Research Briefing — March 7, 2026

As first reported in Research yesterday, Documents 'self-attribution bias' in LLMs: when a model evaluates the safety of its own prior outputs (in-context), it systematically assigns lower risk scores than when evaluating the same outputs in a fresh context. This has direct implications for AI safety pipelines that use the same model as both actor and monitor.

Applies logit lens and tuned lens techniques to probe the latent reasoning steps of CODI (a latent reasoning model) on arithmetic problems. Finds that odd latent steps appear to perform active computation (higher entropy) while even steps serve as storage, and that translators trained on text tokens outperform those trained directly on latent hidden states.

Building on yesterday's Reddit buzz, Zvi Mowshowitz analyzes the US Department of Defense's designation of Anthropic as a supply chain risk, arguing it is retaliatory for Anthropic refusing to give the military unrestricted access to Claude. Describes the legal and political dynamics as an escalation in government-AI company relations.

Proposes that shaping RL exploration of the 'motivation-space' (why a model does something and how it perceives itself) is an understudied and promising lever for AI safety. Distinguishes between action exploration and motivation exploration during RL training, arguing the latter is underdetermined by reward signals and thus shapeable.

Argues that causal variables in mechanistic interpretability are irreducibly subjective — the choice of what variables to study is a pre-formal step that causal inference formalism cannot validate. This means every choice of causal variables induces a different hypothesis space, and reproducibility in mech interp requires reproducing the labeling/variable-selection process, not just code.