Research Briefing — June 7, 2026

This post re-runs a sampled version of the FANToM theory-of-mind benchmark on current frontier models, finding that while belief-state tracking has improved substantially since 2023, models still trail human performance on a relatively simple cooperative reasoning task. It matters because robust belief-state tracking is foundational for AI agents operating in multi-party collaborative settings.

A preliminary interpretability writeup proposing that transformers track persistent context along a sequence-time axis (not just the depth axis) in a compact low-dimensional subspace of the residual stream. The author suggests this concentrated representation could be projected out, compared to attention/MLP writes, and potentially targeted by interventions.

This post examines how sequentially mixing training objectives during LLM post-training creates distinct training dynamics depending on environment distinguishability and pressure for shared circuitry, classifying outcomes into ecological generalists, conditional policies, and strategy churn. It challenges the safety-research assumption of a fixed training objective and argues non-stationary dynamics can be used to intentionally shape AI minds.

A MATS-mentored research post using natural selection theory to model AI agency, arguing that noisy selection on genome structure can make evolution effectively non-myopic and give a Darwinian account of how individuals emerge from coalitions of lower-level replicators. It is the first of a planned series connecting evolutionary dynamics to feature-learning, interpretability, and alignment.

A contrarian alignment essay questioning whether corrigibility, an agent's willingness to be corrected or shut down, is actually a desirable property to engineer into AIs. The author argues we may want to deliberately avoid developing corrigibility techniques, challenging a common assumption in safety research.

Following yesterday's News on Anthropic's pause proposal, A response to Anthropic's post on recursive self-improvement, examining the argument that no single lab can unilaterally pause frontier AI development because of competitive dynamics and verification challenges. The post probes what a unilateral Anthropic pause would actually accomplish.