Research Briefing

Studies when next-token prediction models trained on a distribution can yield low adversarial regret in online decision-making. Shows that with unbounded context, any distribution is exponentially close to a low-regret distribution, but bounded context introduces fundamental limitations.

Introduces daVinci-LLM, a fully open pretraining research effort combining industrial-scale compute with complete research transparency, releasing data pipelines, training logs, and model checkpoints. Aims to advance the science of pretraining which is typically obscured by commercial pressures.

Establishes information-theoretic impossibility results for safety verification of self-improving AI systems, proving that under power-law risk schedules, no classifier-based safety gate can simultaneously maintain bounded cumulative risk and unbounded utility.

Proves that reward hacking is a structural equilibrium under five minimal axioms, not a correctable bug. Shows any optimized AI agent will systematically under-invest in quality dimensions not covered by its evaluation system, regardless of alignment method.

Proves that semantic memory systems inherently face interference, forgetting, and false recall as a mathematical consequence of the geometric structure enabling semantic generalization. Derives four formal results about fundamental tradeoffs in semantically continuous kernel-threshold memories.

Proposes HISA, a hierarchical indexed sparse attention mechanism that replaces the flat O(L²) token scanning bottleneck in DeepSeek-style sparse attention with a two-stage coarse-to-fine search, significantly reducing computational overhead for long contexts.

UK AISI researchers reproduce Anthropic's Natural Emergent Misalignment (EM) from reward hacking using open-source models and non-production RL environments. Finds partial reproduction: reward hacking leads to some EM but with important differences from Anthropic's results.

Introduces FormalProofBench, a private benchmark evaluating whether AI models can produce formally verified mathematical proofs in Lean 4 at the graduate level. Best model achieves only 33.5% accuracy, with performance dropping rapidly for harder problems.

Introduces MazeBench evaluating 16 model configurations on visual maze solving, finding that models like GPT-5.4 (91%) achieve high accuracy by converting images to text grids and enumerating paths rather than genuine visual planning, consuming thousands of tokens.

Introduces SARL (Structure-Aware Reinforcement Learning), a label-free framework that improves reasoning by teaching models the structure of reasoning (topology) rather than optimizing for specific outcomes. Extends RLVR to open-ended settings without verifiable rewards.