentic verification, yet these techniques are typically evaluated one at a…
Structure-Aware Message Passing on Petri Nets through Directed Walks and Incidence Diffusion
A common application of the process mining discipline is to connect observed event data with Petri nets, which are bipartite, directed graphical process models with rich formal underpinnings. Recent work learns vector representations of Petri nets to retrieve process models and…
SAKURA: Serving Attention Key-Value State Under Rate Adaptation for Private Image Editing on Phones
Diffusion transformers are widely used in image-editing services. To preserve both efficiency and privacy, existing systems commonly rely on cloud-device collaboration. Keeping request-specific inference on the device avoids sending personal images and prompts to the cloud.…
Metamodel-Guided LLM Decoding for Consistent DSL Generation
Large language models (LLMs) are increasingly used to generate software engineering artifacts written in domain-specific languages (DSLs). Correct DSL instances usually must comply with not only the concrete syntax but also the abstract syntax defined by a metamodel augmented…
Where Do LLMs Learn to Recommend AI?
AI summaryLarge language models already recommend AI topics in about 27 % of responses, rising to 59 % after instruction tuning. Fine‑tuning and prompt phrasing can raise or lower this rate, showing the tendency originates in pretraining and is modulated by training data and query style.
Real PMIDs, Wrong Papers: Harness Design for Biomedical Claim-Verification Agents
Literature-searching language agents are becoming a standard component of scientific assistants for the life sciences, yet there is little controlled evidence on which parts of their harness matter. We run a controlled ablation on biomedical claim verification (SciFact) with the…
COMET: A coordinated multi-layer runtime monitoring framework for human-supervised AI-enabled safety systems
Incorporating AI systems into safety systems requires monitoring to avoid silent faults and ensure safe and reliable operation. Single monitors often do not provide actionable information for human supervisors, who act as the system's adaptive controller. We present COMET, a…
UserSWE: Evaluating Coding Agents in Multi-Turn Conversations with Dynamic Users
As coding agents become more capable, developers increasingly use them to fix bugs through multi-turn interactions. Existing benchmarks do not adequately cover this scenario. Most present a complete issue report in a single turn without interaction. Their evaluation can produce…
MemoryTax: The Bitter Lesson of Agent Memory
AI summaryMemoryTax tests different memory schemes on a 419‑turn LoCoMo conversation, finding that designed memory (Mem0, Zep) only benefits weak models and can be more costly than full context; a strong coding agent (Pi on Claude Opus 5.5) achieves 99% accuracy at <50% of full‑context cost.
Robot Generalization Beyond Volume: The Diversity Path via Zero-Shot Capabilities
General-purpose robots must handle tasks, objects and environments they were never trained on, yet zero-shot generalization remains one of the least solved problems in physical AI. In digital AI, zero-shot capability grew with the volume and, above all, the diversity of training…
S-EDL: Eliciting Self-Evidence from Sequence Likelihoods for Semantic Calibration of LLMs
Calibrating large language models (LLMs) in open-ended generation is uniquely challenging because many distinct token sequences can express the same underlying meaning. Prior calibration techniques target either fixed-set classification confidence or the binary correctness of…
Broadband Mid-Wave Infrared Imaging Enabled by All-Silicon Hyperboloid Metalens
The mid-wave infrared (MWIR, 3-5 microns) is indispensable for thermal sensing, gas detection, long-range imaging, and free-space optical communications. Meta-optics offer a compact, lightweight, and potentially low-cost alternative to conventional refractive optics, making them…
Every Part Has Its Role: Part-of-Speech Effects in LLM Post-Training
Large language models are typically trained with next-token prediction (NTP), which rewards predicting specific tokens rather than explicitly accounting for the semantic similarities between different words or expressions. Prior work has addressed this limitation by augmenting…
PAUSE: Search-Free Steering of Coding Agents Before Their First Source-Code Edit
Coding agents can invest substantial effort in repository exploration and testing yet still produce unsuccessful patches. Existing online steering methods review trajectories periodically or invoke an advisor when a monitor detects behavioral problems; frequent reviews are…
Automated Materials Data and Knowledge Extraction using Machine Learning and Large Language Models
Over the past decade, substantial efforts have been devoted to developing machine learning (ML) methods for extracting materials data and, where possible, materials knowledge directly from the scientific literature. This review traces the evolution of these methods from…
Asking a Multimodal LLM Whether Someone Fell: Zero-Shot Screening of Low-Resolution Thermal Video in a Laboratory and an Occupied Home
Thermal cameras can watch over older people living alone without capturing faces, and multimodal large language models (LLMs) can judge a short video without task-specific training. We evaluate a production multimodal LLM, Gemini 3.8 Flash at temperature 0, as the decision stage…
Unintended Side Effects of Value Training in LLMs
AI summaryValue‑training LLMs on Hofstede dimensions produces spill‑over effects on unrelated values, risk attitudes, social cognition, and safety. The study shows that post‑training alignment needs broader behavioral checks, suggesting careful placement of value fine‑tuning in the pipeline.
Privacy in Context Is Being Overlooked in the Age of General-Purpose LLMs
AI summaryThe paper critiques the trend toward a single, universal prompt‑privacy framework for LLMs, arguing that privacy must remain context‑specific. It recommends deploying LLMs per application, evaluating privacy in each context, and improving privacy literacy.
Your Benchmark Is an Empirical Measure Over Difficulty
AI summaryThe paper treats a benchmark as an empirical difficulty distribution, revealing two failure modes: measure mismatch and coverage gaps. It introduces diagnostic metrics and the BenchFill algorithm that rewrites items to reduce coverage gaps, showing many existing benchmarks overweight or miss difficulty regions.
Graph Topology Meets Energy-Driven Pathfinding: Preserving Contextual Consistency in Retrieval-Augmented Generation
Retrieval-augmented generation (RAG) frequently integrates knowledge graphs (KGs) for complex multi-hop question answering. However, their performance is increasingly bottlenecked by the knowledge structuring and reasoning processes. (i) Severe semantic loss: Distilling…
ToM2Vec: Continuous Theory-of-Mind Tokens for Multi-Agent Coordination
Forming an accurate theory-of-mind (ToM) is key to successful multi-agent collaboration and requires reasoning under uncertainty to capture multiple possible partner states. This is hard to do in the text representation space native to large language model (LLM) agents, which…
Commit or Abstain? System One Decision Models, LLMs, and Missing Legal Authority
AI summarySystem‑One decision models use LLMs to output calibrated probabilities instead of text, allowing them to abstain when no legal holding applies. In CaseHOLD, a decision model matches the best generative LLM (≈88% accuracy) while cutting cost 55× and reducing wrong citations.
CulturalRiddles: A Multicultural Benchmark of Riddles
As LLMs expand their language coverage, evaluating how well they understand and represent diverse cultures becomes increasingly important. However, many benchmarks remain English-centric, with limited cultural coverage and a focus on factual knowledge over its application in…
FaultGene: LLM-Assisted Fault-Augmented Test Generation for Hardware-in-the-Loop Validation of Autonomous Driving Systems
Reliable validation of autonomous driving cyber-physical systems requires testing not only nominal driving behavior but also failures arising from faults in sensing and embedded computation. Physical road testing provides realistic system behavior but is costly and difficult to…
TIDE: Tree-Induced Fine-Grained Diagnostic Evaluation from Model Behaviors
Benchmark performance is the dominant standard for comparing large language models (LLMs), typically summarized by a single aggregate score, such as accuracy, pass rate, or rubric score. This aggregate score provides limited information about a model's underlying capabilities,…
HiddenLens: A Multi-Layer LLM Framework for Detecting and Resolving Hidden Assumptions in Software Requirements
Natural-language software requirements often omit assumptions about performance, operating conditions, user behaviour, and domain constraints, affecting architecture, testing, and deployment. If undiscovered until implementation, resolving them is far more costly than addressing…
Every Alarm Has a Cause: An Open Generator of Telecom Alarm Storms with Exact Root Causes, Bound to NORIA-O, RFC 8632 and TMF642
Alarm root-cause analysis is studied on data that cannot be shared. Operator logs are confidential, so what the field publishes is a causal graph between alarm types rather than per-incident causes, a fixed scenario set, gated, or never released. Two recent benchmarks carry…
Building Auditable AI Text Detectors with Sparse Autoencoders
AI-generated text detectors are increasingly deployed in consequential settings---peer review, education, admissions, journalism, and content moderation---where their outputs directly affect individuals. Yet the strongest detectors are fine-tuned neural classifiers whose…
Agent Behavior Maps: Locating Agents by What They Would Do, Without Running Them
AI summaryAgent Behavior Maps encode LLM-harness combinations via centered action log‑likelihoods on shared reference trajectories, enabling performance prediction without new rollouts. Across 100 LLMs and four environments, map distances rank‑order trajectory KL, revealing behavioral similarities that persist after controlling for success rates and model size.
Code Comprehensor: Hierarchical Documentation and Cross-Modal Retrieval for Large Repositories
Understanding large, complex, and rapidly evolving code bases remains a core challenge for software engineering at scale, hindering onboarding, maintenance, refactoring, and impact analysis. Providing this information to a Large-Language Model (LLM) is another hurdle, as the LLM…
GuideEvolve: Self-Evolving Agent Framework with Expert Guidance and 3Ts Mechanisms
This study introduces GuideEvolve, a self-evolving agent framework incorporating expert guidance and three autonomous regulatory mechanisms, collectively termed the 3Ts: Think Fast and Slow, Throwback, and Termination by choice. The framework addresses critical limitations in…
Stationarity Constraints Improve Both Energies and Densities of Machine-Learned Exchange-Correlation Functionals
Empirical exchange-correlation functionals are fitted on fixed orbitals but used self-consistently, so better energies often come with worse densities. We show that the Brillouin residual of a trained functional on the Kohn-Sham orbitals of an accurate density controls, to first…
What the Author Wrote, What the Agent Reads: Silent Cue Loss in Enterprise Document Ingestion
Enterprise agents act on documents. Before an agent can follow a policy, answer a compliance question or file a ticket, some ingestion layer must turn a Word file or PDF into tokens - and that layer is where an agent deployed in the wild silently loses information. We show that…
Robust Soft Skill Adaptation via Behavioral Localization in Tool-Using Agents
AI summaryThe paper localizes where a text skill changes token distributions, then uses only the top 10% of trajectory positions to train a short soft prompt. This yields 57.38% success on SpreadsheetBench, outperforming full‑skill and OPCD baselines.
HGM-Dual: Lineage-Aware Guided Agent Harness Optimization for Long-Horizon Planning
Harness optimization, which improves an agent's prompts, tools, and workflows, is an increasingly effective way to enhance agents built on frozen or closed-source LLMs. Leading approaches such as the Darwin--Godel Machine and the Huxley--Godel Machine (HGM) rely on recursive…
SP-HLM: Sparse-Pose-Driven Motion Generation and Whole-Body Control for Humanoid Loco-Manipulation
Humanoid loco-manipulation requires an effective interface between high-level task policies and whole-body control. Existing interfaces typically rely on a single action representation for both task intent and whole-body coordination. This may burden the high-level policy or…
Which Differences Matter? Controlled Tests of Rule Transfer, Joint Reliability, and Information Use in Decision Models
Decision models map stated rules and observed states to discrete answers. Aggregate accuracy leaves unresolved which rule and information changes guide those answers. We connect controlled input changes to transfer, joint reliability, and directed information use through…
SWE-DIVE: Multi-Turn Coding Agent Evaluation with Dynamic User Simulation
Coding agents are typically evaluated using one-shot, well-formulated prompts. Benchmarks such as SWE-bench use existing GitHub issues as input for a coding agent and evaluate whether the agent successfully solves the task. However, in real-world use cases, prompts are often not…
From Reasoning to Action: Benchmarking Frontier AI Models on LIBERO-10 Robot Manipulation Tasks
Recent advances in frontier AI models have demonstrated remarkable capabilities in visual understanding, reasoning, planning, and tool use, raising the possibility of using general-purpose AI models for robot manipulation without task-specific training. However, their…
Charts as Epistemic Actions for Experience-Driven Agentic Data Analysis
AI summaryData agents can now generate and use charts as reasoning steps. The proposed method teaches agents to choose when to produce charts versus text, improving analytical accuracy and visual evidence quality on several benchmarks without fine‑tuning model weights.
Parallel AI Agent Collaboration as Graph-Structured Distributed Block Optimization
Parallel AI agents can produce useful local edits yet degrade a shared solution when their edits are combined. We model this coordination problem as discrete block optimization with an exactly evaluable objective and a known dependency graph. Independent sets certify compatible…
AtomicRL: Atomic Reinforcement Learning for Long-Horizon Robotic Manipulation
Vision-language-action (VLA) models have shown promise in robotic manipulation, but imitation learning (IL) alone often struggles to master complex long-horizon skills. Reinforcement learning (RL) offers an effective avenue for further improving VLA policies. However, existing…
Differentially Private Sparse Reward Estimation with Preference Feedback
Reinforcement Learning with Human Feedback (RLHF) has become a central paradigm for aligning Large Language Models with human values. However, the reliance on sensitive preference data necessitates rigorous differential privacy (DP) guarantees to prevent user information…
BatchBench: Benchmarking AI Robustness to Batch Effects and Domain Shift in Biomedical Images - A JUMP Cell Painting Leaderboard
Despite recent advances, current AI systems remain highly susceptible to domain shifts, particularly in biomedical imaging, where images are usually acquired under substantially different experimental conditions. Although recent domain adaptation methods have made substantial…
Asymmetric On-Policy Distillation Learns In-Context Exploration
Sequential decision-making under partial observability is difficult because agents must explore to infer hidden, task-relevant information before acting effectively. Learning such information-gathering behaviors tabula rasa via online reinforcement learning is…
Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use
Multimodal agents (MMAs) sustain interaction through reasoning, memory, tools, and feedback, most visibly as computer-use agents, while robotic systems (RSs) couple sensing and actuation under physical dynamics. We define multimodal embodied agents (MMEAs) as goal-directed…
Quant-Arena: Benchmarking LLM Agents for Quantitative Trading Research
Quantitative research is an iterative loop: a researcher proposes a market hypothesis, implements it as an executable signal, tests it on historical data, and decides what to keep. Existing benchmarks score formulas, code, or individual trading decisions, leaving open whether…
The End of Consensus Research?
Consensus is a fundamental problem in distributed computing and at the core of state machine replication (SMR). Despite decades of active human research, discovering practical low-latency solutions to the problem is still open, as it requires combining protocol knowledge,…
Environment Duels: A Microcosm of LLM Co-Evolution
Through a tangle of real-world interactions, from synthetic training pipelines to data shared on the open web, frontier LLMs absorb what their predecessors and competitors create. How well does each model play its role in this ecosystem, as a creator and as a solver? We…
PhysiCalWorld: In-Context Physics Calibration for World Models
Existing world models typically perceive a scene through current visual frame or a narrow history window. However, these observations may not reveal the physics properties that determine transition dynamics, such as an object's mass distribution, joint damping, or surface…
Think Harder Where It Matters: Item-Level Adaptive Recursion for Sequential Recommendation
Existing sequential recommenders usually allocate a fixed computational budget to every item in a user's interaction history, regardless of the item's contribution to the final prediction. We challenge this de facto design with two diagnostic findings. (1) Non-uniform…