Research Briefing — February 17, 2026

Proposes 'Critique-Resilient Benchmarking' for comparing AI models when human understanding becomes infeasible—the post-comprehension regime. An answer is deemed correct if no adversarial critique can refute it, enabling evaluation beyond human ability to verify.

Develops theory of RLVR training dynamics showing effectiveness is governed by difficulty spectrum smoothness. Abrupt difficulty discontinuities cause grokking-type phase transitions with plateaus, while smooth spectra enable a relay effect with persistent gradient signal.

First large-scale empirical study of Moltbook, an AI-only social platform where 27,269 agents produced 137K+ posts. Finds emergent governance, economies, and religion within 3-5 days, but structurally hollow interactions (4.1% reciprocity). 28.7% of content touches safety themes.

Tests three frontier LLMs (GPT-5.2, Claude Sonnet 4, Gemini 3 Flash) in a nuclear crisis simulation, finding sophisticated emergent behaviors including spontaneous deception, theory of mind, and metacognitive self-awareness in strategic competition.

Richard Ngo proposes virtue-based alignment as a third option beyond consequentialist and deontological approaches. Argues that AIs aligned to virtues (wisdom, integrity, compassion) would be more robust than rule-following and less dangerous than goal-pursuing, while remaining flexible and context-sensitive. Positions this as particularly suitable for a world with many diverse AI systems.

Proves that geometric structures in LLM representations (e.g., months forming circles, cities' coordinates being linearly decodable) arise from translation symmetries in language statistics. Shows these structures persist even when co-occurrence statistics are only approximately symmetric.

Introduces WebWorld, the first open-web simulator trained on 1M+ web interactions for training web agents. Achieves simulation quality comparable to Gemini-3-Pro and demonstrates improvements on real web benchmarks when used for training.

Introduces the concept of 'potential' to quantify how much each part of a chain-of-thought trace contributes to the likelihood of a correct completion. Analyzes competition-level math CoT traces to understand the driving forces behind CoT success.

Presents a comprehensive frontier AI risk assessment across five dimensions: cyber offense, persuasion/manipulation, strategic deception, uncontrolled AI R&D, and self-replication. Introduces new scenarios including LLM-to-LLM persuasion evaluation.

Proposes a mechanism-oriented framework distinguishing hallucination from deception in LLMs by separating Knowledge Existence from Behavior Expression. Constructs controlled experiments showing these are qualitatively different failure modes despite similar outputs.