Friday, 30 Oct 2026

Scaling Laws for Tool-Use Agents via Controllable Environment Synthesis

Scaling laws for tool-use agents from synthetic environments whose tool and task diversity is controlled.

Spine: Compiling Subtask Trees into Structured Context for Long-Horizon Coding Agents

Compiles nested subtasks into structured context so long-horizon coding agents avoid context inflation.

Bayes-Adaptive MDPs for Rollout Allocation to Enhance GRPO Training

Allocates GRPO rollouts by question difficulty, so hard questions get enough informative signals.

Collective Dynamics of Self-Reinforcement: How Cross-Task Coupling Drives Policy Collapse in Unsupervised RLVR

Explains why self-vote RLVR collapses: cross-task coupling makes the reward feed on itself.

EmbeddingGemma 2: Unified and Efficient Multimodal Representations

Open multimodal embedding model for text, images, audio and video in one unified space.

Vision-Language-Action or World-Action? Just Action Models

Separates backbone, future objective and inference cost in the VLA versus world-model debate for robot control.

From Interaction to Capability: Task Advancement as State and Supervision for GUI Agents

GUI agents record verified task progress as memory, keeping useful operations past failed attempts.

Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness

Learned attention gates separate trusted instructions from untrusted text, blunting prompt injection.

LCM-Promo: An LLM-Based Uplift Model for Promotion Allocation

Fine-tuned commerce model learns promotion response curves from logs to allocate discounts.

Can LLMs Uncover Hidden Physical Laws through Free Exploration?

Benchmark where LLMs explore a world with changed physics and must notice the odd behaviour themselves.

Meta Models: Models that Model Learning

Models that predict how other models will learn, trained on past experiment records.

Compote: Collapse-Resistant Multiple Hypothesis Learning For Sequential Recommendation

Keeps multiple plausible next-item predictions in sequential recommendation instead of collapsing to one.

Ground-to-Satellite Knowledge Distillation for Crop Classification under Temporal and Geographic Shift

Distills ground-level field photos into satellite time series for crop mapping where labels are scarce.

Blindfold: Keeping the Audit Coin Secret from LLM Agents

Shows audit certification of LLM agents also depends on keeping the audit coin secret from the agent.

Exploring the Limits of the Delta Learning Hypothesis: Single-Pair Generator Biases in Direct Preference Optimization

Tests the delta-learning idea in DPO: one model pair as preference data, and where it biases training.

EnergyCal: Taming Loss Spikes in Muon-Family Optimizers for Sparse MoE Training

Lightweight rule that tames loss spikes from Muon-family optimizers in sparse MoE training.

Correctness Conditioning as Selection in Chain-of-Thought Faithfulness Evaluation

Derives the exact bias introduced when chain-of-thought faithfulness is scored only on correct answers.

OpenGray: a benchmark for evaluating the decisions of radiotherapy planning agents

Benchmark of radiotherapy planning tasks with known right decisions, testing when LLM agents should ask for review.

On the Sparsity of Direct Preference Optimization: Weight Disentanglement in the NTK Regime

Theory for why direct preference optimization produces sparse weight updates, in the NTK regime.

Text Generation in Low-Resource Languages Using Human-Curated Vocabularies

Uses human-curated vocabularies to generate text in languages the LLM barely saw in pretraining.

Self-Regularized Flow Language Models

Self-supervised training for continuous flow language models that reuses the denoising trajectory as its own signal.

Prospector: learning a compact graph traversal policy for scientific reasoning

Learns a compact graph traversal policy so LLMs can mine scattered engineering knowledge on demand.

Mechanistic Failure Modes of Weight Pruning in Transformer Language Models

Factorial study of what weight pruning breaks inside transformers, beyond accuracy and perplexity.

What Survives Distillation? A Mechanistic Interpretability Study of GPT-2 and DistilGPT-2

Mechanistic study of what GPT-2 internals survive depth-reduction distillation into DistilGPT-2.

WorkFactory: Turning Professional Intents into Executable Work for Scalable Agent Training

Builds coherent professional tasks from intent descriptions, producing expert demonstrations for agent training.

Coverage Under Contradiction: A Benchmark for Negation-Aware Retrieval

Benchmark for retrievers that must find evidence contradicting what the model already believes.

MLE-World: Predicting Machine Learning Experiment Outcomes with a Learned World Model

A learned world model predicts ML experiment outcomes, so research agents explore more candidates cheaply.

INTR++: INterpretable TRansformer with better Accuracy, Interpretability, and Adaptability

Interpretable transformer that keeps accuracy when it moves beyond its training domain.

Action-Value Calibration via Continuation Sharing for Long-Horizon Agentic Reinforcement Learning

Shares action continuations to calibrate action values, giving fairer credit in long-horizon agent RL.

Does Your Generalist Skill Optimizer Really Incentivize Agentic Tasks? Rethinking Skill Optimization via Bipartite Attribution

Rechecks whether skill optimizer edits really help agents, separating real gains from dev-set noise.

3D-SceneQ: Empowering 3D LLM With Query-Guided Adaptive Pruning and Multi-Modal Feature Enhancement

Prunes task-irrelevant 3D object tokens before the LLM answers, cutting noise and hallucination in scene QA.

SimplexLoRA: Dynamic Rank Updating via Alternating Minimization with Simplex Projection

Grows and shrinks LoRA adapter ranks during fine-tuning through alternating minimization on a simplex.

GraphRAGProv: GraphRAG Meets How-Provenance

GraphRAG answers with how-provenance, so each claim carries its query path as evidence.

Grounding LLM Agents in Business Processes: Process-Aware Knowledge Graphs for Reliable Multi-Step Service Invocation

Grounds LLM API agents in process and schema knowledge graphs, so multi-step service calls stay valid.

Aligning Large Language Models for Safety in African Languages: A Case Study of Hausa

Studies safety alignment for Hausa, testing methods built for high-resource languages in a low-resource setting.

Your Policy Is Secretly a Harness: The Decision Geometry of Agent Interfaces

Asks what an agent policy learns from its harness, and whether that carries to a new harness.

GameVerse: A Minute-Scale Gameplay Dataset for Long-Horizon Interactive World Modeling

Minute-scale gameplay dataset for training interactive world models that stay coherent over longer rollouts.

GradCO: Behavior-Aware Gradient-Conflict Offloading for Text-Based Person Anomaly Search

Offloads conflicting gradients so anomaly search keeps identity cues while separating fine-grained behaviours.

Where Should the Evaluation Budget Go in LLM-Driven Agent Scaffold Evolution?

Measures where the evaluation budget should go when LLM agents evolve their own scaffolds.

Modeling Preference Dependencies with Directed Acyclic Graphs for Narrative-Driven Recommendation

Models narrative preferences as a directed acyclic graph, so dependent user wishes stay hierarchical.

RPSD: Retrieval-Privileged On-Policy Self-Distillation with Optimal Transport

Self-distills LLM recommenders with retrieval-privileged targets and optimal transport, fixing credit-blind training.

ReForest: Interpretability Meets Reasoning in Isolation Forest

Pairs Isolation Forest with an LLM judge so rare numbers become explained, context-aware anomaly alerts.

AnalogScout: Simulation-Efficient PVT-Aware Analog Sizing with LLM Agents

LLM agents size analog circuits with PVT-aware simulation budgets, cutting the corner-simulation time.

CoS-VQ: Coordinated Shared-Scale Vector Quantization for Large Language Models

Shared-scale vector quantization keeps low-bit LLM weights accurate while allowing simple local codebook updates.

Aligning Institutional Power and Operational Ability

Design dimensions that check whether an agent's responsibilities match its real operational ability.

Uncovering Why LLMs Fail: From Internal Mechanisms to Adversarial Reasoning Trajectories

White-box study that traces how ordinary and adversarial reasoning failures develop inside LLM internals.

Agent-Conditioned Routing: Diverse Reasoning Paths Within One Frozen Model

Routes prompts to different experts inside one frozen sparse model, giving diverse reasoning paths without fine-tuning.

Same Verdict, Different Readers: Personalizing Fact-Checking Explanations from Relative Preferences

Rewrites fact-check explanations to fit a reader's background while keeping the verdict and evidence intact.

LinMix: Data Mixing as Linear Domain Attribution

Frames pretraining data mixing as linear domain attribution to set domain weights.

JEPA-Inspired Latent Field Prediction for Sparse-Pilot MIMO-OFDM Channel Reconstruction

Predicts MIMO-OFDM channels in latent space from sparse pilots, avoiding fragile raw-coefficient fitting.

Clearing the FoG: Training KG Agents to Retrieve the Evidence Behind Their Answers

Trains knowledge-graph agents to return the evidence triples behind their answers.

Constraining the Commit Boundary: A DIO/ODI-Informed Authorization Guardrail for Agentic Entity Resolution

Authorization boundary for agentic entity resolution, so weak identity guesses cannot seed their own evidence.

CRISP: Co-designing Pruning and Scheduling for Efficient MoE Inference on Edge Servers

Co-designs pruning and scheduling for MoE inference on edge servers where memory movement dominates.

ETAMem: Decoupling and Aligning Heterogeneous Memory for Long-Term Conversational Agents

Splits mixed memory units and aligns their structure so long-term conversational agents recall the right facts.

M4Bench: Evaluating Procedural Specification for Clinical EHR Derivation Agents

28 clinical data-derivation tasks test whether coding agents can reproduce EHR variable definitions exactly.

Concordat: Typed Sessions for LLM Agents

Python library enforces typed session rules on LLM agents, outside the prompt.

What Carries Forward? Pretraining Disclosure Across Language Model Releases

Audits 41 model releases to see whether later versions keep or drop earlier pretraining disclosure.

Follow the Gradient: Tracing Safety Degradation Back to Training Data

Traces fine-tuning safety loss back to the benign-looking training data that caused it, via gradients.

Teach reality, not ethics: a drop-in continued-pretraining corpus installs coupled human-biosphere-AI welfare dispositions across model families

Continued-pretraining corpus that installs coupled human, biosphere and AI welfare dispositions in model families.

semanticfa: An R Package for Response-Free Semantic Factor Analysis of Psychometric Scales

R package that factors psychometric scale items through language model embeddings, with no responses needed.

A note on data-efficient finetuning of vision foundation models

Notes on data-efficient visual prompting when adapting new foundation model families.

Orthogonal and Parametrized Multi-Attribute Steering of LLMs

Orthogonal, parametrized activation steering so behaviour edits stay strong and comparable in strength.

The Structure Gap: When More Data Cannot Replace Structured Reasoning

Defines problems where known structure beats scale: more data cannot replace structured reasoning.

GTFormer: A Syntax-Free Graph Network Alternative to Transformers for 8192$\times$ Length Extrapolation in Linear Time

Graph network replaces self-attention, giving linear-time sequence modelling that extrapolates to 8192-token lengths.

TeamGym: A Gym for Team Decision-Making by AI Agents with Distributed Information

Environment for training agent teams that must share distributed evidence to decide well.

REGER: Modeling inter-item Relations with Semantic IDs for Generative Recommendation

Models relations between items with semantic IDs so generative recommenders stop training cold embeddings.

Timestral

Family of probabilistic forecasting models from 5M to 209M parameters, topping three time-series benchmarks.

Understanding the limits of latent world models

Probes whether agent world models learn real latent dynamics, or only pass task tests.

Rare, Relevant, and Still Wrong: How Memory Fails LLM Group Decisions

Shows group memory heuristics fail: rare or relevant claims still get dropped before the final decision.

Metacognitive Decoding: Efficient Inference for Diffusion Language Agents

A controller decides how much text a diffusion language agent must generate for each step.

Multi-State Linear Attention with Generalized Second-Order Routing

Multi-state linear attention with second-order routing for cheaper long-sequence language modelling.

MuonGS: Pretraining LLMs with Relaxed Weight-Norm Control

Relaxed weight-norm control improves LLM pretraining, explained through accumulated effective learning rate.

Attention as Lattice Gauge Theory: Relations Interact Only Through Wilson Loops

Models relation composition in attention as lattice gauge theory, where only closed loops are meaningful.

ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

Learns dynamics from imagination over human video, lifting robot policies without more action labels.

GEPA-Curator: Curated Learning for Self Improving Agents

Curates which agent failures are worth learning from, separating prompt weakness from harness limits and noisy judges.

Watermark Basins: Black-Box Detection of Distilled LLMs

Black-box method to detect when a student model was distilled from a watermarked teacher.

GPA-VGGT: Geometry-Preserving Acceleration of VGGT

Accelerates VGGT geometry models on long view sequences while preserving camera and depth accuracy.

RePAIR: Inverting Backdoor Attacks in Semantic Segmentation

Inverts backdoor attacks in segmentation models by recovering the trigger and the redirected victim pixels.

V-LUMEN: Visual Lookup Memory for Embedding Scaling in Vision-Language Models

Adds lookup-based visual memory to vision-language models, scaling capacity without dense compute.

GraftRay: Physically Based Mesh Insertion into Gaussian Splatting Scenes via Path Tracing

Path tracing relights inserted objects in Gaussian splat scenes so they cast and receive real shadows.

Zooming without Zooming v2: From Seeing Details to Connecting Evidence for Generalizable Multimodal Perception

Links fine detail with evidence connection to generalize spatial perception, using cheaper shared training data.

Learning to Look Beyond Screens: Transferable Visual Observation from GUI Grounding

Trains GUI grounding to teach transferable visual observation behaviours that work beyond screens.

TriDifNet: A Unified Three-Task Diffusion Network for End-to-End RGB-T Perception

One diffusion network handles infrared-visible alignment, fusion and detection together for hard lighting.

SplatScout: Query-Driven Part-Level Referring Segmentation in Gaussian Scenes

Query-driven loop segments object parts in Gaussian scenes instead of picking from fixed candidates.

A First-Step Bias Restores Lost Diversity in Diffusion Model Distillation

Restores lost diversity in few-step distilled diffusion models by re-weighting the first noise step.

Multibranch Ensemble Head in Vision Foundation Models for Efficient Cucumber Phenological Stage Classification

Multibranch ensemble head adapts DINOv2 to tiny crop datasets for cucumber stage classification.

MRI-to-CT Translation of Spinal Images via Anatomy-Guided Pseudo-Pairing and Adversarial Diffusion

Synthesizes spine CT from MRI with anatomy-guided pairing, avoiding extra radiation for surgical planning.

Physics Guided Diffusion Model Based Exposure Interpolation for High Dynamic Range Imaging

Physics-guided diffusion interpolates between two exposure images to avoid ghosting in single-shot HDR.

MixSIS: Mixed-supervised instance segmentation

Instance segmentation that mixes box, point and image supervision instead of picking one annotation level.

Beyond Parameter Protection: Fisher-Shaped EWC for Replay-Free Continual Learning

Reshapes Fisher protection in EWC so it still holds after finite parameter movement.

Variational Optimal Control for Diffusion Guidance

Links diffusion sampling guidance and reward adaptation through one shared variational control framework.

GCSR-OCC: Geometry Calibration and Semantic Routing for Open-Vocabulary Occupancy Prediction

Calibrates Gaussian geometry and routes semantics for open-vocabulary occupancy grids without category limits.

FastBEVFusion: Simple Multi-Modal Fusion for 3D Object Detection at 50 FPS

Simple LiDAR-camera fusion that runs 3D object detection at 50 frames per second.

VGIQA: Fine-Grained Image Quality Assessment with Visual Grounding Rewards

Rewards image quality checks with visual grounding, so multimodal models justify their judgements.

How Much Can Best-of-N Help a World Action Model? Measuring Selectable Headroom and Realizing It in Closed Loop

Measures how much best-of-N candidate selection can improve a world action model, then uses it closed loop.

Uncovering Shortcut Amplification in Video RLVR through the Lens of Perceptual Uncertainty Collapse

Shows video RLVR can amplify text-prior shortcuts when visual evidence is weak.

Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution

Retrieves motion primitives during execution so VLA policies resist out-of-distribution visual shifts.

Zhengrong-Spatial-Aware 1.0: Advancing 3D Multimodal Intelligence

Two adaptation pipelines fine-tune 3D-aware VQA models on a custom spatial dataset with point maps.

TopoField: Technical Report for ECCV 2026 SceneFun3D Challenge

ECCV challenge report that turns a point-wise probability field into ranked instance masks for interactive parts.

Subspace Clifford Networks

Grouped subspace interactions replace global cyclic rolling in Clifford-style channel mixing.

Doubly Robust Deep Kernel Tests for Structured Outcomes

Doubly robust kernel test for whether two interventions produce the same structured outcome distribution.

Decision Models as Judges for Hallucination Detection

Cheap decision models replace free-form LLM judges for hallucination detection, returning verdicts with probabilities.

HDVS: semi-supervised semantic segmentation via heterogeneous dual-branch voting supervision

Dual-branch voting supervision that distrusts single-branch pseudo-labels in semi-supervised segmentation.

REGLAT at ArGuard Shared Tasks: An Ensemble-Based Model Integrating NER for Harmful Content Detection in Arabic

Two-stage Arabic harmful-content classifier that first extracts harm entities, then sorts prompts into seven categories.

Taming Noise Initialization for Zero-shot Video Restoration with Image Diffusion Priors

Tunes noise starts so image diffusion models restore video without video pretraining.

TRUCO: A Scalable Lock-Free Algorithm for Parallel Contour Extraction

Lock-free parallel contour extraction that beats sequential Suzuki-Abe on high-resolution masks.

MvT: Event-based Multi-view Projection for Multiple Object Tracking

Multi-view projection of event-camera data for multiple object tracking, avoiding pseudo-frame tricks.

A Lightweight Attention-Based CNN with Saliency-Map Interpretability for Skin Cancer Detection

Lightweight attention CNN with saliency maps for explainable skin cancer classification from dermoscopy images.

Hierarchical Sparse Graph for Multi-dimensional Assignment in Event-based Tracking

Wide-window assignment graph for event-camera tracking that survives long occlusions and missed detections.

ScIntBench: Physical Interaction Plausibility in 3D Scene Generation

Benchmark that tests whether generated 3D indoor scenes work when you actually interact with them.

Normative Deviation Modeling for EEG-based Depression Detection

Detects depression from EEG by modelling deviation from a healthy norm, across heterogeneous datasets.

TimeLineXR: An Interactive Exploration of Time in Vision-Language Models

Interactive tool to explore how vision-language models encode time inside their embeddings.

PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations

Adds lightweight physical dynamics tokens to a vision-language-action model, avoiding slow video generation in the loop.

CTBakeGS: training-free baking of photorealistic CT volume renderings into 3D Gaussian splats

Training-free bake of photorealistic CT volume renderings into Gaussian splats that run on weak devices.

Displacement Moments and Path-Terminal Separation in Brownian Interpolant Sampling

Proves sampling error bounds for Brownian interpolants in terms of displacement moments.

Beyond the Readout: A Latent Numerical Frame Shared Across Modality and Operation

Asks whether multimodal models share one internal number concept across text and image arithmetic.

When Visual Evidence Is Insufficient: How Instructions, Post-Training, and Internal Activations Shape Refusal in Multimodal Language Models

Studies when multimodal models refuse unanswerable image questions, and what shapes that refusal.

Segmentation Confidence for Arbitrary CNNs

Adds a per-pixel confidence measure to existing medical segmentation CNNs, with no retraining.

Representing the Unseen: Rethinking Visual Imagination for Spatial Reasoning

Argues imagined states should feed reasoning as maps or graphs, not always as generated RGB views.

PerceptRL: Making Visual Perception Rewardable in MLLMs from Answers to Interactions

Makes visual perception a reward signal during multimodal post-training.

QD-PINN:A Physics-Informed Neural Network Approach to Bayesian Quickest Detection in High-Dimensional Diffusions

Physics-informed neural networks estimate optimal stopping boundaries for quickest detection in high-dimensional diffusions.

ALADDIN: Adversarial Latent Diffusion Distillation for One-Step Identity Swapping

One-step face swapping by distilling diffusion quality into a fast adversarial model.

Which Hand Was Late? Diagnosis for Robot Failure Recovery in Bimanual Tool Use

Names bimanual tool failures, like slips and mistimed gestures, that single-arm benchmarks miss.

GroundControl: Satellite-Conditioned Multi-View Diffusion for Reconstruction-Free Novel View Synthesis

Conditions multi-view diffusion on satellite views so novel views stay plausible outside the input set.

Gym-Anything-RT: Scaling Real-Time Environments for Vision-Language Agents

Tests vision-language agents in fast settings like games, where the world changes while they think.

The Noise Plane: A Diffusion Bridge Across Exponent and Dispersion

Diffusion framework for sensor noise, replacing Gaussian assumptions with signal-dependent Tweedie noise models.

Spectral Braids: A Topological Descriptor for Dynamic Directed Networks

Tracks moving graph eigenvalues as braids, giving a topological descriptor for changing networks.

Zero-Shot Multimodal Event Spotting in Soccer Broadcasts

Zero-shot spotting and timing of soccer events in full broadcasts, with no soccer-specific training.

MoI-Bench: A Benchmark for Image Generation and Editing on Mobile Devices

Benchmark of on-device image generation and editing on mobile, built from real user needs.

Beyond Answer Changes: Matched Same-Class Controls for Local Audio Evidence

Matched same-class sound edits separate real audio evidence from waveform artefacts in audio-language model answers.

VideoLVR: Latent Visual Reasoning over Grounded Video Evidence

Keeps video evidence reachable during reasoning, so multimodal models stop forgetting what they saw.

Perceiving and Predicting Dynamic Sports: Spatiotemporal Intelligence in Vision Models for Egocentric Tennis Video

Annotated egocentric tennis video tests how well vision models perceive and predict fast sports action.

Know, then Draw: A New Learning Paradigm for Visual Generation

Splits visual generation into autoregressive knowledge pretraining and diffusion rendering for high-fidelity images.

What Improves During Pretraining? Tracking Knowledge Acquisition with Point-in-Time Models

Tracks when general and domain knowledge appears during pretraining with point-in-time web models.

Repurposing Video Diffusion Transformers for Cross-View Temporal Object Correspondence

Reuses video diffusion transformers to keep object identity across camera views and time from a query mask.

Pushing Dexterity and Safety for In-Hand Object Pose Reaching That Works Remarkably Well

Studies dexterity and drop safety together for in-hand object pose reaching with multi-fingered hands.

Honeycomb-Surfels: Compact Hexagonal Primitives for Efficient Scene Rendering and Reconstruction

Hexagonal surface elements reconstruct 3D scenes with less memory and faster training than Gaussian splatting.

Can Discrete Tokens Serve as Semantically Lossless Representations of Visual Signals?

Asks whether discrete visual tokens lose semantics, and argues the loss may be small.

From Ethernet to Wi-Fi: Flow Feature Selection and Construction for Traffic Classification

Five-stage feature procedure that transfers traffic classifiers trained on Ethernet to Wi-Fi flows.

Evaluating Priority Dispatch for Premium Membership in On-demand Taxi-sharing

Simulates premium membership dispatch for taxi-sharing using deadline shortening and dedicated vehicles.

Crop Yield Prediction Using Regression Models: A Comparative Evaluation of Model Performance

Compares five regression models for crop yield prediction on a cleaned million-row farm dataset.

Learning to Code the World: Harness Evolution for Compositional 3D Reconstruction

Represents 3D scenes as code and evolves the harness around a small open agent to rebuild them.

Hierarchical Low-Rank Reparameterization for Parameter-Efficient

Compresses parameter-efficient fine-tuning further with hierarchical low-rank reparameterization.

Systematic Morphometric Brain Autoencoding for Representation Learning in Neurodegenerative Diseases

Self-supervised morphometric brain autoencoder for early neurodegenerative detection across scanners, without many labels.

MCP-Net: Real-Time Coronary Perception Network and Topological Analysis for Robotic Guidewire Navigation

Real-time coronary anatomy and topology network for robotic guidewire navigation under X-ray.

What the Primary Lund Plane Omits: A Model-Stitching Diagnosis of a Jet-Tagger Representation Gap

Model stitching shows the jet-tagger gap lies in the representation, not the downstream head.

Generative Replay for Equilibrium Propagation: Toward Edge-Friendly Continual Learning in Hopfield Networks

Generative replay fights catastrophic forgetting in equilibrium-propagated Hopfield networks on edge hardware.

Sharpness-Aware Muon: Matrix-Aware Optimization with Momentum-Guided Perturbations

Adds momentum-guided sharpness-aware perturbations to the Muon optimizer for flatter minima.

TCM-DiagCommit: A TCM-SD-Derived Benchmark for Evidence-Conditioned Syndrome Selection and Commitment

Benchmark that watches whether TCM diagnosis models update syndromes as evidence fields arrive, then commit.

Low-degree polynomials lower bounds for clustering in small dimension

Low-degree polynomial lower bounds for clustering many clusters in small dimension.

Bayesian Inference of Expensive Simulators: Survey and Comparison of Design Choices

Survey and comparison of Bayesian inference choices for prohibitively expensive simulators.

Multi-Label Agreement: A Flexible Framework for Inter-Annotator Reliability

Flexible framework for inter-annotator agreement when tasks use multiple labels per item.

CodeCap-Bench: Can Code Serve as a Good Video Caption?

Tests whether executable code can serve as a compact, precise video caption.

The Geometry of Manifold Learning: Singular foliations via Deep Learning models

Deep learning study of singular foliations as the intrinsic geometry of training data.

Euclidean Crowding-Driven Adaptive Euclidean-Hyperbolic Graph Learning

Adaptive Euclidean-hyperbolic graph learning that shifts representation space by how crowded the hierarchy gets.

Channel Selection for Mixture-of-Experts Decoding Needs No Calibration Data

Chooses which MoE expert channels to load at decode time without any calibration data.

FeatureArc: Benchmarking Coding Agents on Evolving Features

Benchmarks coding agents on successive feature extensions to their own earlier code, not single fixed requests.

HGPRGDel: GPR-Aware Learning and Ranking for Growth-Coupled Gene Deletion Design

Ranks growth-coupled gene deletion designs with Gaussian-process ranking over metabolic models.

Regression Functions as Fields on Covariate Space: Evolution Equations, the Hessian, and Identification

Treats the regression function as a field on covariate space and derives its evolution equations.

ReferStream: Interactive Subject Injection for Streaming Video Generation

Injects new subjects into live video generation without breaking the scene or the subjects already there.

As the Story Unfolds: Watching a Film and Identifying Characters as a Human Does

Names film characters the way viewers do, from dialogue and the film itself, with no cast list.

When Should We Intervene at Test Time?

Asks when test-time adaptation helps, and shows average accuracy hides the batches it damages.

AI-Assisted Discovery of CNOT Decompositions for Quantum Permutations on Constrained Topologies

AI-assisted search finds CNOT decompositions for quantum permutations on qubit graphs with tight connectivity.

Quantized Edge Inference of a Self-Supervised SAR-Optical Encoder on Commodity x86 Hardware

Benchmarks a quantized self-supervised SAR-optical encoder on cheap x86 field hardware.

SFE-O: An Ontology for Knowledge-Grounded AI in Supercritical CO$_2$ Biomass Extraction and Fractionation

Ontology for knowledge-grounded AI over supercritical CO2 biomass extraction data.

Physics-Guided Flow Matching: A Generative AI Framework for Global Total Electron Content Forecasting

Physics-guided flow matching forecasts global total electron content a day ahead for space weather.

LiveSearch: Questions as Executable Verifier Code for a Live Benchmark of Professional-Source Financial Search Agents

Benchmarks financial search agents on live sources, with questions written as verifier code.

Rènyi-Regularized Flow-Based Variational Inference for Bayesian Calibration

Renyi-regularized flow inference calibrates Bayesian posteriors that have several separated modes.

Structured Representation Probing for Risk Anticipation in End-to-End Autonomous Driving

Reads a driving policy's internal traffic and ego representations to anticipate failures before they happen.

Scale-aware and spatially-localised optimisation of kilometre-scale MLWP models

Wavelet-based loss keeps fine-scale structure in kilometre-scale weather models without losing accuracy.

Online Interval Selection with Binary Predictions: Optimal Consistency and a Continuous Robustness Tradeoff

Proves consistency bounds for online interval selection when predictions arrive in arbitrary order.

TRIDENT-Seq: Profile-Matched Prefix-Future Contrast for Label-Efficient Payment Risk Modeling

Pretrains a payment risk encoder on unlabeled transaction histories, then adapts it with scarce labels.

How Many and How Much? Contrastive Learning with Negative Samples for Knowledge Graph Embedding

Asks how many negatives and how hard they should be for contrastive knowledge graph embedding.

An end-to-end framework for uncertainty-aware validation of generative models

Framework that quantifies generative model uncertainty and propagates it into a proper validation test.

Test-Time Communication in Trading Agent Swarms

Runs 200 trading agents to ask when independent decision-makers should talk, with no shared goal.

Graph Representation Learning via Adaptive Expectile Aggregation

Learnable expectile aggregation captures class-relevant spread in heterophilic graphs, where means fail.

How Reliable Is Individualized Functional-Network Attribution in Resting-State fMRI Cognitive Prediction in ADNI?

Measures whether individualized fMRI network attributions reproduce across scans in Alzheimer's prediction.

Keeping Inference in the Radio: Design and Measurement of a Spiking Accelerator for mmWave ISAC Activity Recognition

Spiking accelerator runs mmWave activity recognition inside the radio, saving the main processor's energy.

When uncertainty quantification fails: structural vulnerabilities and corrections for selective prediction under clinical class imbalance

Shows clinical uncertainty scores fail under class imbalance, and tests fixes for deferral systems.

Pneuma: A Large-Scale Real-World Time Series Benchmark for Commercial Aircraft Predictive Maintenance

First public large-scale time-series benchmark for commercial aircraft predictive maintenance.

GAS-PFN: Graph-Aware Spatiotemporal Prior-Data Fitted Networks for Zero-Shot Traffic Forecasting

Zero-shot traffic forecasting with graph-aware prior-data networks that transfer to new road networks.

Guided-RLVR: Guided Exploration in Reinforcement Learning with Verifiable Rewards

Guides exploration in RLVR training so language models keep varied reasoning paths instead of collapsing.

When Is a Relation Worth Trusting? Manipulation Costs, Exposure, and the Trust Trap in Multi-Relational Graph Learning

Shows when an attacked relation is worse to use than to ignore in multi-relational graph learning.

UCVE: Graph Structured Retrieval for Compliance Validation

Graph retrieval that checks compliance obligations for coverage instead of treating each requirement alone.

WCC: A Learning-Based Framework for Deciding When to Apply Cooperative Coevolution in Large-Scale Optimization

Learns when to switch between cooperative coevolution and non-decomposition in large-scale optimization.

Game-Theoretic Solvers with Bayesian Opponent Modeling and Moral Rhetoric Framing for GLEE

Bargaining agent combining a game-theoretic solver, online Bayesian opponent modelling and moral framing.

Multi-Agent Reinforcement Learning of Co-Adaptive Hip Exoskeleton Control (MARL-Exo) for Metabolic Energy Reduction

Trains hip exoskeleton control with multi-agent RL against a metabolic energy reward, then tests new users.

Interpretable Meta-Policies for Neural Decision Making

Hierarchical meta-policies that keep neural control decisions readable by translating them into symbols.

Autonomous fixed-wing interception using reinforcement learning

RL flies cheap fixed-wing interceptors against medium-altitude strike drones that quadcopters cannot reach.

Strategyproof Deep Reinforcement Learning for Online Resource Allocation

RL resource allocation that also keeps agents truthful about their valuations.

Kinematic Constraint-Referenced Evaluation for Articulated Object Manipulation

Scores manipulation against the object's own joint axis, separating real progress from off-axis motion.

EnvSched: Environment-Aware Scheduling for Rollout-Efficient Agent Reinforcement Learning

Schedules task instances and their environments together to make agent RL rollouts cheaper.

Interpretable Deep Learning Models for Breast Cancer Diagnosis Using Medical Imaging

Benchmarks ten vision backbones with interpretability checks for benign versus malignant breast imaging.

Hardware-Matched Spiking Neural Networks for Real-Time Purkinje Cell Spike Classification

Hardware-matched spiking networks classify Purkinje cell spikes on untethered headstages.

Retrospective macromolecular pool fraction mapping from conventional clinical MRI: a physics-guided self-supervised deep learning framework

Estimates macromolecular pool fraction from routine clinical MRI with physics-guided self-supervision, no special scan.

Para-Gaussian Approximations of Random Deep Neural Networks

Slightly non-Gaussian approximation for random deep networks, where Gaussian limits degrade with depth.

Learning Quality-Controllable Robot Policies with Iterative Hindsight Relabeling

Relabels robot rollouts so policies learn controllable quality like speed and smoothness, not just success.

An Input Skip Can Reparameterize a VAE Without Improving the Generator

Proves an input skip can reparameterize a Gaussian VAE without changing its generated samples.

When, Where, and What to Correct: Human-Guided Adaptive Residual Correction for Robot Policies

Human-guided residual correction fixes the failure points of a robot policy without retraining the whole thing.

KPPax: preprocess differentiable kinetic ODE for machine learning in atmospheric chemistry

Kinetic preprocessor that turns atmospheric chemistry mechanisms into differentiable, compiled code for ML models.