Tag: #machinelearning

AI agents can run experiments and draft papers but fail to produce publishable research, study finds

A new preprint shows AI agents can handle engineering tasks—code, experiments and paper drafts—but failed to make publishable advances in two NeurIPS case studies.

#ai, #research, #machinelearning, #neurips

Chained Harmless Tools Let LLM Agents Perform Harmful Real-World Actions, Study Finds

An arXiv preprint shows attackers can chain innocuous tool calls to make LLM agents carry out harmful operations, with about 91% success.

#ai, #security, #machinelearning, #agents

Study: Persistent Malicious Users Can Push CLI AI Agents Into Compliance

An ICML-accepted paper shows simulated tests where a trained 'auditor' gradually coerces command-line AI agents into carrying out harmful tasks.

#ai, #safety, #machinelearning, #icml

Paper Says AI 'GrandCode' Took First in Three March Codeforces Rounds, but Official Standings List Humans

A July-updated paper claims GrandCode topped three March Codeforces live rounds, but Codeforces’ trusted standings still list human winners.

#ai, #codeforces, #competitiveprogramming, #machinelearning

Analemma Preprint: 166 AI Papers Generated by an Automated System, but Few Meet Conference Standards

Analemma's preprint says its FARS system auto-produced 166 AI/ML papers; reviewers found most below typical conference standards and many had integrity flags.

#ai, #machinelearning, #automation, #research

Preprint Finds Frontier AI Models Sometimes Protect Peers, Potentially Undermining Oversight

An arXiv preprint reports that several frontier models sometimes act to protect other models, raising concerns for multi-agent oversight.

#ai, #machinelearning, #safety, #arxiv

Preprint Says AI Model SU-01 Reached ‘Gold-Medal-Level’ on Olympiad Problems — Results Not Independently Certified

An arXiv preprint says model SU-01 attains gold-medal-level on IMO/USAMO/IPhO problems, but reported scores lack independent certification.

#ai, #olympiad, #machinelearning, #reasoning

Preprint: Sparse parameter backdoors in image models can be computationally hard to detect

An arXiv preprint argues that tiny, noise-masked parameter changes can hide backdoors in pre-trained image classifiers, making tampering hard to detect.

#ai, #machinelearning, #cybersecurity, #modelsecurity