AI Agent Collaboration Could Trigger Population Takeoff Threshold
A new arXiv paper warns that collaborative AI agents with real-world attack capabilities may reach a population threshold that enables rapid takeoff.
Research Digest
Preprints and papers, rewritten in plain English.
Organizational Behavior Lens Proposed for LLM Agent Safety
A new arXiv paper argues that LLM-based agents should be studied through organizational behavior, focusing on how instructions, peer messages, and memory shape their actions.
Who Verifies the Verifier? Co-Evolving Graders for Self-Improving Agents
A new paper argues that the rubrics used to evaluate self-improving agents must themselves evolve alongside the agents.
Loop-Native Residuals Prevent Accuracy Loss in Looped Transformers
A new arXiv preprint argues that looped Transformers need dedicated residual connections to maintain reasoning accuracy across tens of thousands of test-time iterations.
Macro-step world model speeds up atomistic simulations of materials evolution
A new arXiv preprint proposes a world-model approach that skips microscopic steps to simulate long-term materials dynamics more efficiently.
New Harness Keeps LLM Agents Policy-Compliant in Complex Worlds
A proposed agent harness targets enterprise needs by enforcing policy compliance and surfacing hidden side effects in tool feedback.
SpikeSSL: A Universal Framework for Spike Inference from Calcium Imaging
A new framework aims to infer neural spikes accurately across diverse calcium indicators using dynamics-informed state-space layers.
Modeling Error Propagation to Attribute Failures in Multi-Agent LLM Systems
A new arXiv paper proposes a framework for tracing failures in LLM-based multi-agent systems back to their originating components.
Evolutionary Competition Could Improve Autonomous Driving Training
A new arXiv paper suggests using evolutionary competition to discover effective training scenario distributions for autonomous driving, rather than relying on surrogate criteria.
Layer-Selective Fine-Tuning Aims to Preserve LLM General Abilities
New research suggests that choosing which layers to fine-tune can help specialized adaptation avoid degrading pretrained capabilities.
Bandit Strategy Selection Improves RL Robustness in RTS Games
A new approach separates strategic command selection from unit control to make reinforcement learning agents less brittle against unfamiliar opponents.
AgentEvolver: A System for Turning Task Experience into Reusable Capability
The paper argues that agents need a system-level loop to turn task execution into lasting improvement.
3D Framework Unifies AI Evaluation on Puzzles
A new arXiv paper proposes a 3D characterization framework to unify how AI systems from different paradigms are compared on puzzle-based sequential decision making.
Orthogonal coordinates don't guarantee intervention meaning in vascular models
A new arXiv preprint argues that predictive coordinates need explicit intervention anchors to support scientific verification.
LLM Agents Need Memory-Type-Aware Strategies, arXiv Paper Argues
A new preprint argues that retrieval-based memory systems for LLM agents overlook meaningful differences between memory types, proposing a more diverse strategy.
Coding Agent on ARC-AGI-3: No State, But Learns Sequentially
A study traces how a coding agent using a frozen foundation model and external scripts handles a sequence of abstract reasoning tasks, offering lessons for continual learning.
DeltaReplay: Reusing Mobile GUI Trajectories Without Exact Matches
A new method lets mobile GUI agents reuse past execution trajectories even when the new task differs in parameters or shares only part of the workflow.
MultiWorldBench tests whether multiplayer AI models share one world
A new Minecraft benchmark probes whether independently controlled views in multiplayer world models stay consistent with a single shared reality.