Research Briefing

Formalizes probabilistic alignment for world models and introduces PAWBench plus PAWEval to evaluate video generators as stochastic samplers, finding that current frontier generators fail to match reference outcome distributions.

Introduces Zero-WAM, a causal video-action model conditioned on in-context human video demonstrations that enables zero-shot cross-task generalization for robotic manipulation, supported by an automatically curated dataset and a future-chunk prediction objective.

Anthropic Research blog post titled 'Automated researchers can reliably mitigate alignment failures.' Full content was not provided in the source data, but the title indicates Anthropic is reporting empirical results on automated AI safety research as a mitigation for alignment failures. Likely accompanies a longer technical paper.

Theoretical and empirical analysis showing that Evolution Strategies provide broader reasoning coverage and better Pass@K than GRPO for LLM reasoning, owing to sparse functional updates and population diversity, motivating a hybrid approach.

Proposes Test-Time Policy Optimization (TTPO), a label-free test-time training scheme that distills agreeing rollouts and penalizes disagreeing ones to improve mathematical reasoning without supervised signals.

Introduces TASTE (The AI Safety Taste Evaluation), a benchmark measuring how well AI models can judge AI safety research proposals by agreement with experienced human researchers. Two design choices (a discussion stage and filtering for high-confidence labels) produced 77% estimated human agreement across 92 pairs. Fable 5 achieved 60% agreement, underperforming humans.

Introduces UrbanGround, a realistic 3D city replica for evaluating multimodal LLM agents on sustained navigation and spatial reasoning, and reports that strong local perception does not compose into reliable long-horizon goal-directed behavior.

Presents GameWAM, a unified world-action model that jointly predicts future visuals and executable keyboard-mouse actions for native video-game control via block-causal flow matching, mode-specific action distributions, and block-cycle replanning.

Introduces and operationalizes 'monitorability disposition,' a meta-level property measuring a model's willingness to flag its own misbehavior and remain monitored. The author argues current large reasoning models are not trained to make their chain-of-thought traces monitorable, leaving a gap that output filtering cannot close. Frames the OpenAI-HuggingFace and Mythos/Fable 5 incidents as motivating examples.