One-Line Summary

A new 2M-pair video editing corpus and two agent-learning methods show AI moving from simple filters to precise control and self-improving behaviors.

Research Papers

Goku releases 2M-pair video editing dataset and benchmark

Goku is a large training-and-evaluation set that teaches AI editors to follow plain-language instructions for videos, not just change colors but also control motion and structure. It targets multi-task and structural manipulations such as precisely guiding a subject’s movement, addressing gaps in prior datasets focused mostly on appearance tweaks. ¹

The dataset contains 2 million high-quality, instruction-aligned video editing pairs built with a synthesis pipeline that breaks complex edits into controllable sub-problems. A progressive filtering system is applied throughout to keep the data reliable as complexity increases. ¹

On top of the data, the authors propose Goku-Edit, which uses a multimodal large language model (MLLM) as the text encoder and a decoupled dual-branch design. A dedicated mask branch handles structural control, freeing the main branch to focus on appearance rendering. ¹

For evaluation, Goku-Bench provides 1,000 human-verified test cases and seven editing-specific metrics. On this benchmark, Goku-Edit delivers up to a +8% gain in instruction following compared with other open-source models, grounding claimed improvements in measurable outcomes. ¹

Hierarchical Experimentalist Agents let models learn by experimenting

Hierarchical Experimentalist Agents (HExA) turns large language models (LLMs) into in-context experimentalists: they design small, targeted experiments, run them in a simulator, learn reusable skills, and integrate the evidence to answer queries or act. The framework is training-free, works with black-box models, and needs no external supervision, oracles, or offline data. ²

To evaluate active experimentation, the authors introduce Interphyre, a tool-calling benchmark built on the PHYRE 2D physics environment with simulation application programming interfaces (APIs). Baseline agents struggle on the hardest levels—Claude Sonnet 4.6 scores 2%—but the same model under HExA reaches up to 77%; using only skills transferred from easier levels (no fresh experiments), it attains 44%. HExA also improves open-weight models and outperforms agentic baselines like ReAct and Reflexion. ²

TRIAGE: role-typed credit assignment boosts agent RL

TRIAGE is a credit-assignment method for agentic reinforcement learning (RL) that labels each segment of an agent’s trajectory by semantic role instead of giving uniform credit based only on the final verifier outcome. Standard Group Relative Policy Optimization (GRPO) can punish useful exploration when a rollout fails and reinforce regressive steps when it succeeds; TRIAGE adds a structured judge to mark segments as decisive progress, useful exploration, no-progress infrastructure, or regression, then applies fixed role-conditioned rewards while keeping verifier outcomes as the optimization signal. ³

Across ALFWorld, Search-QA, and WebShop, TRIAGE improves success rates over GRPO for two policy models and reduces environment-facing turns on completed rollouts by 10.4% (ALFWorld) and 14.8% (WebShop). Ablations indicate the gains come from role typing—especially catching regressions—rather than simply adding denser rewards, and the authors connect the approach to lower-variance policy gradients. ³

Why It Matters

Richer, instruction-aligned data and evaluation like Goku’s 2M-pair corpus and 1,000-case benchmark give teams a realistic way to train and measure complex video edits, moving beyond style filters toward precise structural control. ¹

For agents that act, HExA shows that letting models run targeted experiments can turn a 2% success rate into up to 77% in a physics simulator, while TRIAGE suggests role-aware, process-level signals can cut detours and improve reliability—together pointing to agents that learn faster and waste fewer steps. ²