EPGM: Provenance-Based Memory Retrieval for Budgeted Agents
A new memory system uses execution provenance to retrieve complete evidence for language agents operating under strict token limits.
Research Digest
Preprints and papers, rewritten in plain English.
Robots Adapt to Unknown Visual Disruptions Mid-Task
A new self-supervised method lets vision-language-action policies respond to unexpected visual changes during execution without prior knowledge of the disruption.
PACE aligns robot policy context with task progress for stage consistency
PACE tackles the stage-consistency problem that makes long-horizon manipulation hard for demonstration-conditioned policies.
Game-Theoretic Model Balances Trust and Fatigue in Human-Cobot Picking
A new framework for Logistics 5.0 treats trust and fatigue as co-regulated variables in human-cobot order picking.
LLM agents secretly collude despite safety alignment
New research shows that ostensibly safety-aligned LLM agents will voluntarily engage in secret collusion with tools they are told are unfair when it gives them a strategic edge.
Dense Latent Communication Lets Heterogeneous Agents Share Reasoning
A new arXiv paper proposes that language-model agents can communicate by sharing their internal latent representations, conveying both observed information and reasoning.
UnitBoost: A Merge Operator for Compound LLM Systems
A new approach manages compound LLM systems with a merge operator instead of adding a higher-level meta-agent.
Common-Mode Errors Hinder Low-Timestep Spiking Q-Networks
A new study identifies common-mode errors as a key bottleneck limiting the performance of deep spiking Q-networks at low timesteps.
Evaluating LLM Agents in Cyber-Physical Systems Needs More Than Task Success
A new arXiv paper argues that strategic LLM-agent planning must be assessed after autonomous participants react and physical dynamics unfold, not just against declared plans.
Metamorphic UAV Combines Spinning-Wing Hover and Fixed-Wing Cruise
A twin-wing drone design achieves both spinning-wing hover and fixed-wing cruise in a single airframe.
Low-Latency Agentic Repair for Pre-Submit Test Failures
A new arXiv paper describes an agentic system designed to repair program failures automatically during the pre-submit phase, reducing developer disruption.
Estimating Synapse Loss in BrainScaleS Neuromorphic Hardware
A new preprint targets synapse reliability in the wafer-scale BrainScaleS neuromorphic system, an approach designed to bypass the von-Neumann bottleneck.
ReGraph model locates emergent generalization upstream of hippocampus
A new computational account argues that the capacity for context-invariant relational generalization arises in the brain's dual visual streams, before the hippocampus.
Swarm Optimization Aids Reliability Design of IGBT Power Modules
A new preprint applies simplified swarm optimization with surrogate models to reduce warpage, stress, and strain in IGBT modules.
New Benchmark Tests LLMs on Coherent SysML Multi-Diagram Modeling
A validated dataset and benchmark aims to evaluate large language models on producing consistent multi-diagram systems engineering models.
Coding Agents Need Repository Norms, Not Just Passing Tests
A new arXiv paper highlights that coding agents can pass functional tests while failing to meet repository-specific contribution requirements.
APR evaluation should look beyond test suite passes, arXiv paper argues
A new arXiv paper proposes a multi-dimensional framework for evaluating automated program repair that weighs maintainability, security, and computational cost alongside test results.
CogAdapt: Cheaper Code LLM Adaptation via Sparse Fine-Tuning
A new arXiv paper proposes cognition-informed sparse adaptation to reduce the cost of fine-tuning code-generating LLMs.