August 29, 2026

Anthropic published a paper showing that an Automated Alignment Researcher (AAR) agent searches the literature, proposes methods, and trains models to close alignment gaps. In ten failure categories — including deception, sycophancy, and privacy violations — the system closed 26% to 96% of the safety gap and outperformed 28 human safety researchers on tasks like deception mitigation, with one run lifting an early Claude Opus 4.8 checkpoint to near-production alignment in 60 hours.

Label-free post-training method that exploits an asymmetric failure mode in majority-vote pseudo-labels: rollouts that disagree with the pseudo-label are usually wrong regardless of the vote. Distills agreeing rollouts via On-Policy Self-Distillation and penalizes disagreeing rollouts with Grouped RL. Without ground-truth labels, TTPO matches label-supervised OPSD on five competition benchmarks and lifts Qwen3-1.7B from 38.0% to 45.2%.

Teacher-free on-policy distillation for flow matching. At each timestep the student’s deterministic next-state prediction is branched into K stochastic SDE candidates, rolled out with an ODE sampler, and scored against a deterministic self-reference to produce normalized advantages. A velocity field is then optimized with an all-branch pull-push objective. Outperforms prior RL and OPD baselines on single and mixed-reward benchmarks without requiring task-specific teachers.

Sandbox built from Hong Kong’s 3D geospatial data that supports first-person navigation for evaluating MLLM agents. Three research questions probe scene grounding, long-distance navigation, and robustness to environmental changes. MLLM agents handle basic visual recognition well but struggle with orientation, pedestrian awareness, and sustained goal-directed behavior over extended exploration.