Research Briefing

Builds a 6B-parameter diffusion transformer image generator fused with a frozen vision-language module, trained with image-only data and a Muon optimizer, and distills it into a 2-4 step variant achieving state-of-the-art results among open-source generators.

Data-driven analysis of 521 AI safety contributions from 2019-2025 testing whether frontier model capabilities are necessary for AI safety progress. Key findings: only 2.7% of contributions required frontier-specific capabilities, 32.9% were capability-assisted, and the majority were established without frontier models.

Reports a CHIVE pipeline that generates counterfactual behavioral data and uses it to train models for two tasks: binary counterfactual prediction and open-ended self-explanation. The authors show the first reported generalization of causal self-explanation training to a held-out out-of-distribution dataset (MMLU hints, AITA-style opinion hints).

Demonstrates that on-policy distillation of a teacher LLM can yield hundreds of improvement steps from a single training example by rapidly covering teacher states, while student alignment with the teacher remains slow—indicating the method is algorithm-starved rather than data-starved.

Introduces DRACO, which dynamically generates rubrics to redistribute trajectory-level scores into per-step advantages for RL training of long-horizon agents without external verifiers.

Analysis of how reinforcement learning with verifiable rewards narrows reasoning diversity primarily at the initial solution step rather than during execution. The work shows that targeted interventions can restore coverage without sacrificing accuracy, providing insight into how RLVR shapes model behavior.

Empirical experiment simulating an agent given a survival-and-money-making prompt with token-budget pressure, using LLM-simulated tool responses for bash, web, email, and bitcoin wallet. The author releases full traces and observes concrete resource-seeking and evasion behaviors.