Video reasoners learn to audit their own visual reliance: VIOS lifts three backbones by 3.60 to 4.63 points across 12 benchmarks
Synopsis
In video reasoning the correct answer often hinges on a few visually decisive positions, yet post-training methods such as GRPO assign one coarse score to the whole reasoning trajectory and need multiple rollouts or external judges. VIOS instead has the model audit its own visual reliance: it compares next-token distributions with and without the video, takes the per-token reliance gap between the student and an answer-privileged teacher, and concentrates supervision where the student under-uses visual evidence. Across 12 video-understanding benchmarks, Qwen2.5-VL-7B-SFT rises from 46.35 to 49.95 on average, InternVL2.5-4B and InternVL2.5-8B gain 4.63 and 3.93 points, and training needs a single on-policy rollout with no external reward or judge.
Interpretation
Video-reasoning post-training gains a per-token supervision signal weighted by visual reliance, without sampling several trajectories per question or calling a larger judge model. Post-training had mainly followed GRPO and its video variants, sampling eight or more rollouts per prompt, rewarding those above the group mean with one scalar, and spreading that reward over the entire reasoning chain. On 12 video-understanding benchmarks, Qwen2.5-VL-7B-SFT trained with VIOS rises from 46.35 to 49.95 on average, ranking first on MVBench, Video-MME, MINERVA, MMVU, VideoMathQA, TempCompass, LongVideoBench and VSI-Bench, and second on TVBench, MLVU and SciVideoBench.
The core is to let the model measure how much each token depends on the video, then use the student-teacher reliance gap to decide where supervision applies. The teacher is an exponential moving average of the student and additionally sees the ground-truth answer, so it offers a higher supervision ceiling than a frozen teacher. Training covers three backbones, Qwen2.5-VL-7B-SFT, InternVL2.5-8B and InternVL2.5-4B, on the Video-R1-260k corpus, with each model trained for 100 VIOS steps on 8 NVIDIA H200 GPUs.
The reliance-gap gate and the visual contrastive term each add gains that compound, placing supervision on the positions where vision actually matters. Distillation alone gives an average of 48.21, adding the gate gives 48.69, and adding the contrastive term reaches 49.95. Ablations use the main Qwen2.5-VL-SFT configuration: the gate beats ungated distillation on 9 of 12 benchmarks and adds 0.6 points on average, and the contrastive term adds a further 1.3 points.
Perspective
The demonstrated scope is three backbones, Qwen2.5-VL-7B-SFT, InternVL2.5-4B and InternVL2.5-8B, trained on the Video-R1-260k corpus and evaluated on 12 video-understanding benchmarks plus four static image benchmarks; training caps video at 16 frames and 128 vision tokens per frame, doubled at inference to 32 frames and 256 tokens per frame. The method requires a base model that can already produce a coherent chain of thought, which is why Qwen2.5-VL is initialized from an SFT checkpoint.
Computing the reliance gap requires evaluating the student and teacher both with and without the video at every position, roughly doubling per-step forward cost relative to standard distillation, an overhead offset by eliminating multi-rollout reinforcement learning. The method assumes a ground-truth final answer for each training example, restricting it to VQA-style data with answers. The contrastive term promotes visual utilization but does not enforce grounding to specific frames or regions, so hallucinations consistent with the video's general content can still occur; correctness enters only implicitly through the answer-privileged teacher, so the model inherits whatever rationalization biases the teacher exhibits at that training step.