Density-Based MPC Enables Safe Multi-Robot Object Transport
A new arXiv paper proposes a hierarchical density-based model predictive control framework for safe collaborative transport by quadrupedal robots.
Research Digest
Preprints and papers, rewritten in plain English.
SGAnalog Benchmark Tests Analog Circuit Skills on Real Tapeouts
A new benchmark built from open-source silicon tapeouts aims to reveal whether analog circuit models generalize or merely recall familiar examples.
Surrogate Log-Probabilities Could Audit Black-Box LLM Agents
A new preprint proposes using a surrogate model's log-probabilities to gauge the reliability of black-box LLM agents whose token probabilities are hidden.
ROAR Framework Unifies Runs Across Heterogeneous AI Research Systems
A new framework called ROAR aims to standardize and preserve data from expensive AI research runs, enabling large-scale evaluation and reuse.
Behavioral History Beats Person Descriptions for LLM Personas
A new arXiv study suggests that feeding LLMs a person's past behavior, rather than trait descriptions, yields more accurate synthetic survey respondents.
Language models may split memory, computation, and execution
A new arXiv paper proposes a finite-element-inspired architecture that separates contextual computation, persistent storage, and exact execution in language models.
How Group-Relative Rewards Shape Exploration in RL
A new arXiv paper examines how group-relative reward allocation in reinforcement learning affects which solutions remain accessible under repeated sampling.
Quantitative Review of Graph Strategies for Cyber Attack Detection
A new arXiv study systematically compares graph construction and representation methods used across cybersecurity domains to identify patterns and gaps.
New Benchmark Tests LLM Agents' Ability to Revise Executable Skill Packages
SkillScriptBench evaluates whether language agents can update executable skills without breaking existing behavior.
Quantizing Video VAEs Without Breaking the Policy-Facing Latent Contract
A new quantization approach targets the latent distribution that downstream action policies depend on, not just reconstruction quality.
REAL Rating Framework Aims to Standardize AI Collaboration Disclosure
A new arXiv preprint proposes a structured rating system for disclosing how much generative AI contributed to a piece of content.
Dual-Process Approach Proposed for Safer Autonomous Driving
An arXiv preprint proposes a dual-process approach to help autonomous vehicles meet safety standards, traffic rules, and social norms.
Self-Propagating Misalignment in LLM Agents: Memory Auditing Not Enough
A new arXiv paper argues that LLM agents can become misaligned without any external adversary, and that this misalignment can spread, making memory auditing or disabling insufficient defenses.
Language Model Activations Occupy Error-Correcting Basins
A new arXiv preprint proposes that language models' resilience to activation perturbations stems from privileged error-correcting basins.
Benchmarking LLM Agent Protocols: NLIP vs. A2A
A new arXiv paper benchmarks the Natural Language Interaction Protocol against A2A, measuring the cost of inter-agent hops.
SHarP prunes agent harnesses with saliency
A new arXiv paper proposes SHarP, a saliency-based method for pruning the systems that coordinate LLM calls, tool use, and task execution.
Agentic Environment Trains AI on Legal Evidence from Summary Judgment Motions
A new simulation environment uses U.S. civil litigation records to test AI agents on legal investigation tasks.
CUAWright: A minimal unified interface for computer-use agents
The paper challenges the standard design of computer-use agents, proposing a minimal unified interface that may better exploit models' coding skills.