A paper-dominated day on the research feeds. The clearest throughline is verifiability: OpenComputer and EnvFactory build executable, checkable software worlds for computer-use agents, while a cluster of RLVR papers attacks the crude single-reward signal with learned-reliability process rewards, policy-aware rubrics, and contrastive evidence optimization. Video generation kept pushing past plausibility toward verification, with work on reasoning under verifiable rewards and benchmarks for detecting AI-video artifacts. Evaluation itself is being reframed as a design science rather than a leaderboard. The press cycle was quiet; the signal today is in the methods.
Two frameworks build executable, checkable environments so agent training and evaluation can be grounded in real outcomes rather than synthetic stubs. Paired with new GUI and long-context agent work, the day's agent research is converging on verifiability and reusable skills.
A verifier-grounded framework for constructing checkable software worlds where computer-use agents can be trained and evaluated against real outcomes.
Synthesizes executable environments at scale to unblock the two main bottlenecks in agentic RL: robust execution and reward signal.
Derives reusable skill programs from past experience to tackle long-horizon tasks more reliably than re-planning each time.
GUI-agent benchmark using live omni-modal smartphone interaction rather than static screenshots.
Maintains a reusable orientation cache across invocations so long-context agents don't re-explore the same corpora or repos each time.
Several papers attack the same weakness: RLVR's single binary reward gives every token the same credit. Learned-reliability process rewards, policy-aware rubric rewards, and contrastive evidence optimization all try to make the signal reflect which steps actually mattered.
Process reward models that emit a reliability estimate alongside each step score, so downstream methods can weight feedback by confidence.
Weights rubric-based rewards by what the current policy actually needs to learn, extending RLVR beyond automatically-checkable correctness.
Differentiates token credit within a correct RLVR rollout via contrastive evidence, instead of rewarding every token equally.
Fully open-source RLVR post-training recipe aimed specifically at long-context capability.
Shows that inserting self-generated data before and during RL improves outcomes by shaping data diversity.
Video models are being pulled from 'looks real' toward 'is correct': reasoning under verifiable rewards, controllable generation from sparse intent, and benchmarks that score artifacts and multi-shot coherence rather than single-clip realism.
Applies verifiable-reward training to video diffusion so models optimize for verifiable correctness, not just plausible frames.
Benchmark for whether multimodal models can detect temporal and structural artifacts in AI-generated video — a prerequisite for trustworthy video QA.
Wraps unified diffusion video editing in a tool-using agent loop for more controllable, multi-step edits.
As models act over time through tools and users, static benchmarks lose meaning. A position paper argues interactive evaluation needs design-science rigor, while ThoughtTrace captures what users actually think (not just say) and Editor's Choice tests abstract editing intent.
Argues that evaluating agents acting over time through tools and users requires design-science methods, not static leaderboards.
First large-scale dataset capturing what users think (not only what they type) during real LLM interactions.