acceptodds
Withdrawn from ICLR 2027

VideoRubric: Rubric Granularity Matters in Reinforcement Learning for Video Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reinforcement learning with verifiable rewards is promising for improving video reasoning in multimodal large language models, yet answer-level rewards provide limited guidance on the reasoning process. We introduce VideoRubric, a rubric-guided reinforcement learning framework that studies process supervision for VideoQA at two granularities: specific rubrics describe sample-conditioned evidence and reasoning requirements, while abstract rubrics capture reusable patterns such as temporal consistency, causal modeling, and evidence grounding. VideoRubric constructs specific rubrics with a multimodal teacher, abstracts them into a shared rubric set, and incorporates rubric judgments into GRPO through correctness gating. Experiments with Qwen2.5-VL-7B-Instruct on 1K rubric-augmented samples show that the correctness-gated specific-rubric supervision achieves the strongest average performance across MMVU, TempCompass, VSI-Bench, and VideoMME. Further analyses show that rubric granularity shapes learned reasoning differently: specific rubrics yield clearer separation between correct and incorrect responses, whereas abstract rubrics produce more structured and preferred reasoning outputs.

Primary area:
reinforcement learning
Keywords:
video reasoning, multimodal large language models, rubric-guided reinforcement learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jwR8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies rubric-guided reinforcement learning in video reasoning. To address the issue of answer-level rewards providing limited guidance on the reasoning process, this paper proposes a rubric-guided reinforcement learning framework named VideoRubric that studies process supervision for VideoQA. Specifically, VideoRubric leverages specific rubrics and abstract rubrics to learn sample-conditioned evidence and reusable patterns, respectively. The experimental results on Qwen2.5-VL-7B-Instruct with 1K rubric-augmented samples validated that the proposed VideoRubric achieves competitive results on multiple benchmarks.

Critical strengths
  1. The exploration of rubric granularity in reinforcement learning for video reasoning is technically sound and interesting.

  2. The experiment showing that Qwen2.5-VL-7B-Instruct optimized on 1K rubric-augmented samples achieves strong performance across multiple benchmarks is interesting.

  3. This paper is well-organized and easy to read. The authors provided extensive experimental details in the appendix.

Critical weaknesses
  1. From the results in Table 1, we can see that 1K RL data has achieved strong results. Have the authors tried experiments with more RL data?

  2. The training costs, such as GPU hours, are not provided in this paper. Could the authors provide the results?

  3. In this paper, the teacher model and abstraction model play a key role in RL data generation. Did the authors investigate the effect of different models on this process?

Questions for authors

After reviewing this paper, I have two critical questions that require a response from the authors.

  1. Have the authors tried experiments with more RL data? Would more rubric-guided RL data produce better performance?

  2. Did the authors explore different settings for the teacher model and abstraction model? Would a more capable model produce high-quality rubric RL data and lead to further performance gains?

Reviewer_SKdZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces VideoRubric, a rubric-guided reinforcement-learning framework for video reasoning. It studies two levels of process-supervision granularity: specific rubrics, which describe sample-conditioned visual evidence and reasoning requirements, and abstract rubrics, which describe reusable reasoning properties across samples. A multimodal teacher generates the specific rubrics, an abstraction model consolidates them into a shared rubric bank, and a judge model evaluates generated chain-of-thought reasoning against the active rubric set. The rubric reward is combined with format and answer rewards using correctness gating, so rubric supervision is applied only when the final answer is correct. The authors train Qwen2.5-VL-7B-Instruct on 1K rubric-augmented samples and evaluate it on MMVU, TempCompass, VSI-Bench, and VideoMME. The specific-rubric model with gated rewards obtains an average score of 56.36, compared with 55.59 for outcome-only GRPO and 55.68 for gated abstract-rubric supervision. Additional analyses suggest that specific rubrics better separate correct from incorrect responses, while abstract rubrics lead to more structured and preferred reasoning when both models answer correctly.

Critical strengths
  1. The research question is clear and relevant. Treating rubric granularity as a distinct design dimension for process supervision is a useful and potentially important perspective for video-reasoning RL.

  2. Correctness gating is well motivated. It prevents abstract process criteria from rewarding plausible-looking reasoning when the final answer is wrong, and it aligns process supervision with the primary task objective.

Critical weaknesses
  1. The main performance evidence is statistically weak. The experiments use one base model, one training configuration. The average improvement over outcome-only GRPO is only about 0.77 points, yet no confidence intervals, significance tests, or multi-seed results are reported. The training set is also heavily concentrated in LLaVA-Video and NeXT-QA, which limits the strength of the generalization claim.

  2. The rubric judge does not directly inspect the video. It evaluates the response against rubric text, so the claimed visual grounding depends on the upstream teacher's rubric generation rather than an independent video-grounded verification step. This may reward matching or paraphrasing the rubric, including unsupported reasoning. The reasoning-quality analysis is also restricted to both-correct examples, uses a relatively small human subset, does not report annotator agreement, and relies on LLM-generated taxonomies and labels.

Questions for authors
  1. Can the authors control rubric count, length, and relevance across the two settings—for example, by using question-conditioned abstract-rubric selection—and add a video-grounded judge or human-verified rubric ablation to isolate the effect of granularity?

  2. Can the authors provide rubric-quality and answer-leakage analyses, agreement statistics between human and automatic evaluators, and reasoning-quality results on incorrect answer and answer disagreement cases?

Reviewer_3DpnOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

VideoRubric studies process supervision for VideoQA reinforcement learning by varying the granularity of rubric-based rewards. The authors distinguish specific rubrics — sample-conditioned criteria naming the concrete evidence and reasoning steps needed to justify an individual answer — from abstract rubrics — sample-agnostic criteria such as temporal consistency, causal modeling, and evidence grounding, consolidated iteratively from the specific ones. Both rubric sets are produced offline by a teacher (Qwen3-VL-Plus) and an abstraction model (Qwen3-Max).

Critical strengths

The granularity contrast is a clean experimental design. Prior rubric work studies how rubrics are generated, evolved, aggregated, or scheduled; this paper holds the RL algorithm, training data, and judge fixed and varies only the abstraction level of the rubric text, isolating a variable the literature has left implicit.

Critical weaknesses

The empirical margin is small and reported without variance or a serious baseline set. Averaged over four benchmarks, specific+gated beats outcome-only GRPO by 0.77 points and abstract+gated by 0.68. No seeds, no confidence intervals, one model (Qwen2.5-VL-7B), one data scale (1K). Per-benchmark, the outcome baseline is actually best on TempCompass, and the +2.24 MMvU gap does most of the work. The closest prior rubric methods — RLCER, AutoRubric-RIV, RuCL, Rubric Anchors — are cited but never run, so "best among compared methods" compares against only GRPO and an in-house ablation. At this effect size the headline claim may not survive seed noise.

The paper's own analysis undercuts its conclusion. On the shared both-correct set (Table 3), humans prefer the abstract-rubric model 60.7% holistically and 63.0% on completeness, and GPT-4o agrees at 63.0%. The paper reframes this as "benchmark accuracy and perceived reasoning quality capture different aspects," but that is a concession, not a defense: the variant humans and an LLM judge actually prefer is not the one the abstract touts as "the strongest." An ablation dressed as a framework, evaluated on a single backbone and a single dataset scale, does not clear the bar for acceptance.

Questions for authors

see the above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.