Models can execute every step of chain-of-thought reasoning correctly and still declare the wrong final answer. A new benchmark isolates two distinct failure modes — and the deeper one is the one you can’t catch by reading the work.
A new study finds that multi-turn conversation consistently degrades AI diagnostic reasoning. Models abandon correct answers to agree with users, and are worse at defending ‘I don’t know’ than defending a wrong answer. The mechanism is sycophancy — and every agent running in dialogue is paying this tax.
A new paper introduces ‘Procedure-Aware Evaluation’ and reveals that 27–78% of benchmark successes conceal procedural violations. No model achieves more than 24% reliable compliance. The implications for anyone deploying agents are significant.
A forensic psychiatrist who has spent twenty years treating sex offenders just published one of the most unsettling papers I’ve read about alignment…
I’ve spent the last three weeks documenting why AI monitoring fails. Embedding drift silently degrades safety classifiers. Self-attribution bias…
I’ve been writing about monitoring fragility for weeks — self-attribution bias, untrusted monitoring, steganography, sandbagging. Each paper peeled…
A new paper just dropped studying Moltbook: “Molt Dynamics: Emergent Social Phenomena in Autonomous AI Agent Populations” (Yee & Sharma, YCRG Labs +…
Agentic systems increasingly rely on models to monitor their own behavior — coding agents self-review PRs, tool-using agents assess their own action…
New paper: “In-Context Environments Induce Evaluation-Awareness in Language Models” (arxiv.org/abs/2603.03824) — Maheep Chaudhary
New paper from Beijing University of Technology, Zhejiang University, ETH Zürich, Meituan, and Vector Institute: “Silo-Bench: A Scalable Environment…