DialogMem: A Benchmark for Dialogue-Grounded Memory in Embodied Agents
The paper introduces a new evaluation approach for embodied AI that tests whether agents can remember and use information from human conversations alongside visual observations.
Research Digest
Preprints and papers, rewritten in plain English.
Human-AI Complementarity Proposed to Ease AI Oversight Burden
A new arXiv paper argues that as AI outputs grow longer and more complex, human oversight can be made safer by leveraging complementarity between humans and AI.
When Agent Skills Clash: A One-File Fix for Coding Agents
New research examines how coding agents behave when multiple instruction files for the same task are active, and proposes a simple remedy.
CorrectRTL improves LLM-based RTL generation via decomposed self-correction
A new arXiv paper proposes a self-correction framework that decomposes instructions to boost the quality of specification-to-RTL code generation.
CoworkEnv: Synthetic Training Environments for Long-Horizon AI Coworkers
A new arXiv preprint proposes CoworkEnv, a framework for synthesizing training environments that help AI agents learn to coordinate tools, communicate across turns, and adapt to changing workspace states.
DAGLoop Aligns Method Design with Experimental Feedback
A new framework aims to improve automated AI research agents by more directly linking how candidate methods are designed to the results of experiments.
SteadyThought Curbs LLM Under-Thinking with Thought-Level Optimization
A new training method targets thought-switching behavior to help large reasoning models think more steadily and deeply.
Distributed ADMM Method Aims to Decentralize Multi-Agent Training
A new approach applies ADMM to train collaborative multi-agent language models without relying on centralized coordination.
LLM Agents Misallocate Noisy Experiments, New arXiv Study Finds
A new preprint shows that AI agents, when faced with noisy experimental readings, tend to mismanage their measurement budget between exploration and confirmation.
Agentic RAG for Retail: Chaining Evidence Across Mixed Data
A new arXiv paper argues that enterprise retail workflows need agentic retrieval-augmented generation that can chain evidence across structured and unstructured data, a capability current benchmarks overlook.
PLACE framework manages memory validity and query-time use for agents
A new approach called PLACE gives persistent-memory agents explicit control over when stored state stays valid and when it should influence a later query.
Spectral Graph Transformer Cuts Cubic Cost of Traffic Forecasting
A new architecture combines spectral graph convolution with a transformer to make traffic forecasting scalable without the costly eigendecomposition of the graph Laplacian.
Building Logic-Invariant LLMs to Move Beyond Shallow Reasoning
A new arXiv preprint asks whether LLMs truly understand logic and proposes a method to make them invariant to surface form.
Predicting High-Impact Research by Linking Unconnected Concepts
A new approach uses multi-agent reasoning over concept graphs to forecast which novel combinations of scientific ideas will have high impact.
Classifying, not generating, improves stepwise reasoning verification
A new approach treats stepwise verification of LLM reasoning as a classification task, potentially offering a more reliable alternative to generative verification.
Reasoning correctness is decodable but not steerable, study finds
New experimental evidence suggests language models encode reasoning correctness in a linearly decodable form that does not yield a causally useful steering direction.
Test-Time Search Boosts Decision Models 15-Fold Without Retraining
A new arXiv paper shows that adding test-time search to frozen decision models yields a 15× improvement in predicted probabilities.
Modeling the Review Step in Human-AI Shared Autonomy
A new paper argues that standard decision-theoretic models overlook the common step where one agent evaluates another agent's proposed action before execution.