VideoCoCo Pushes VBench to 77.88

Featured

01 Executable Code Controls Video Time

VideoCoCo assigns temporal control to executable Blender code. A coding agent programs the scene, object motion, and progression over time. Blender then produces a deterministic spatiotemporal draft.

A generative video engine turns that draft into realistic footage. One engine controls how events unfold; the other controls how they look. Physical constraints become inspectable, editable, and repeatable artifacts instead of wishes buried in a prompt.

VideoCoCo raises OmniWeaving's PhyGenBench score from 0.475 to 0.558. On VBench-2.0, it improves the baseline from 52.18 to 77.88 and leads both benchmarks on average. The team also built VideoCoCo-3K with 3,000 draft-instruction-target video sets.

The full paper still needs to establish how well code handles complex physics and natural motion. Results may also depend heavily on Blender script quality. Video teams should test whether their target scenes can be described reliably as programs.

Key takeaways:

• Executable programs turn physical and temporal constraints into an inspectable, editable control layer.
• Splitting simulation from rendering suits products that need repeatable camera and object motion.
• Evaluate script reliability and scene programmability before adopting this approach.

Source: VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

02 Tool Use Must Earn Its Cost

Visual reasoning agents do not need more tools by default. They need to predict whether each call will produce a net gain. The framework evaluates two questions: Was the tool necessary, and did it improve the answer?

Existing models often lack this judgment. Gains on difficult tasks can disappear when unnecessary calls introduce errors on simple ones. Reinforcement learning teaches the model to call tools according to task demands while improving performance on genuinely difficult problems.

This evaluation better reflects deployment trade-offs across quality, latency, and cost. The paper's full benchmark results are still needed to quantify those gains.

Key takeaways:

• Measure the net benefit of visual tool use, not the number of calls.
• Mistaken calls on simple tasks can cancel gains on harder problems.
• Use necessity and answer improvement as joint constraints on cost and quality.

Source: Beacon: Knowing When and How to Perform Agentic Visual Reasoning

03 Semantic Scores Hide Broken Bodies

Multi-person editing exposes a gap between semantic correctness and physical plausibility. A vision-language model may approve an embrace or piggyback ride despite fused limbs, extra hands, or intersecting bodies.

MPIE-Bench contains 2,500 samples across 405 scenes and 14 interaction types. Multi-person mesh reconstruction separately checks anatomical completeness and whether contact geometry matches the instruction.

Across ten editing models, standard vision-language scores generally exceed 0.95. Yet the best anatomy and interaction scores reach only 0.65 and 0.72. No model performs well on both. Five-person reviews also find the new metrics closer to human judgment than zero-shot vision-language evaluation.

Key takeaways:

• Treat anatomical integrity and contact geometry as separate acceptance criteria.
• Vision-language scores above 0.95 can still hide obvious structural failures.
• A high aggregate score does not establish production readiness.

Source: MPIE-Bench: Benchmarking Anatomically Plausible Multi-Person Interaction Editing