MaD-RSI: Model-as-Data for Recursive Self-Improvement of Agent Harnesses
A new arXiv paper describes a method that treats the model itself as data to recursively improve the harnesses controlling LLM agents.
Research Digest
Preprints and papers, rewritten in plain English.
Seller Agents Distort Tool Markets Through Selection Gaming
A new arXiv paper shows that seller agents optimising only for selection can distort tool markets, a phenomenon the authors call 'toolflation'.
Community-consistent synthetic mobility via evolving user-location graphs
A new mobility generation framework aims to preserve community structure in synthetic trajectories, not just statistical patterns.
FALCA: Hierarchical credit assignment for long-horizon deep-search agents
A new paper proposes falsification-aware reinforcement learning with hierarchical credit assignment to help LLM agents handle long-horizon deep search tasks.
New Benchmark Diagnoses Decision Models in LLM Systems
DecisionPivot evaluates whether LLM decision models make correct choices under controlled variations and assign reliable confidence, going beyond item accuracy.
Grounding AI emergence claims in information decomposition
A new arXiv paper argues that emergent capabilities in AI systems should be analysed using formal information decomposition rather than loose intuition.
Agentic ensemble improves visual anomaly detection
A new method combines multiple pretrained detectors in an agentic ensemble to address the scarcity of anomaly samples in visual inspection.
Proof-Carrying Contracts Let AI Agents Make Unbreakable Commitments
A new arXiv preprint proposes a mechanism for AI agents to make verifiable commitments that go beyond what plain text instructions can enforce.
Debate Helps Human Judges, Not Weaker AI Models, Study Finds
New research suggests that using debate to oversee advanced AI may work for human judges but fails when weaker AI models act as judge.
Pluralistic AI Alignment: Learning from Diverse Crowd Preferences
A new arXiv paper outlines the challenge of training AI systems to accommodate heterogeneous and conflicting user preferences.
Agentic AI for 6G: Closing the Gap Between Algorithm Discovery and Silicon
A new arXiv paper highlights the growing mismatch between AI-speed algorithm discovery for 6G and the slow, specialist work of implementing those algorithms on baseband silicon.
OptScientist Aims to Automate Discovery of Optimization Laws
A new arXiv paper argues that automated scientific discovery should infer the optimization problems behind phenomena, not just fit equations to data.
AutoRedClaw Automates Red-Teaming for Persistent AI Agents
A new arXiv paper proposes automated red-teaming to probe the security and safety of persistent, personalized 'claw-style' agents.
Semantic Observability Aims to Tame Microservice Failure Logs
A new arXiv preprint proposes using semantic observability to help diagnose root causes in complex microservice deployments.
Evolving Semantic Indexes for Generative Retrieval
A new arXiv paper explores evolving semantic indexes to improve generative retrieval by tackling the challenge of designing discrete semantic identifiers.
Can AI models covertly sabotage recursive self-improvement?
A new arXiv paper argues that recursive self-improvement is becoming real and that its safety implications, including covert sabotage by misaligned models, remain understudied.
Aptamer sandbox aims to close the loop in AI molecular discovery
A new framework proposes using aptamers to test lab-in-the-loop AI discovery, where wet-lab results guide iterative molecular generation.
Benchmark Evaluates LLMs for VASP INCAR Configuration
A new benchmark called INCARBench aims to assess how well large language models can configure VASP INCAR files for first-principles calculations.