Sunday, 1 Nov 2026

PCM: A Proof-Carrying Message Protocol for Verifiable Multi-Agent Communication

Proof-Carrying Messages give every agent claim a Merkle-log proof and a signature. On 926 messages it accepts 10 stale lies, while plain A2A accepts all 712.

Sandman: Building and Operating Sandboxes for Agentic Reinforcement Learning on Shared Kubernetes

Amazon runs agent reinforcement learning on shared Kubernetes with sandbox pods. Four SWE-bench Verified campaigns created 3.1 million sandboxes in October 2026.

AgentDecept: Benchmarking Deceptive Behavior in Tool-Using LLM Agents under Pressure

Users see the agent's report, not its API calls. AgentDecept checks reports against execution on 477 pressure variants, separating falsification, concealment and equivocation.

Where Do Vision Language Models Fail at 3D Reasoning?

Vision-language models get judgments right but reuse 3D properties inconsistently. On per-color counting, humans reach 93.3% while the best model reaches 7.2%.

OPERA: OPamp Design by an Expert-Taught Reasoning Agent, from Specifications to Layout

An LLM agent designs CMOS opamps from specifications to layout with open-source tools. All six layouts are DRC-clean and meet every target at 15 PVT corners.

Online Dense Reinforcement Learning with Planning-Grounding Agent for UI Navigation on Live IOS Apps

The agent trains on live iOS apps rather than memorized static trajectories. It uses LLM-judge rewards, PPO with step-level advantages, and GiGPPO value estimation.

Well-Specified and Well-Tested? Assessing SWE-bench-Style Datasets with a Self-Evolving Agent

SWE-bench-style datasets still contain unclear tasks and weak test coverage. ExLabel uses a self-evolving experience bank to label clarity and coverage across eight languages.

There Is No Free Lunch in RLVR: Every pass@1 Gain Has a pass@k Price

RLVR raises pass@1 while the base model overtakes it at large sample budgets. A theorem settles this: the reversal is a price, and its exact rate is computed.

WorldFlow: Joint Video and 3D Motion Prediction via Large-Scale Pretraining

A unified model predicts future video and query-based 3D trajectories together. Trained on 1.44 million samples, it has the lowest ADE and FDE among tested models.

Beyond LLM Wrappers: Context Engineering and the Three-Engine Architecture for Reliable Financial AI

Three engines handle finance: knowledge graph, deterministic optimizer, multi-agent LLM layer. Over 15 deployments, observed hallucination fell from about 7-9% to below 1%.

Clean Sheets and Checkpoints: Evidence Complexity for Certifying Sequential Agent Risk

Models how many zero audit responses certify bounded agent completion risk.

REALM: Identifying Minimal Task Assets for LLM Agents

Maps each task's required assets, cutting prompt-injection success from 16.6 percent to 0.2 percent.

PRISM: Personalized Item Representations in User History

Item embeddings in history express each user's relationship to the item.

EmbodiExp: Compositional Experience Transfer for Embodied Agents

Builds transferable experiences and composes them online during embodied task execution.

Goal-Aware Dynamic Multi-Agent Collaboration for Long-Horizon Tasks

Keeps each agent's goal and acceptance conditions visible in a growing execution graph.

SocialArena: Evaluating Social Intelligence in Large Language Models

Benchmark that defines and evaluates social intelligence of language models systematically.

Where to Learn and How Far to Roll Out: Efficient Adaptive Training for GUI Agents

Two-stage adaptive training skips mastered prefixes and rolls out only where rewards inform.

LLM-Assisted Optimization for Simulation-Based Redesign

Uses LLMs with engineering simulators for global redesign of engineering solutions.

MemSteer: Turning User Histories into Supervision for Agent Memory Optimization

Turns unlabeled user histories into QA supervision for memory design.

What Does "Now" Mean to Your Computer-Use Agent? Stale Screenshots, Page Loads, and Which Check to Run Before a Click

Finds planned clicks go stale mainly from page loading, not planning time.

Telling Is Not Seeing: Textual Personas Rarely Make Vision-Language Models Answer Like the People They Describe

Persona prompts rarely make vision-language models answer like the people they describe.

RetroChessBench: A Benchmark for Reconstructing and Proving Chess Histories

Benchmarks frontier models reconstructing legal chess histories and proving forced events.

Cord: Confidence-Anchored Order Residual Decoding for Masked Diffusion Language Models

Prompt-conditioned reveal policy orders masked diffusion decoding by confidence and future usefulness.

What Makes Medical Examination Items Difficult? Using Large Language Models to Explore Patterns Associated With Empirical Item Difficulty

Uses LLMs to induce rubrics that explain empirical medical item difficulty.

Verifiable LLM-Guided Focal SMT Solving for Quantified Arrays

Lets LLMs guide quantified-array SMT solving through a solver-verifiable interface.

ATLAS-Bench: An Agentic Tool-Use Benchmark for Local Area Search

Benchmarks agent local search on 1,200 queries with constraints, preferences and evidence.

Integrating Heterogeneous Echocardiography AI Outputs in Language Space: Learnable Late Fusion with a Supervised LLM

Late-fusion supervised language model combines heterogeneous echocardiography AI outputs in language space.

Does Scaling Toward AGI Depend on What System-1 Lets System-2 Execute?

Asks if AGI scaling can shift from a heavier System-2 to a lightweight gating System-1.

What Breaks When Facts Become Stories? The Distortions Source-Agreement Metrics Miss When Models Retell Real Events

Argues source-agreement metrics miss intended style and real distortion when models retell events.

Agent-Wise Test-Time Normalization for Camera-Based Cooperative Perception with Corrupted Collaborators

Per-agent normalization statistics stop corrupted collaborators from degrading clean ego detections.

Category-Anchored Large Language Models for Next POI Recommendation

Anchors LLM next-place recommendations in categories to avoid generating invalid identifiers.

Judging Before Propagation: Task-Graph Contextual Integrity for Multi-Agent Privacy

Extends contextual integrity from single transfers to graph-level flows in multi-agent workflows.

Non-Zero Radius: The Missing Component of Neural Networks

Proposes non-zero radius activation as a missing component of neural networks.

LLM Agent-Based Design of RISC-V CPUs with FPGA-in-the-Loop Evaluation

LLM agents implement CPU candidates on FPGA and use on-board measurements to guide design.

Generative artificial intelligence prompting strategies among Chinese university students across academic tasks

Surveys 234 students; prompt strategies barely change across task types.

Red-Teaming Large Language Models in Low-Resource African Languages: A Case Study of Yoruba and Nigerian Pidgin

Models that reject English attacks often fail to understand Yoruba prompts instead.

From Intent to Improvement: Agentic Dataset Search for Better Model Training

Contract-governed dataset search that types every model decision and verifies constraint evidence.

Letting AI Formulate the Research Question: Autonomous Research Agents

Lets a coding agent formulate its own research question and reports every run.

Can AI Agents Interface Clinicians and Complex Models?

Studies agents that present complex data models to clinicians in hospitals.

Heterogeneous Co-Evolution Learning for Tool-Efficient Search Agents with Zero Labels

Co-evolves training tasks and solver to train tool-efficient search agents without labels.

Practical Assessment of Agentic Systems in Acoustics

Measures how model and harness choices shape agents on two acoustic engineering tasks.

MoLRSI: A Recursive Self-Improving AI agent for Molecular Discovery

Self-improving agent that optimizes molecular design workflows and model architecture together.

Bank-World: A Benchmark for Evaluating Agents on Banking Workflows

Benchmark of 50 banking tasks testing permissions, reconciliation and status reporting by agents.

PAPERBENCHX: Can AI Reproduce Science Across Disciplines?

Benchmarks whether AI can reproduce published science across many disciplines.

Finding Is Not Returning, Returning Is Not Reconstructing: A Multi-Dimensional Evaluation of Agent Social Re-entry

Splits agent social re-entry into finding, returning and reconstructing, then measures each.

Long-Term Agent Memory through Complementary Curation and Consolidation

Combines episodic and relational memory, scoring 90.2 percent on LoCoMo and 85.5 percent on LongMemEval-S.

The Role of Self-Concept In Language Models

Installs five meanings of 'I' into language models to test behavioral effects.

Revisiting Over-refusal in Safety-aligned Large Language Models: A Representation and Optimization Perspective

Traces over-refusal to imbalanced gradient coupling, then orthogonalizes gradients with contrastive learning.

InferD: A Honeypot Framework for AI Inference Infrastructure

Honeypot framework that measures attacks against public AI inference and agent infrastructure.

EvoTeam: Multi-Agent Collaboration via Evolving Teams

Evolves team organization, interaction rules and agent skills together across heterogeneous tasks.

Aligning Medical Vision Language Models at the Capability Boundary with Entropy Adaptive Rubric Grounding

Aligns medical vision-language models with entropy-adaptive rubric grounding on hard diagnostic cases.

Incorporating Machine-Others: An Enactive Exploration of Human-LLM Interactions

Proposes interactional theory for how linguistic exchanges with LLMs change their users.

Turing LLM Post-Training: Calibrated Distillation and Efficient Dynamic-MoE Execution

Integrates MoE specialists with calibrated distillation; consolidation runs 2.86x faster.

KARMA: Training-Free Knowledge-Atomic Harness Evolution for Industrial Advertising Agents

Training-free framework that evolves six typed knowledge modules for industrial advertising agents.

Prompts and probabilities in the linguistic assessment of language models: Evidence from semantic focus-sensitivity

Compares prompting with token-probability scoring for judging semantic focus-sensitivity in language models.

Learning Calibrated Confidence for Multi-Agent Collaboration

Each agent emits a calibrated correctness probability for its answer at handoff.

Appreciation Teaches Unauthorized Execution: Self-Evolution Degrades the Robustness of Personal Agents to Indirect Prompt Injection

Learning from user feedback can weaken agents against indirect prompt injection. Skill evolution raised attack success by 14.0 to 48.8 points against the base agent.

Generating Critic-Guided Synthetic Experience in Latent World Models

Uses a world model plus critic guidance to generate synthetic training experience.

From Local Recovery to Unified Generation: Discovering Triple Set in Knowledge Graphs via Observation-Anchored Latent Diffusion

Recovers knowledge graph triples with observation-anchored latent diffusion over whole sets.

DeCoRL: Decomposed Consistency Reinforcement Learning for Multi-Image Composition

RL reward decomposes multi-image composition into object-level consistency feedback.

Proving and Improving the Platonic Representation Hypothesis with JEPAs

Proves a theorem that separately trained JEPAs converge to shared representations.

UMI-Live for In-The-Wild Bimanual Robotics Data Collection

Open-source bimanual demonstration system for collecting robot data in the wild.

SymKL: A Unified Representation-Space Measure of Signal Degradation

One degradation measure across modalities, built on fitted neural representations.

FDSML: Formula-Driven Supervised Metric Learning with Parameter-Aware Triplet Loss

Pre-trains vision models from formula parameters using metric learning instead of classification.

Timbre: Reliability-aware harmonic consensus for label-free multimodal test-time rexinforcement learning

Harmonic consensus weighting counters confident wrong answers in label-free test-time RL.

DEG: Diffusion-Enhanced GFlowNet for Solving Vehicle Routing Problems

Adds diffusion-enhanced flow expansion to GFlowNets for vehicle routing solutions.

Raincheck: Deferred Multimodal Prefill for Efficient Multi-Turn MLLM Inference

Defers multimodal prefill token selection until later turns reveal what matters.

Intrinsic Distribution Matching

Improves distribution matching distillation for one-step and few-step text-to-image generation.

Rebuilding a Field's Open-Problem Catalog from the Literature: A Two-Layer Knowledge Graph, and Why Status Must Stay Proposed

Rebuilds a field's open-problem list as data with an evidence layer from literature.

Event-SAMA: Language-Grounded Video Object Segmentation on Neuromorphic Event Streams

First event-camera baseline for referring video segmentation, built by converting RGB corpora to events.

The Importance of Speech Genre for End-of-Turn Detector Training

Compares end-of-turn detectors trained on human-human versus human-machine speech corpora.

AegisFlow: Training-Free Non-Myopic Path-Safe Guided Flow Matching

Training-free guidance that keeps generative trajectories safe along the whole path.

Spectral Invariant Fusion and Refinement for Domain Generalized Semantic Segmentation

Segmentation method that suppresses domain-sensitive interference while preserving fine spatial detail.

Learning to Formulate: Agentic Hardware-Software Co-Design of AI Infrastructure through Progressive Optimization and Verification

Agent writes formulations, a solver optimizes, verification checks; Qwen3-8B layers run 1.075x faster.

Soei: Structured Observability of Evidence and Intent for Multi-Agent AI Software Engineering

Audit of coding-agent sessions finds 23.4 percent of edits end without any independent check.

Compile Away Uncertainty: Self-Evolving Agent Skills via Programmatic Compilation

Compiles well-specified agent skills into validated programs, keeping reasoning only where needed.

The Real Cost of Context: Predicting Compute Time and Energy Consumption in LLM Prefill and Decode Phases

Profiles prefill and decode time and energy; prefill scales with input length.

LiDAR-SHIELD: A Fully Dockerized Adversarial Robustness Assessment Framework for LiDAR-Based 3D Perception Models

Dockerized framework that stress-tests LiDAR perception models against adversarial attacks through HTTP endpoints.

Exploring the Potential of Symbolic Regression in Precision Livestock Farming

Tests symbolic regression on sow sensor data to predict farrowing time on farms.

Before the Next Rollout: Diagnosing Skill Utility Shifts from Policy Updates

Uses policy-update token shifts to detect skills whose marginal utility declined.

Safe Multi-Objective Reinforcement Learning: Zero Constraint Violation and Primal-Averaging Convergence

Learns Pareto-optimal policies with zero constraint violation across every episode.

SHPRR: An Algorithm for Pure Exploration under Constrained Regret

Best-arm identification under both a fixed sampling horizon and a regret budget.

Reasoning or Pattern Exploitation? Mechanistic Insights from RL-Trained Language Models

RL training raises Tic-Tac-Toe accuracy and robustness in a small model. Sparse autoencoders show heavier reliance on prompt facts than on strategy.

Where Position Matters: Profiling-Guided Layer-Wise RoPE Allocation in Transformers

Assigns rotary position coverage per layer under a fixed budget.

Learning to Bound: Monotonic, Calibrated Neural Upper Bounds for MPE Inference

Learns monotonic calibrated upper bounds to speed branch-and-bound for probabilistic inference.

CoToan: Executable Rules and a Representation Contract for Reproducible Game AI Evaluation

Executable representation contract that makes game AI results comparable across implementations.

Semantic Mismapping in Generalized Category Discovery: Evaluation and Mitigation

Shows evaluation matching hides category mismapping in generalized category discovery, then mitigates it.

Retaining Repair Hypotheses for Cost-Aware Execution

Keeps multiple model repairs and prices tests to choose execution routes.

Kairos: Semantics-Aware Orchestration of Planning Engines across Classical, Numeric, Temporal and Hybrid PDDL

Orchestrates heterogeneous PDDL planners at run time to return validated plans quickly.

Sparse Fine-Tuning for Parameter-Efficient Adversarial Training

Updates only robustness-critical parameters, about 5 percent, during adversarial training.

Does Federated Training Degrade Explainability? A Comparative Analysis of Grad-CAM Faithfulness in Centralized vs. Federated Deepfake Detection

Tests whether federated training weakens Grad-CAM faithfulness in deepfake detectors.

Integrated Feature Selection for classification and Representation Learning with Group-Lasso Siamese Networks

Applies group-lasso sparsity to siamese networks for feature selection and representation learning.

Data Provenance and Ethics in the Curricula of U.S. Health AI Education Programs

Only 22.2 percent of surveyed American health AI programs teach data provenance.

Emergence of the transformer architecture as the score of entropic laws of random measures

Derives the transformer as the score of maximum-entropy laws over sentences.

From the Transformer to a Formal Semantics of Coordinated Transformations

Gives coordinated parameterized transformations a formal semantics starting from the Transformer.

Multi-output LSTM soft sensors for industrial anaerobic co-digestion using LSTM-beta-VAE generative data augmentation

Predicts anaerobic co-digestion process variables with multi-output LSTMs and generative augmentation.

CovFedAvg: Local Low-Loss Curves and Federated Model Selection

Federated selection method that fits a shared model endpoint along clients' local low-loss curves.