#1 ·
Turns Matter: Advancing Streaming Video RL via Turn-Level Credit
Shuai Dong, Kanghui Tian, Linghan Chen, Wenqi Shao
Reinforcement LearningVideo Understanding
Why it matters
Streaming video models must decide, turn by turn, whether to answer or keep watching. Applying GRPO with a single trajectory-level reward means every turn inherits the same credit — the model gets reinforced for redundant answers after it already succeeded. Turn-level credit assignment is the fix, and the same bug exists in any multi-turn agent (tool calls, chat, browsing) where the trajectory reward hides which step actually worked.
The paper shows that sharing one advantage across all turns in a streaming video RL run reinforces redundant replies and discourages continued observation after an early success. By assigning credit at the turn level instead, the model learns a cleaner 'respond vs. keep watching' policy. The result is more accurate answers and fewer wasted responses in long-horizon video streams.
arXiv | HuggingFace
#2 ·
From Pixels to Actions: Towards Native Vision-Language-Action Policies
Weichen Fan, Haiwen Diao, Penghao Wu, Siyuan Hu
RoboticsVision-Language Models
Why it matters
Almost every manipulation model runs pixels through a pretrained visual encoder first, then reasons about actions on the compressed features. But the signals manipulation depends on — local geometry, boundaries, depth, subtle state changes — are exactly what that abstraction destroys. A policy that ingests raw pixels and outputs actions end-to-end removes the single biggest information bottleneck in robot learning pipelines.
The authors challenge the standard VLA recipe — encode, then act — and argue the encoder is where manipulation-relevant detail dies. Their native VLA policy takes raw visual input directly and learns to produce actions without an intermediate perceptual summary. On manipulation tasks that hinge on fine visual detail, the end-to-end policy outperforms encoder-based baselines, suggesting the encoder bottleneck is real and costly.
arXiv | HuggingFace
#3 ·
FrontierPhysics: Evaluating Agents for End-to-End Physics Research
Bingran You, Wenjun Ke, Ria Rosenauer, Bo Liu
AI AgentsScience Benchmarks
Why it matters
Every physics benchmark so far asks fixed questions, tests isolated sub-tasks, or replays a paper whose solution is given. None measure whether an agent can actually do research. With 56 tasks authored by 42 working physicists from their own research across nine domains — and scoring research quality, not just correctness — this is the first benchmark that treats the agent as a researcher rather than a quiz taker.
FrontierPhysics packages real research problems — derived from the authors' own work — into an end-to-end benchmark: the agent gets a research question and must plan, compute, and produce an answer that physicists grade on quality, not just numeric correctness. It covers nine physics domains and exposes where current agents fall short when the task was never designed for them.
arXiv | HuggingFace
Outlook
Yesterday's theme is less raw capability and more precise credit: where exactly did the model earn its reward (which turn of a video conversation), where exactly does the pipeline lose information (the encoder bottleneck in VLA models), and whether agents can survive a task that was not designed for them (real physics research). Expect follow-ups on turn-level RL to spread beyond video to tool-calling agents.