Also Worth Noting - 2026-07-17

Also Worth Noting

02 [Training] SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning Sparse trajectory-level rewards leave a supervision gap that outcome-based RL alone cannot close for long-horizon agents. SEED fills it by converting completed on-policy rollouts into hindsight skills, injecting dense token-level guidance mid-training rather than waiting for episode outcomes. The result treats distillation and RL as complementary signals rather than competing post-training recipes. Teams training multi-turn agents on tool-use tasks should consider this pairing before assuming RL alone is sufficient. link

03 [Theory] Demystifying On-Policy Distillation: Roles, Pathologies, and Regulations On-policy distillation does not raise a student model's capability ceiling. It steers the student toward correct reasoning paths it could already reach, acting as an exploration catalyst rather than a capability expander. Prompt diversity matters more than per-problem sampling volume, and the guiding signal quality determines everything. Teams using OPD to unlock genuinely new skills are misreading what the method does. link

04 [RAG] GRASP: GRanularity-Aware Search Policy for Agentic RAG Most agentic RAG pipelines hard-code retrieval granularity and treat lexical versus semantic mode as a fixed configuration. GRASP trains agents via RL to select chunk size and retrieval type adaptively per query step, cutting irrelevant context that otherwise interferes with multi-step reasoning. Letting the model control granularity rather than the pipeline engineer is the mechanism that moves the needle. If your RAG system fixes these parameters at build time, that decision is worth revisiting. link

05 [Agent] RoboTTT: Context Scaling for Robot Policies Robot policies have operated on single-step or short-history context; RoboTTT scales visuomotor context to 8K timesteps, three orders of magnitude beyond current state-of-the-art policies, without growing inference latency. Test-time training on that extended context unlocks one-shot in-context imitation from human video demonstrations and on-the-fly policy improvement at inference time. This is the robot equivalent of what long-context LLMs do for text, and the scale at which it becomes practically useful is now established. link