CAS Reframe tops Mix-RL and MOPD on 14 visual benchmarks

Researchers at the Institute of Automation, Chinese Academy of Sciences report that their new Reframe post-training framework leads all comparisons across 14 image and video benchmarks, raising an observation-dependence score on the V* benchmark from a field-best 41.93 to 75.53. The paper, posted to Hugging Face, targets what the authors call observation neglect, the pattern where a visual agent invokes a tool and then disregards what it returns.

Reframe pairs Ground-Drop-GRPO, which randomly removes the original visual input and rewards evidence-aware answers, with Observation-Gated-MOPD, which "replaces deficient tool actions with canonical ones mid-rollout and resumes distillation from the repaired observation," per the abstract. Neither stage requires a task-specific SFT cold start.

Across seven image benchmarks, enabling tools lifts Reframe's accuracy by 5.23 points, "versus at most 1.47 for the other models," the paper states, and relative to a tool-free Qwen3.5-9B base at 63.46, "87% of the overall gain stems from tool use." On images, Reframe beats vanilla MOPD by 1.29 points; on videos by 0.78, with the largest single-task jump of 4.86 points on VisualProbe.

The study restricts its toolset to two primitives, CropImage and SelectFrames, and the comparison table covers four open agents (Pixel-Reasoner, Thyme, DeepEyes, DeepEyesV2) rather than closed frontier systems. Code and the Reframe-100K corpus, 63K image QA pairs plus 37K video QA pairs, are released on GitHub. Agentic tooling has been a steady beat on our tracker, with 557 stories in the agents column over the past 90 days.