New Machine Learning (cs.LG) papers, Thursday 22 October 2026
239 papers announced on arXiv. Get them ranked for you, free →
Interdomain Attention: Beyond Token-Level Key-Value Memory
AI summaryInterdomain Attention mixes softmax attention and state‑space recurrent memory in a single layer, compressing keys/values into an SSM. Across 125M–1.3B models, IA beats S4D baselines and softmax Transformers on language, reasoning, and genomic tasks.
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…
Machine-learned spectral PDE solvers for fluids and plasmas
AI summaryWARP learns PDE solvers by differentiating through spectral Galerkin rollouts anchored to data, yielding numerically stable predictions from new initial conditions. In eight fluid and plasma benchmarks, it improves coefficient recovery, rollout accuracy, and statistical metrics over single-step weak-form SINDy.
To Keep or Not to Keep: Learning KV Cache Retention in Disaggregated LLM Serving Systems
Disaggregated \textsc{LLM} serving separates prefill and decode across node pools, placing a network fabric between \textsc{KV}-cache computation and consumption. This breaks a key assumption of classical cache policies: a miss no longer incurs only local recomputation, but also…
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…
Solar System Surface Technosignatures: Searching for Signs of Artificial Design on Extraterrestrial Terrain
Artificial radio emissions have historically been a major focus of technosignature searches. Here, we apply an anthropological lens to expand this scope to surface technosignatures: non-random, intentional patterning on planetary terrain, similar to long-lasting human-made…
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.
Scalable Stress-Constrained Topology Optimization: A Super-resolution and Simultaneous Approach Based on Machine Learning
AI summarySNTO trains neural networks to represent state, adjoint, and design fields in a single loop, enabling stress-constrained topology optimization at millions of degrees of freedom. It achieves super‑resolution, sharper boundaries, lower volume fractions, and scales to 3D problems on multiple GPUs.
SimHack: Predicting the onset of detectable reward hacking before RL training
AI summarySimHack predicts when reward hacking first appears during RL training, learning dynamics from a few small runs and simulating unseen recipes. It achieves 11.3% relative error on median onset and 0.85 rank correlation, enabling safer curriculum choices that avoid hacking.
Borrowing Strength Across Stores: Transfer Learning for Choice-Based Revenue Management
AI summaryThe paper introduces a shrinkage estimator for the multinomial logit demand model that balances store‑specific estimation with cross‑store pooling. It proves finite‑sample bounds on estimation error and downstream assortment/regret, showing empirically that moderate shrinkage boosts revenue for hundreds of apparel stores.
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.
Big Claims, Small Margins: Do Reported GraphLLM Gains Survive Simple Baselines?
AI summaryThey benchmark LLM‑based node classifiers against inexpensive baselines on three recent TAG datasets, finding that simple TF‑IDF or frozen e5‑large‑v2 models achieve within 3% of reported LLM gains on most tasks, but some reported improvements remain significant.
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.
A Reverse-Gold Theorem: Positive-Data Identification by Substrate-Bounded Contrastive Learners
Gold's theorem shows that no learner can identify an arbitrary language in a superfinite class from positive presentations alone. We prove a positive result of a different shape. Fix an innate \emph{recursive contrast operator} $R$ mapping any string to a finite set of…
Understanding the Inner Workings of a State-Space Foundation Model for Limit Order Books
Autoregressive foundation models offer state-of-the-art capabilities to simulate the market by generating the message flow of limit order books (LOBs). Yet, without understanding their inner workings, research directions are hard to narrow and practical use in real-world trading…
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.
Motion Equivariant Convolutional LSTM: When Motion Becomes a Gauge Symmetry
AI summaryMotion Equivariant Convolutional LSTM (MELSTM) models sequences as gauge fields, making the network equivariant to any time‑varying motion. It outperforms standard CLSTM and flow‑equivariant CLSTM on future‑frame prediction, motion‑based digit classification, and action recognition, generalizing to unseen motions with fewer parameters.
Building a Nested LADDER for LLM Unstructured Post-Training Pruning
AI summaryLADDER allocates sparsity layer‑wise by iteratively pruning the least‑degrading fraction of weights, creating nested masks for multiple budgets. It matches uniform pruning at low sparsity but outperforms it from 70%+ sparsity, maintaining higher zero‑shot accuracy after distillation.
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.
Extreme Low-State Quantized Transformer for Electricity-Load Forecasting
AI summaryThe paper quantizes Transformer weights to 3–16 discrete states for electricity‑load forecasting, showing post‑training quantization degrades error by ~6–7%, while fine‑tuning recovers 90–95%. An accumulated‑update thresholding method cuts state changes by 80% with minimal error increase.
Measure First, Then Allocate: Calibrated Rollout Budgets for MCP Agents in Reinforcement Learning
AI summaryThe paper introduces Discord for evaluating harnesses and BRACE for allocating rollout budgets using reward disagreement. BRACE improves policy training efficiency in tool-using language models, outperforming strong baselines with fewer rollouts across multiple benchmarks.
Is Anyone Enforcing the Cartel? What Reinforcement Learning Teaches LLM Agents in Markets
AI summaryLLMs trained by RL collude when enforced, but only if they see a punishment or imitate a rival’s cuts, not through strategic deterrence. An audit shows that only agents that directly match large cuts can enforce a cartel, while others merely follow static rules.
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.
An Explicit Feedback Contract for Iterative ML-Agent Harnesses
AI summaryIntroduces an explicit feedback contract that releases validation scores only when noise is below a threshold, preventing adaptive data leakage. On real and null-control tasks, the contract reduces the validation‑test gap, improves test performance, and stops agents early (4‑12 vs 30 submissions).
Depth Separation for Neural Networks under Gaussian Measure
AI summaryDepth‑3 ReLU networks of polynomial size can approximate a particular radial function under Gaussian measure to fixed small L² error, while any depth‑2 network achieving the same error must have exponential size or subexponential weights; the result follows from a Hermite‑frequency lower bound verified in Lean 4.
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…
A Mechanistic Study of Automated Failure Attribution in LLM-based Agentic Systems
AI summaryThe paper shows that large language models internally encode failure‑attribution information but rarely surface it in outputs. Introducing a simple grafting trick that activates this latent knowledge, they achieve higher accuracy than direct prompting across several LLMs and benchmarks.
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…
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.
KIRA: A Full-Stack Reading Screening System with Dual-Pathway Analysis for Children
AI summaryKIRA pairs word‑level fluency scoring with phoneme‑level error detection for early‑grade reading screening, using child‑adapted recognizers that cut word error rates to 8.84% and phoneme error rates to 23%. It provides teachers with fluency scores and phoneme feedback via a web interface.
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.
Self-Supervised Tuning of Low-Fidelity Optimization Models via High-Fidelity Feedback
Low-fidelity (LF) optimization models are widely used in large-scale operations because they can be solved efficiently and reliably, but they often exhibit systematic errors relative to high-fidelity (HF) formulations that better capture physical and operational constraints.…
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.
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.
How Much Gradient Does Truncated Backprop Lose in Looped Transformers?
Looped transformers reuse one block of layers many times and are often trained by backpropagating through only the last k loops. We measure how much of the weight gradient this throws away, exactly and for every k from one backward pass, on nine looped language models, and…
LEAN: Loss-Engineered Modulation of the Model Weight-Sensitivity Distribution for Efficient, Error-Tolerant Neural Networks
AI summaryThe paper introduces LEAN, a regularizer that reshapes a trained network’s weight‑sensitivity distribution, reducing the most fragile weights by 5‑13× with negligible accuracy loss. This allows smaller parity codes and multi‑bit formats, giving models up to 3.3‑bit storage while surviving thousands of bit‑flip attacks.
Actually, All Learnable Distribution Classes Are Privately Learnable
AI summaryAny class of distributions learnable with finitely many samples up to arbitrary total‑variation error is also privately learnable with finitely many samples, proving Ashtiani’s conjecture under the standard learnability definition.
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.
Combating Catastrophic Forgetting in Continual Domain Adaptation via Knowledge Recasting
AI summarySkill Schema Transport (SST) converts new domain examples into reusable procedural schemas, allowing a language model to learn new knowledge without altering its core skill space. SST reduces catastrophic forgetting by 7.10‑42.41 points and boosts final performance by 2.53‑5.75 points in continual domain adaptation.
CSD: Common-Support Distillation for Cross-Tokenizer Knowledge Transfer
AI summaryCommon-Support Distillation aligns teacher and student vocabularies by building a shared support for each input chunk and matching induced logits with a single objective. It outperforms prior cross-tokenizer methods, boosting student accuracy by up to 4.2 points in several Llama/Gemma setups.
Is Refusal Direction Universal across Low-Resourced Languages?
AI summaryThe paper tests refusal directions in three low‑resource languages (Belarusian, Bashkir, Tajik) on Qwen2.5‑14B-Instruct and finds that transfer is graded, not universal: safe‑aligned languages transfer well, but strongly misaligned ones fail to modulate refusal, weakening safety guarantees.
Minimax-Optimal Predictive Flow Matching under Distribution Drift
AI summaryPredictive flow matching under distribution drift is tackled with DyProd‑FM, which refits flow fields on dyadic prefixes and selects among them using a sleeping Prod algorithm. It achieves minimax‑optimal dynamic regret rates \(\widetilde{O}(\min\{N_\omega+1,1+\sqrt{T\Omega}\})\) and \(\widetilde{O}(\min\{N_\mu+1,1+T^{1/3}V_\mu^{2/3}\})\) for velocity and law variations, respectively, even for nonconvex neural classes.
Tropical Cyclone Attributes in Sparse Autoencoder Features of an AI Weather Model
AI summaryThe study uses sparse autoencoders to probe GraphCast’s internal representation of tropical cyclones, mapping activations to cyclone position, intensity, and pressure. It finds consistent, reproducible attribute–feature patterns across layers, with deeper layers better predicting future position and pressure.
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…
Backtracking Direct, Indirect and Total Effects
Natural direct and indirect effects are defined through cross-world interventional counterfactuals. We propose an alternative construction based on backtracking counterfactuals. Rather than intervening on the treatment or mediator, we use a backtracking conditional distribution…
A Model-Free Continuous-Time RL Algorithm from Discrete-Time Off-Policy Data with End-to-End Guarantees
In this paper, we study continuous-time reinforcement learning, where the state follows an unknown stochastic differential equation but only a fixed off-policy batch of discrete-time transitions is available. Standard discrete-time Q-learning becomes problematic as the sampling…
Query Feedback and Teaching for Stable and Certifiable Language Generation
AI summaryThe paper defines stability and a stricter notion, certifiability, for language generators and introduces teaching dimensions for correctness, stability, and certifiability. It characterizes stable and certifiable generation with finite witnesses, compares them to identification in the limit, and evaluates teaching costs for regular and erasing pattern languages.
One Policy, Many Styles: Where to Aggregate Preferences for Steerable Humanoid Control?
AI summaryThe paper evaluates how to aggregate user preferences in steerable humanoid control, comparing reward-, advantage-, loss-level, and PCGrad aggregations. It finds no single aggregation dominates, with early aggregation yielding broader style changes but more unintended effects, while later aggregation offers precision but lower responsiveness.
The Data Differentiator Framework (DDF): Quantifying Marginal Data Value with Information-Theoretic and Game-Theoretic Measures in an Agentic Evaluation Pipeline (SDSAE)
AI summaryThe Data Differentiator Framework scores candidate data assets by penalizing mutual‑information redundancy, normalizing marginal utility from ablation studies, and applying business governance terms. On a credit‑card fraud benchmark only the genuinely differentiated asset passes significance and is recommended for acquisition.
Auditing Conformal Prediction for Language Models Under Inference-Time Model Shift
AI summaryThis paper formalizes inference-time model shift in language model conformal prediction, demonstrating that changing generation settings between calibration and deployment breaks theoretical guarantees. Consequently, standard evaluations substantially overstate true deployment coverage.
Multi-Objective Protein Optimization via Modular Diffusion Steering
AI summaryMODS is a diffusion‑steering framework that samples scalar Chebyshev weights and surrogate members to generate protein sequences targeting different objective tradeoffs. It achieves competitive hypervolume and outperforms baselines on six protein optimization tasks with limited experimental feedback.
Is 2:4 Sparsity Still Worth It? Semi-Structured Sparsity versus Quantization in LLM Inference
AI summaryThe paper compares adding 2:4 sparsity versus lower-precision quantization for transformer inference on NVIDIA GPUs, finding sparsity hurts accuracy and speed unless quantization is at the hardware minimum. At lowest precision, 2:4 can still boost throughput by 27% with retraining.
Online Open-World Wafer Defect Discovery using Hyperdimensional Computing
AI summaryWaferHDC applies hyperdimensional computing to continuously detect and cluster unseen wafer defects online, avoiding retraining. It achieves 13.3 p.p. higher adjusted Rand index and 10.5 p.p. higher NMI than the best online baseline on WM‑811K.
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…
Transformers as Generic Deep Neural Networks: Exact Compilation, the Cost of Removing Activation-Activation Products, and Multi-Agent Attention Clustering
AI summaryWe show a Transformer block can be exactly compiled into a generic DNN using only affine, routing, square, sigmoid, inverse‐sqrt, and GELU or its sigmoid approximation. Removing activation‑activation products incurs an $N^2$ cost; keeping a single operation type eliminates it, and exact tail‑mass estimation outperforms tail dropping with fewer keys.
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…
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%.
How Does Orthogonalization Adapt to the Neural Network Hessian Structure? A Gradient Self Outer-Product Analysis at Initialization
AI summaryAnalyzes how Muon’s orthogonalization aligns with neural‑network Hessians, showing that overparameterization drives diagonal dominance of the momentum’s Gram matrix. In four canonical models, diagonal entries grow faster than off‑diagonals, matching empirical scaling in language‑model pretraining.
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.
Is In-Context Reinforcement Learning A Posterior Sampler? An Empirical Study
AI summaryICRL, specifically DPT and variants, is tested against two Bayesian posterior sampler conditions: sufficient‑statistic invariance and martingale property. In Gaussian, Bernoulli, and linear bandit experiments, both conditions fail, indicating ICRL does not act as a posterior sampler in practice.
From Common Information to Canonical Correlation Analysis and Beyond: Bridging Information Theory, Signal Processing, and Machine Learning
Many problems in signal processing, machine learning, and neuroimaging provide several jointly observed measurements of one process, each mixing a component common to all of them with a component private to itself, and require the common component to be recovered without…
Attention-based Classification for Indirect Prompt Injection
AI summaryAttention Tracker fails to reliably detect indirect prompt injection in agentic LLM settings, achieving near-chance performance on AgentDojo. Replacing its scalar focus-score threshold with a Fisher Linear Discriminant (AttnDA) improves detection across benchmarks, but still requires domain‑specific training data.
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.
Doob-Based Generative Policy Optimization
Diffusion and flow models provide expressive policy classes for robotic and sequential decision-making systems. However, their action likelihoods are generally intractable, limiting the direct use of widely adopted policy optimization methods such as PPO and GRPO. We propose…
The Optimal Sampler for Diverse TSP Solutions: Characterization, Limits, and Learning
We address the Diverse Traveling Salesman Problem (D-TSP) that seeks a set of $k$ distinct TSP tours. Each selected tour must not exceed a prescribed multiple of the optimal tour length, while the maximum pairwise Jaccard similarity within the selected set is minimized. This…
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.
FlashNext-RL: A System Optimization Study for Long-Horizon Agentic RL of Qwen3.8-Flash-Next
AI summaryFlashNext‑RL is a model‑specific RL training framework for Qwen3.8‑Flash‑Next that freezes early layers, offloads memory, and optimizes mixture‑of‑experts and sparse‑attention kernels. It boosts non‑padding throughput to 4,438 tokens/s/GPU (213 TFLOPs) and enables 131 k‑token context long‑horizon RL in 22 min per iteration.
A Jump-Modified Wasserstein Gradient Flow for Maximum Mean Discrepancy with Energy Kernels
We study a jump-modified Wasserstein gradient flow of the squared maximum mean discrepancy generated by the energy kernels $K(z)=-|z|^q$, $0<q<2$, in every dimension $d\ge2$ as well as the one-dimensional range $1\le q<2$. Under subcritical density and finite-moment assumptions,…
Local and Sparse Kernel Operator Learning for Incompressible Flows: An Operator Lagrange Formulation
AI summaryLocal sparse kernel learning uses divergence‑free polynomials and compactly supported kernels to build operator Lagrange weights that avoid dense global reconstruction. It yields accurate, low‑storage models—often matching or beating neural operators—while scaling efficiently on GPUs.
Target Regeneration Shrinks the Gains of Draft-Training Objectives in Speculative Decoding
AI summarySpeculative decoding uses a draft model whose tokens are verified by a target model; new draft‑training objectives improve acceptance on externally written responses, but their benefit largely disappears after the target model greedily regenerates the draft. The paper shows this shrinkage across data, verifiers, and decoding settings, concluding that gains should be evaluated against a target‑regenerated baseline.
Dissimilarity-Based Markov Decision Processes for Stabilizing State Distributions
AI summaryDMDP framework learns policies that stabilize state distributions by minimizing dissimilarity to a target. The paper proves asymptotic stability under sufficient conditions, provides finite-dimensional distribution representations and certificates, and demonstrates benefits on HalfCheetah with PPO and SAC.
Better Policies, Worse Verifier Audits
AI summaryThe authors show that as reinforcement learning policies improve, standard on-policy verifier audits degrade because successful generations obscure component-level failures. They prove this audit collapse leads to suboptimal verifier selection and propose a mixture-based repair to preserve failure support.
Gate Initialisation versus Recurrent Dropout and Zoneout in Learning Long Dependencies: A Controlled Study of LSTM and GRU Networks on Synthetic Tasks
AI summaryThis study evaluates LSTM and GRU regularizers and gate initializations on long-dependency synthetic tasks through a large factorial grid. It finds that chrono-initialization consistently enables successful learning across difficult sequence lengths, vastly outperforming conventional initialization and standard recurrent dropout or zoneout.
ART-MPG: Autoregressive R/T Manifold Path Guiding
Specular chains concentrate important transport in small regions of path space and the structure depends on chain topology. We introduce ART-MPG, a compact neural model that samples chain length, an autoregressive reflection/transmission (R/T) string, and a direction from a…
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.
Dynamics of Hidden Learning Structure
AI summaryThe paper introduces Fiber Regulation Theory and Fiber Acquisition Theory to explain how a learner’s hidden learning structure evolves during training, and shows that combining protected repair and fixed-capacity LoRA renewal on Qwen2.5-32B yields higher teacher‑forced accuracy and lower macro NLL than either method alone.
OccJEPA: Occlusion as Masking for Amodal Video Segmentation
AI summaryOccJEPA treats occlusion as masking, using a pre‑trained V‑JEPA 2.1 predictor to fill in unseen pixels and a light head to decode amodal masks. It is deterministic, 33.5× faster than diffusion methods, and outperforms them on SAIL‑VOS, TAO‑Amodal, and Kubric.
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.
The Price of Choice in Selective Prediction
AI summaryThe paper studies auditing selective classifiers that must satisfy a risk–coverage contract using a menu of decision rules. It shows the minimax sample size for K‑interval menus is \(K(1-\kappa)/(\kappa\gamma^2)\) (up to logs), achieved by the intersection‑union test, and that naive procedures can incur linear‑\(K\) costs.
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.
Internal Guidance as a Sampling Method for Tilted Diffusion Models
AI summaryThe paper introduces internal guidance for diffusion models, shifting the denoiser’s input instead of adding guidance terms. By training with random diagonal rescalings, a single denoiser samples binary reward tilts, class labels, and infilling without fine‑tuning or external guidance.
Causal world models under physical parameter shifts
AI summaryCausal world models that encode physical parameters explicitly (e.g., link masses in affine inertia terms) keep OOD prediction accuracy, boosting joint R² from 0.609 to 0.727 for out-of-range masses, whereas fully connected models collapse to 0.697 RMSE on a synthetic benchmark.
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…
DriftNeRF: Cross-Ray Density Drifting for Sparse-View Neural Radiance Fields
Neural Radiance Fields (NeRF) synthesize photorealistic novel views from dense image collections, but degrade sharply when few views are available. We attribute a specific, measurable defect to the objective itself: the rendering loss is a sum over independent rays and contains…
SU(2)-Equivariant Spinor Message Passing for Molecular Magnetism
Predicting molecular magnetism often requires electronic-structure calculations with spin-orbit coupling (SOC), which are costly to repeat across many molecules. The underlying electronic states can transform as spinors, yet equivariant surrogates typically use only integer-rank…
MHeatmap: Two-Walk Laplacian Reordering and Mosaic Layout for Rectangular Matrix Visualization
AI summaryMHeatmap reorganizes rectangular matrices by using Two‑Walk spectral seriation to jointly order rows and columns, then displays them with a Mosaic layout that sizes cells by row/column totals. It achieves the best seriation metrics on synthetic data and real datasets.
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.
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…
SAFTER: Safety-Aware Filtering Through Explicit Recoverability for Robust Reinforcement Learning under Distribution Shift
Learning safe policies that generalize to unseen environments is difficult under bounded actuation, where an unsafe decision may become unrecoverable. Conventional control barrier function (CBF) filters enforce only instantaneous safety and ignore future recoverability. We…
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.
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.
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.
ROCOCO: Robust Outline Completion of Occluded Curvilinear Objects
AI summaryROCOCO uses flow-matching generative models to inpaint occluded curvilinear objects without a prior mask, augmented with a UNet that predicts the mask from the observation. It outperforms baseline inpainting methods, achieving state‑of‑the‑art full‑image error with far fewer parameters.
A Decade of Learning-Based Control: From Theoretic Foundations to Practical Deployment
Motivated by ten years of Learning-Based Control Invited Sessions at the IEEE Conference on Decision and Control (CDC), this tutorial paper reviews key advances at the intersection of machine learning and control. We focus on methods that combine the flexibility of machine…
Curvature-Scaled Sharpness-Aware Minimization: Convergence Beyond Smoothness
Sharpness-Aware Minimization (SAM) has established itself as a widely used method for training deep neural networks. Despite its popularity, existing convergence guarantees for SAM-type methods require smoothness, a restrictive assumption. Generalized smoothness conditions have…
Truncated Out-of-Bag Probability Gain: A Noise-Robust Criterion for Random-Forest Wrapper Feature Selection
AI summaryTOP‑Gain scores candidate features by the mean increase in OOB probability for the true class, truncating negative changes to reduce noise. It outperforms 14 baselines on 50+ datasets, yielding higher ROC‑AUC, accuracy, and more stable feature sets with far less computation.
A Unified Evaluation Framework of Physical Adversarial Robustness in the Edge VLM Era
AI summaryThe paper introduces a unified metric, conditional attack success rate (cASR), to evaluate physical adversarial patches on edge vision‑language models and object detectors. It finds VLMs are as vulnerable as detectors, with failure modes shifting from omission to displacement/substitution under distribution shift.
Reconstruction differences rival individual differences in the fly compass connectome and break trained circuit models
Neural network models are increasingly constructed from synapse-resolution connectomes, and the same circuit is now available in reconstructions from other animals and pipelines. We do not know whether such models are dependent on the reconstruction they are built from, and…
On the gradient conjecture at infinity: Finite spherical length for gradient trajectories of definable potentials
AI summaryThe paper proves that any gradient trajectory of a definable \(C^1\) potential has finite spherical length, thus always converges directionally. This solves the gradient conjecture at infinity and the definable Thom gradient conjecture, and implies directional convergence for DNN training flows.
Preference Is Not Permission: Personalizing the Auto-Purchase Line in Agentic Commerce
AI summaryThe paper defines a personalized auto‑purchase line with a three‑state decision (don’t buy, confirm, buy) based on warrant, reversibility, price, and urgency, and shows that shallow models trained on a few dozen personal labels can predict this boundary with 0.79 accuracy, outperforming all prompt‑model baselines.
PerSPL: Personalized Semi-supervised Preference Learning
AI summaryPerSPL learns personalized reward models from unlabeled responses by pair‑mean score centering, turning them into user‑specific signals. Across four benchmarks, it outperforms seven baselines, achieving the highest accuracy for both seen and unseen users.
Delay-Guided Attention for Quality Prediction in Multistage Manufacturing
AI summaryDAMF estimates per-stage lag distributions and uses them to guide cross-stage attention, improving quality prediction in multistage manufacturing. On LiquidSugar and benchmark datasets, it outperforms baselines and yields accurate, stable lag recovery.
Schrödinger-Bass Bridge Barycenter Models for Measure-Valued Data
AI summarySchrödinger‑Bass Bridge barycenter models use directional asymmetry and dual drift‑diffusion control to encode, decode, symmetrize, and Markovianize measure‑valued data. They guarantee barycenter existence, compute Gaussian barycenters via finite‑dimensional optimization, and outperform existing methods on image and scRNA interpolation tasks.
Cache Locally, Recompute Selectively for Streaming Video Understanding
Streaming video understanding (SVU) requires video-LLMs to process incoming video continuously while answering user queries with low latency. Recent methods reduce query-time cost by prefilling frames into KV caches during idle time and reusing the cached KVs at query time. To…
MA-UITD: Offline Multi-Agent Reward Inference for Characterizing Continuous Social Behavior
AI summaryMA‑UITD infers agent‑specific reward functions from offline joint demonstrations in continuous‑state environments without known dynamics. It achieves state‑of‑the‑art reward recovery in discrete and continuous tasks and reveals distinct reward structures for leaders and followers in real social‑foraging data.
Extending FAIR Principles with Observability for Trustworthy Federated Learning
AI summaryExtends FAIR principles to federated learning by adding an observability layer that produces signed trust evidence from each participant per round. Demonstrates a three-level trust model and a Python prototype for evidence generation and trust assessment, enabling future participant selection and aggregation.
Group Distributionally Robust Optimization over Decentralized Networks
We study group distributionally robust optimization (GDRO) over decentralized networks, where workers collaborate to learn robust models by minimizing the maximum local risk. Existing methods optimize only dual-regularized surrogates and establish guarantees for a globally…
Estimating Performance for Safe Deployment of Machine Learning Models
AI summaryMOVAL estimates model performance from unlabeled data by class-wise confidence calibration and an ensemble of calibrated scores, achieving state‑of‑the‑art accuracy, F1, AUC, and pixel‑level metrics across distribution shifts in natural and medical imaging.
Design and Tuning of Multi-Agent Systems Using Bayesian Evolutionary Optimization
AI summaryBayesian Evolutionary Optimization automatically designs multi‑agent system topologies by evolving workflow graphs and using a graph‑aware Gaussian Process with Expected Improvement to prioritize expensive evaluations. The method outperforms existing systems on code‑generation and tabular pre‑processing tasks under limited budgets.
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.
Approximate Solomonoff Induction: On the Robustness of Learned Universal Predictors
AI summaryThe paper bounds how close an approximate universal predictor must be to the Solomonoff mixture to keep prediction error low. It shows worst‑case excess error is Θ(η/w_μ), but averaged over the prior it is only O(η), and extends to distribution shift and finite‑sample settings.
Compressing Onboard Language Models for Space Robots under a Behavioral-Fidelity Requirement
Language models are being explored as the interface through which space robots interpret operator commands and select among verified skills. Running them onboard calls for low-bit compression, and flight use adds requirements that generic post-training quantization leaves…
ExpEvolve: Structuring and Reusing Experience for Open-Ended Discovery
AI summaryExpEvolve injects idea-level extraction into evolutionary coding agents, identifying high‑level factors that explain program performance and using them for guided exploitation and exploration. Across 11 math tasks it boosts final performance and sample efficiency by up to 5×, with transfer to other LLMs and MLS‑Bench.
Data-driven and physics-guided hindcasting of measured wave spectra
Short-range hindcasts of measured wave spectra are evaluated using a recurrent model initialized by one hourly buoy spectrum and supplied with local wind, fetch and depth. Measurements from three Baltic Sea stations are combined for training in 2017--2020, validation in 2021 and…
Interpretable Cutting Plane Selection for Mixed-Integer Linear Programming Using Genetic Programming
AI summaryGPSCS evolves an interpretable symbolic scoring function via genetic programming to rank cuts and a statistics‑based rule to choose how many cuts to add. On SCIP it reduces solving time by 6 % versus default cuts and doubles speed compared to a neural‑network method.
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.
DISRL: Dual-Scale Incremental Semi-Supervised Reinforcement Learning for Dynamic Community Tracking
Dynamic local community detection tracks the evolving community of designated seed nodes across graph snapshots. However, it faces two critical challenges: temporal supervision uncertainty, where sparse and outdated labels fail to provide seed-specific guidance under topology…
BrainCodes: A Lightweight EEG Foundation Model over Discrete Neural Codes
Recent electroencephalography (EEG) foundation models increasingly emphasize model scale, potentially limiting their deployment on resource-constrained devices and in everyday environments. We argue that transferable EEG features can instead be learned with a compact…
Reader--Writer Coupling in Transformer MLPs
A transformer MLP gives each hidden channel two roles: a reader that decides when the channel responds, and a writer that decides what it adds back to the residual stream. Read as a key--value memory, the layer can be measured one factor set at a time: the pairing of readers to…
Decodability Is Not Controllability: Local and Anisotropic Speaker Control in Continuous-Latent TTS
High linear decodability does not entail causal controllability in continuous-latent text-to-speech (TTS). In VoxCPM2, position-matched probes and equal-norm interventions across 30 speakers and 60 directed dyads expose this separation: reference-audio states at the final…
Benchmarking and Error Analysis of Deep Search and Deep Research LLM Agents
As Large Language Models (LLMs) evolve into autonomous Deep Search and Deep Research Agents capable of executing long-horizon, multi-step tasks, evaluating their reliability becomes increasingly critical. However, existing evaluation paradigms predominantly focus on final task…
Data Schedules and Learning Rate Schedules Should Be Designed Together
Mid-training (MT) is a training phase in LLM training pipelines that continues pretraining (PT) on smaller datasets of higher value. Taking the PT and MT phases as a whole, training usually starts with only PT data and then transitions completely to MT data. In this work, we…
Closed-Loop Dual-Reference Guidance for Accent Control in Cross-Lingual Voice Cloning
Zero-shot cross-lingual voice cloning often carries the reference speaker's accent into the target language. We propose closed-loop dual-reference guidance, which pairs the original reference with a reusable native anchor constructed by voice conversion (VC). The references…
GaP-MoE : Explicit Expert Specialization in Mixture-of-Experts via Orthogonal Gradient Accumulation and Projection
Sparse Mixture-of-Experts (SMoE) models improve scalability by routing input tokens to only a subset of expert networks, but they often suffer from representation collapse, in which multiple experts converge toward similar functions and lose their intended diversity. Prior work…
Compositional neural networks for multimodal X-ray and neutron scattering inversion
X-ray and neutron scattering are indispensable for structural characterization across condensed-matter physics and materials science. Yet, because scattering experiments generally record intensities without phase information, distinct structures can yield similar scattering…
Right On Q: Recovering Reasoning under Quantization via Self-Generated Trajectories
Quantization-aware training (QAT) aims to recover reasoning capabilities in low-bit large language models (LLMs). Existing recipes typically supervise the quantized student with the original full-precision model on fixed response sequences. However, the resulting low-bit…
Cascading Bandits with Noisy Observations: Algorithms and Regret Bounds
We propose a novel stochastic cascading bandit setup where probing an ordered shortlist of $K$ out of $L$ arms incurs position-dependent costs, and the learner observes only noisy feedback regarding each arm's true output. We identify a sub-class of noise models for which the…
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…
Nominate-then-Adjudicate: LLM-Assisted One-Pass and Per-Instance MIP Solver Tuning
Mixed-integer programming solvers expose hundreds or even thousands of parameters whose best setting varies per instance, yet per-instance search is prohibitively expensive. We frame this as configuration for heterogeneous Big Data optimization workloads and propose a…
Progressive Permutation Alignment with Temperature-Controlled Spectral Sharpening
In implicit neural representations (INRs), 3D signals are encoded directly in network weights, making stable weight-space operations essential. However, permutation symmetry causes independently trained yet functionally equivalent models to appear distant in weight space, often…
Region-Aware Multi-View Inspection Using Deep Learning and Domain-Informed Decision Fusion
Automated optical inspection of precision components with tiny surface defects remains challenging when different regions of the same component exhibit different defect characteristics and, therefore, different acceptance criteria. This paper presents a region-aware, dual-model…
The Predictive--Optimization Gap in GuanDan Self-Play
AI summaryThe paper shows that in GuanDan, a predictive signal (state–gain interactions) improves outcome prediction but does not consistently improve policy optimization: a single checkpoint outperforms a baseline, yet across multiple seeds and iterations the effect averages to zero.
Alternative Strategies for Active Learning with Bayesian Bradley Terry Model
AI summaryBayesian Bradley‑Terry inference with Polya‑Gamma Gibbs sampling plus three knowledge‑gradient active learning rules (Kendall‑tau, rank entropy, posterior separation) improves rank recovery from pairwise judgments. Entropy selection is best at moderate noise; posterior‑separation wins at high noise, and the PSR metric tracks accuracy.
When Thoughts Become Memory: Unifying In-Context Learning and Fine-Tuning
AI summaryProposes a certified rule that quantifies whether new data should be added to a model’s context or its weights by measuring the code length saved on held‑out queries. The rule correctly selects between in‑context learning and fine‑tuning across transformers, TabPFN, occupancy and stereo networks, achieving 61,948/64,000 correct decisions and ≤3.3% error on null streams.
Generalization Bounds via Vector-Valued $L_p$ Correlation $V_\alpha$
Understanding generalization of randomized learning algorithms remains a central challenge. Mutual-information and total-variation bounds provide complementary ways to quantify data--hypothesis dependence, but they can behave quite differently in high-dimensional regimes. We…
BIRD-RL: Scaling Agentic Reinforcement Learning over Stateful Data-Centric Environments
AI summaryBIRD‑RL introduces a four‑stage curriculum and a trajectory‑scoped execution infrastructure that allow large language models to learn data‑centric tasks like SQL debugging without isolated databases, achieving a 12.8× spatial efficiency gain and 44–48% success rates on the BIRD‑CRITIC‑SQLite benchmark.
One Act, Many Masks: Effect-Marginal Policy Optimization for Long-Horizon Agents
AI summaryEffect-Marginal Policy Optimization (EMPO) aligns RL updates to execution effects rather than specific text, correcting credit‑misalignment in long‑horizon LLM agents. EMPO boosts task success and learning speed on ALFWorld, WebShop, and Sokoban without extra critics or rollouts.
How Far Can Foundation Models Be Distilled for Efficient Jet Tagging?
Jet tagging is one of the most important applications of machine learning at the Large Hadron Collider. Recent models build in more physics knowledge or rely on large-scale pretraining, but the experiments place strict limits on their inference cost. Knowledge distillation can…
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,…
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.
Asymmetry Tunneling Beyond Adaptive Equivariance Relaxation
AI summaryA residual path is added to a strictly equivariant model, with a deliberately limited interface to test equivariance relaxation. Across five benchmarks the study finds “asymmetry tunneling,” where residuals survive despite restrictions, indicating adaptive relaxation is not required to capture symmetry-breaking structure.
Density-Level Backward Error for Low-NFE Generative Flows
AI summaryDensity-level backward error identifies which part of a flow’s numerical defect alters the density, enabling a solver‑order‑aware compute‑allocation rule. The method allocates evaluations more efficiently than curvature‑based schemes and improves low‑NFE Euler sampling on pretrained ODE flows.
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.
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.
Asynchronous, Generation-Free ABC: Streaming AMIS Reweighting for Heterogeneous Simulator Workloads on HPC
Approximate Bayesian computation (ABC) enables Bayesian inference for models with intractable likelihoods by estimating parameter posterior distributions from simulated data. Sequential ABC methods improve sampling efficiency by iteratively adapting the proposal distribution…
Shortcut-Resistant CAM Distillation for Long-Tailed Recognition
AI summaryShortcut‑Resistant CAM Distillation (SRCD) transfers object‑centric explanations from head classes to tail classes by aggregating head‑class CAMs with energy‑based weighting, then distilling them to tail‑class CAMs, reducing shortcut reliance and improving long‑tailed recognition.
A unified approach for estimation of unnormalized models via Stein operators
AI summaryThe paper unifies estimation of unnormalized models by linking proper scoring rules (e.g., score matching, pseudolikelihood, NCE) to Stein operators. It shows these rules induce Stein operators, providing a geometric framework that characterizes the asymptotic efficiency of the estimators.
BRIDGE: An Acceptance Gate to Block False Improvements in Model Updates, with a Retrospective Clinical Evaluation
Models are updated repeatedly, and an update that looks better on validation data but not on new data is a false improvement that can misguide clinical decisions. The naive rule accepts every positive validation gain and is optimistic when the same data choose and judge the…
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…
Supervision-Adaptive Curriculum and Guarded Anti-Curriculum Learning for Dense Prediction
Curriculum learning orders samples by difficulty, but difficulty alone does not establish supervision reliability for dense prediction. Easy samples can expose shortcuts, hard samples can be informative, and noisy or ambiguous samples can mislead training. Building on our…
pplx-embed-v2-late Technical Report
In this report, we introduce pplx-embed-v2-late, a pair of 0.6B and 9B multilingual and multimodal late-interaction models. Both models achieve state-of-the-art results within their respective size classes across a range of text, image, and visual document retrieval benchmarks.…
Geometric knowledge spaces, Attention pairing and Architectures
AI summaryThe paper formalizes knowledge worlds and attention pairings geometrically, showing that standard and linear, causal, tensor, and RNN attention arise naturally from this framework. It then proposes new geometric spaces that could lead to novel transformer‑style architectures.
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…
Residual-Based Causal Discovery via Pretrained In-Context Quantile Regression
AI summaryLoscar extends RESIT to location‑scale noise models by estimating conditional mean and scale with pretrained in‑context quantile regression, avoiding parametric assumptions. It outperforms an FGLS variant on bivariate and multivariate causal‑discovery benchmarks.
Phase Transitions in Learning Driven by Extremal Data
AI summaryThe paper analyzes a minimal multiple‑instance learning model with softmax‑based max pooling, deriving a low‑dimensional gradient‑flow description in the joint limit of large bag size and dimension. It finds two phase transitions—a freezing transition from disordered systems and a retrieval transition linked to Hopfield networks—matching simulations.
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.
Hyperspherical Simplex Representations for Backward-Compatible Retrieval
Training modern AI models has become increasingly expensive, and updating deployed models can alter their internal feature representations, affecting applications built upon them. In retrieval systems, such updates often require complete gallery re-indexing, a computationally…
WRDO: Weighted Refusal Direction Optimization in Aligned Language Models
Aligned large language models are trained to comply with benign requests while refusing harmful ones. However, recent work shows that this refusal behavior is fragile and can often be disrupted by simple activation- or weight-space interventions. In particular, refusal has been…
SOAR: Social and Opinion-Aware Recommendations
One of the enduring challenges in online platforms is recommendation. Both communities and advertisers face difficulties in analyzing complex user behavior, identifying habits and evolving preferences, and providing timely recommendations that are relevant to user interests. In…
Certified Neural Approximation of Robust Nonlinear Model Predictive Control
AI summaryThis paper constructs neural network controllers that inherit the safety and stability of a robust MPC policy. By treating the approximation error as bounded disturbance and using offline Lipschitz‑based synthesis, the resulting NN preserves recursive feasibility and reduces online computation.
Implicit Feature Reweighting for Label-Free Graph Generalization
Graph-level out-of-distribution (OOD) generalization is critical for deploying graph neural networks in real-world applications where distribution shifts are inevitable. Existing approaches typically require class labels during representation learning to decide which…
Private Online Prediction from Experts with Small Losses
We consider online prediction from experts, a fundamental problem in machine learning, under differential privacy constraints. Existing private algorithms achieve near-optimal regret $\tilde{O}(\sqrt{T})$, where $T$ is the time horizon. However, this bound becomes suboptimal…
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…
What Prediction-Market Contracts Reveal at Launch: Trading Volume and Dispute Risk
Using a near-census of 802,863 resolved markets listed on Polymarket between January 2023 and May 2026, and only information available at listing (the market's title and description, named resolution source, and contract metadata), we predict lifetime trading volume and whether…
Fast Bayesian Optimization Through Linear Models
Bayesian optimization (BO) is typically used only for problems where the underlying function is deemed "expensive" (e.g. in time or money), with the speed of the BO algorithm being of little importance. In the real world, however, problems span a continuous range of values,…
Unlocking Spatial Grounding in Flow-Matching Models via SNR Tuning
Despite the success of foundation models in image synthesis, achieving fine-grained spatial control remains a significant challenge, typically requiring auxiliary structural adapters or complex inference-time steering. While modern vision-language (VLM) text encoders possess the…
Fragment-Level Graph Encoding for Interpretable Molecular Property Prediction
Graph neural networks (GNNs) are among the standard tools for molecular property prediction, but the atom-level explanations of their predictions are hard for medicinal chemists to act on. We propose a change of input representation rather than a new model: molecules are encoded…
A Transfer Learning-Driven Methodology for Efficient and Sustainable Indoor Localization Using VLC
The rapid expansion of the Internet of Things enables seamless collaboration among connected devices, making indoor localization a critical component of industrial automation. However, conventional localization techniques struggle in dynamic environments. These methods depend on…
Near Minimax Rate for Quantile Distributional Reinforcement Learning
In this paper, we study quantile distributional policy evaluation, where the return distribution under a given policy is approximated by the quantile Bellman fixed point $\bm\eta_m$. We investigate the statistical accuracy of estimating $\bm\eta_m$ across quantile resolutions…
A Diagnostic Study of Faithfulness in Diffusion Language Models for Table-to-Text Generation
AI summaryDiffusion language models are evaluated for faithfulness in table-to-text generation using ToTTo. While QLoRA‑adapted LLaDA achieves the highest BLEU (24.15), its grounding remains weak; autoregressive models drift from the table, indicating that reference similarity alone fails to capture faithfulness.
Local and Global Recursive Ridge Regression for Online Residual Correction of Frozen Multivariate Forecasters under Label Delay
AI summaryDRiC trains a ridge regression on DCT coefficients of residual blocks to correct frozen forecasters online, even with delayed labels. DRiC‑L and DRiC‑G achieve up to 14.5% MSE reduction across datasets while using 15–17× less memory than the best competitor.
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…
No Frames, Just Frequencies: Bandwidth-Efficient Deployment-Time Adaptation for Smart Glasses
AI summarySmart glasses adapt to distribution shifts by sending tiny frequency descriptors instead of frames, enabling a host to synthesize out‑of‑distribution data and fine‑tune the model. The method restores 93–95% of supervised adaptation gains with 213×–784× bandwidth savings.
Watch Your Step: Anticipated Perception for Humanoid Locomotion
AI summaryThe paper introduces a lightweight foothold prediction model that samples elevation maps only around predicted future steps, reducing perception input size. This anticipatory preprocessing cuts training time to one‑third and improves locomotion on sparse footholds, narrow corridors, and unseen terrains.
Reliability-Aware Probabilistic Fusion for Multimodal Learning with Incomplete Evidence
AI summaryReliability‑aware Gaussian Product‑of‑Experts weights each modality’s latent precision by an input‑dependent reliability score and omits missing experts. The method matches full‑data performance and outperforms baselines in AUROC when modalities are absent, especially on HAIM‑MIMIC‑MM.
Space weather dependence of decameter-scale ionospheric irregularities spatial distribution
This paper describes an approach to constructing nowcast and forecast predictive models for the occurrence of high-latitude decameter ionospheric irregularities based on SuperDARN and SECIRA radar data and space weather parameters. The method relies on self-learning pattern…
Cross-Area Transfer Learning for VLC-based Indoor Localization with a Transfer Efficiency Score
Indoor localization systems are critical for Industry-4.0 applications. Visible Light Communication (VLC)-based systems offer advantages such as immunity to electromagnetic interference, high accuracy, and energy efficiency, but their performance is hindered by environmental…
Recursive Code Evolution for Task-Oriented Semantic Communications
Semantic communication jointly optimizes inference accuracy, channel use, and computation, yet the underlying communication program is typically specified before training. This paper presents ComRSI (Communications-Oriented Recursive Self-Improvement), a framework for…
Probabilistic Machine Learning for Short-Term Atmospheric Seeing Forecasting
Accurate short-term forecasting of atmospheric seeing is of significant interest for astronomical observations and free-space optical communication systems, where rapid variations in optical turbulence can affect system performance and operational planning. In this work, we…
SPARE: Sparsity Allocation via Routing Entropy for Efficient MoEs
Mixture-of-Experts (MoE) architectures scale model capacity efficiently, yet their substantial memory footprint poses a major barrier to deployment. Expert pruning reduces this footprint, but most methods focus on which experts to remove and prune every layer equally. We propose…
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…
Leveraging Time-Series Foundation Models for Financial Crime Detection
Financial crime detection requires identifying illicit activity within large transaction networks, where suspicious behaviour is rare and unfolds over time. Existing graph-based methods capture relationships among accounts but typically underuse the temporal behaviour encoded in…
Characterizing Oblivious Common-Reward Multi-Player Bandits: Universal Stationarity, Coordination Traps, and Certified Global Rates
AI summaryThe paper analyzes oblivious common-reward multi‑player bandits, proving that independent gradient updates provide an unbiased estimator of a hidden product‑policy objective and guarantee universal finite‑time stationarity. It shows stationarity cannot guarantee global optimality, but with finite‑table certificates on product‑gradient dominance and noise, global optimality and standard $O(1/t)$ gap/$O(\log T)$ regret rates are achieved.
DriveMLP: When Training Strategy Outperforms Model Complexity in Autonomous Driving
AI summaryDriveMLP trains a lightweight MLP policy using Expert‑Guided RL, where RL optimizes a trajectory‑distance reward from logged expert data. It outperforms the previous best IL planner by +3.2 points on nuPlan Val14 Reactive while being 15× faster.
Lifting PCA in the Wasserstein Space
Principal Component Analysis (PCA) is a fundamental tool for statistical dimension reduction, yet no canonical extension to distribution-valued data has emerged within the framework of optimal transport. While Wasserstein barycenters and Wasserstein least-squares regression…
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…
Rethinking Post-hoc Debiasing of Large Pretrained Models
Large pretrained models such as CLIP are widely used as frozen feature extractors. However, their embeddings inherit spurious correlations from training data, introducing bias into the features. A growing body of research has sought to address this issue through post hoc…
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…
Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies
Vision-language-action (VLA) policies need to recognize gaps in their training knowledge to request useful demonstrations or assistance in new situations, especially when failures are costly. Disagreement within an ensemble of independently trained policies can guide these…
Monitorability Is Trainable: Post-Training for Chain-of-Thought Monitorability
Chain-of-thought (CoT) monitoring is one of the few practical tools available for overseeing capable AI systems, but it is fragile: a model's reasoning is only useful to a monitor if it remains coupled to what the model actually does. This coupling can be degraded during…
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.
Asymmetric Likelihood-Free Hypothesis Testing
AI summaryThis paper characterizes sample budgets in asymmetric likelihood‑free hypothesis testing: with different numbers of samples from reference distributions, one can test a third distribution using only $\tilde O(n^{1/4})$ samples instead of the $\Omega(\sqrt n)$ needed when budgets are equal. The results also give matching bounds for non‑adaptive 2‑clustering and show an adaptivity gap of $\tilde\Theta(k^{1/4})$.
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…
NeuroCodec: EEG-Guided Latent-Domain Generative Model for High-Fidelity Target Speaker Extraction
Driven by recent advancements in hearing assistive devices, target speaker extraction (TSE) in multi-talker environments has drawn significant attention. To address this, hearing technologies identify attention using neural signals recorded via electroencephalography (EEG) to…
RangeNav: Task-Aligned LiDAR Range-Image Pretraining for Robot Navigation
Autonomous mobile robots operating in dense crowds and cluttered sites require $360^\circ$ perception to react to surrounding hazards. While panoramic LiDAR offers full-surround coverage, learning representations within reinforcement learning (RL) is sample-inefficient under…
Silent Alignment in Two-Layer Linear Networks: A High-Dimensional Perspective
AI summaryThe paper analyzes gradient flow in high-dimensional two-layer linear networks, deriving exact asymptotic spectral dynamics. It shows a silent alignment phase where singular directions align quickly while values stay zero, followed by a slower spectral learning phase where singular values grow and loss decreases.
A Sharp Dynamic Haussler Lemma
Haussler's covering lemma (JCTA 1995) gives the sharp $(C/\delta)^d$ metric-entropy bound for binary VC classes under normalized $\ell_1$ distance. We prove a sharp global dynamic analogue for bounded real-valued classes in normalized $\ell_p$, for every $1\le p<\infty$. Writing…
Tractable Inference with Kolmogorov-Arnold Networks
Machine-learning surrogates in computational science often require not only accurate predictions but also expectations, variances, conserved integrals, and normalization constants. Conventional neural networks, including Kolmogorov--Arnold networks (KANs), generally evaluate…
Preventing Model Extraction in Contextual Linear Optimization
AI summaryThis paper studies model extraction attacks on contextual linear optimization PTO systems, proving that attackers can recover the underlying linear predictor up to scale by observing decisions across contexts. It introduces a boundary‑aware perturbation defense with provable regret and extraction‑error bounds.
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.
FrontierCS 2: A Benchmark for Computer Science R&D
AI summaryFrontierCS 2 is a 48‑task benchmark where LLM agents patch real research repositories to replicate deleted contributions, judged by hidden test suites. Claude Opus 5.5 passes 58% of runs; many tasks remain unsolved.
Mastering Model Fit: Applications of the InterModel Vigorish for Interpretable Machine Learning
Accuracy alone gives little information about which features or model components improve probabilistic fit, which outcomes benefit, or how fit changes during training. We introduce the InterModel Vigorish (IMV) into the Machine Intelligence space in order to express differences…
CORD: Conflict-Aware Optimization for Joint RL and Distillation in Agents
Joint reinforcement learning and self-distillation combines task rewards with dense teacher guidance for agent training, but their gradients can partly cancel. A mixing coefficient controls relative scale without selectively removing opposing components. We present CORD, a…
The Price of Current-Effect Inference in Nonstationary Multi-Armed Bandits
Adaptive experiments use bandit algorithms to favor interventions with higher expected rewards. However, decision-makers also require reliable inference on current treatment effects: differences between intervention and reference-arm mean rewards at the present time, rather than…
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,…
SCDM: Scalable Causal Discovery in Nonlinear Temporal Systems with Meta-Learning
AI summarySCDM learns a shared temporal predictor and causal‑strength matrix via meta‑learning across target tasks, avoiding exhaustive target‑wise optimization. It produces continuous causal scores that, under a risk‑gap condition, recover the true graph when errors stay below the population margin.
EPIC: Rethinking Trust Regions for Autoregressive LLM Policies via Exact-KL Prefix Importance-Ratio Clipping
Reinforcement learning (RL) based on PPO-style clipping has become an effective approach to improving the reasoning capabilities of large language models (LLMs). However, standard token-level clipping measures policy change only through the conditional probability of the current…
Mitigating Model Collapse through Dreaming Learning
Training generative models on synthetic data causes loss of diversity and a decrease in performance, a phenomenon known as model collapse. We study this degeneration spectrally in a controlled Markov setting. The eigenvalue spectrum of the learned transition operator offers a…
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…
ExGrad: Beyond Suppression - Explicit Clean-Gradient Subtraction in Adversarial Training
Adversarial training derives updates from surrogate losses, but source-loss descent does not directly characterize an update's effect on adversarial margins. We introduce ExGrad (extrapolative gradient recomposition), which constructs clean-input, adversarial-input, and…
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…
Looped Uniform Diffusion Language Models
Uniform diffusion language models (UDLMs) generate text by iteratively denoising sequences of ordinary vocabulary tokens, any of which may be corrupted and subsequently revised. During training, UDLMs must denoise from a context that may itself be unreliable, whereas masked…
Loopholing All the Way: Language Modeling with pure Self-Conditioning
AI summaryLAW uses pure Self‑Conditioning in a continuous diffusion language model, propagating expected embeddings and decoding once. It surpasses prior discrete/continuous models on Sudoku, Maze, GSM8K, and improves a pre‑trained diffusion model, matching ARMs on OpenWebText with 64 NFE.
From Classifier Heads to Encoder Prototypes: Personalized Federated Learning with Head-Derived Clustering
Personalized federated learning must decide which clients should share representations. We introduce HDC-FL (Head-Derived Clustering for Personalized Federated Learning), which reads that decision off a signal shared-encoder training already produces: the client-specific…
Circuits of LLM Classifiers
Language models make it easy to build classifiers by simply prompting a model with the sample to classify. However, their predictions are neither deterministic nor uniformly reliable across samples. How can we combine predictions from different models to improve reliability? We…
Orbitol: Crystal Generation on Wyckoff Orbits
Generative models propose crystalline materials by learning composition and geometry, but unconditional sampling gives no direct control over the space group, which constrains whether responses such as piezoelectricity are symmetry-allowed. We present Orbitol, a crystal…
VX-One: Typed, Calibrated Decision Heads for Real-Time Perimeter Triage - An empirical study of unknown-attack honesty, cascade gray zones, and calibration under drift
Perimeter triage faces a latency/semantics trade-off: signature engines are fast and format-safe but blind to unseen attacks, while generative models are semantic but slow and produce unparseable output. In late 2026 a class of commercial System One decision models (TypeSafe…
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…
Velocity Ambiguity Profiles: Time-Resolved Bayes-Risk Diagnostics for Flow Matching
Flow matching trains a continuum of velocity regression problems but usually reports a loss averaged over interpolation time, obscuring which regions are limited by intrinsic ambiguity and which remain data-limited. We introduce the Velocity Ambiguity Profile (VAP),…
Minimax Optimality of the NPMLE for Gaussian Mixtures in Hellinger Distance
We analyze the problem of estimating univariate Gaussian mixture densities using the nonparametric maximum likelihood estimator (NPMLE) based on $n$ independent and identically distributed samples. We show that the unconstrained NPMLE achieves the minimax optimal rate…
The role of locality and weight sharing in convolutional neural networks
AI summaryThe paper analyzes a teacher–student convolutional model to separate the effects of locality and weight sharing. It shows that for uncorrelated data convolution is P‑times faster; for correlated data weight sharing hurts and locally connected nets eventually outperform CNNs, with a predictable transition.
Self-normalized Bernstein Inequalities for Regenerative Processes and Applications to Learning Without Independence
AI summaryProves self‑normalized Bernstein inequalities for bounded functions of regenerative processes using observed cycle lengths, yielding empirical bounds with a second‑order tail term; applies them to derive minimax‑optimal generalization error bounds for agnostic learning from non‑i.i.d. data.
A Quasipolynomial Algorithm for $\ell_\infty$ Contractions
AI summaryComputes an ε‑approximate fixed point for any d‑dimensional nonexpansion under ℓ∞ in time log(1/ε)^{O(log² d)}, using a boosting technique to shrink the search region; this yields a quasipolynomial algorithm for simple stochastic games.
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…
LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data
Reinforcement learning (RL) is an effective way to strengthen the long-context reasoning of large language models (LLMs), but it needs training data that are long, diverse, challenging, and verifiable. Human annotation of such data is expensive, naturally occurring data are too…