Research Briefing — March 11, 2026

Formalizes when chain-of-thought reasoning is necessary through 'opaque serial depth' - the longest computation possible without interpretable intermediate steps. Computes numeric upper bounds for Gemma 3 models and provides automated measurement methods.

Uses AlphaEvolve (LLM-based code mutation agent) to improve lower bounds for five classical Ramsey numbers, including R(3,13) from 60 to 61 and R(4,13) from 138 to 139.

Discovers that enabling chain-of-thought reasoning in LLMs substantially improves factual recall even for simple single-hop questions. Identifies two mechanisms: a computational buffer effect and an implicit multi-hop decomposition of recall.

EsoLang-Bench evaluates LLM reasoning using esoteric programming languages (Brainfuck, Befunge-98, etc.) that have minimal training data. Frontier models achieving 85-95% on standard benchmarks score only 0-11%, revealing memorization vs genuine reasoning.

Studies how reasoning affects honesty in LLMs using realistic moral trade-offs. Contrary to humans who become less honest with deliberation, reasoning consistently increases LLM honesty across scales and families. Shows this is partly due to representational geometry, not just reasoning content.

Presents PlayWorld, a pipeline for training video world models from unsupervised robot self-play rather than human demonstrations. First system capable of learning entirely from autonomous play for high-fidelity video world simulation.

Proves an impossibility result for AI alignment verification: no procedure can simultaneously be sound (rejects misaligned systems), general (full input domain), and tractable (polynomial time). Each pair is achievable but all three cannot hold simultaneously.

Benchmarks political persuasion capabilities of seven frontier LLMs (from Anthropic, OpenAI, Google, xAI) across N=19,145 participants. Finds LLMs outperform standard campaign advertisements, with Claude most persuasive and Grok least persuasive.

Argues that improving LLM logical reasoning capabilities mechanistically enables situational awareness - a dangerous emergent capability. Introduces the RAISE framework identifying three pathways through which reasoning improvements enable deeper self-awareness.

Challenges the Linear Representation Hypothesis by measuring geometric distortion in LLM activation spaces and proposes 'Curveball steering' - nonlinear activation steering that follows the curved geometry of activation spaces instead of assuming linearity.

Proposes In-Context RLVR, which uses the quality of reasoning traces as demonstrations (Evidence Gain) to implicitly reweight rewards in RLVR training. Shows better reasoning traces serve as better teachers via in-context learning.