acceptodds
Withdrawn from ICLR 2027

VeriRL: Grounding Video Reasoning for Depression Assessment with Verifiable Facial-Dynamics Rewards

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reinforcement learning with verifiable rewards (RLVR) has shown strong potential for improving reasoning in multimodal large language models (MLLMs), yet extending it to medical video reasoning is challenging: task labels verify final predictions but not whether the behavioral evidence reported alongside them agrees with observable video evidence. We propose VeriRL, a post-training framework that grounds video reasoning through verifiable facial-dynamics rewards. Selected temporal facial Action Unit (AU) dynamics serve as measurable, clinically motivated behavioral references for auxiliary verification. To support progressive training, we construct two facial-evidence-aligned Chain-of-Thought corpora, GroundFace-CoT ( 4K images) and GroundVideo-CoT (1,130 video clips). During RL, the dynamics reward measures agreement between structured model reports and precomputed OpenFace-derived references. We further introduce Dynamics-Grounded Sign-GRPO (DGS-GRPO), which uses a fixed-sign classification advantage to preserve correctness feedback under homogeneous rollouts, while a bounded dynamics term adjusts its magnitude without reversing the correctness-determined sign. Experiments on AVEC 2014 and our curated Vlog dataset show gains over the supervised baseline in binary recognition, improving Accuracy/Macro-F1 from 0.61/0.61 to 0.63/0.63 on Vlog and from 0.57/0.51 to 0.60/0.55 on AVEC. On held-out Vlog videos, dynamics agreement further increases from 0.64 to 0.72, with exact channel agreement improving from 0.50 to 0.60. These results support measurable behavioral signals as auxiliary verification sources for evidence-grounded medical video reasoning.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Multimodal Large Language Models, Reinforcement Learning with Verifiable Rewards, Medical Video Reasoning, Evidence Grounding, Facial Dynamics, Depression Assessment

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_puTKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VeriRL, a reinforcement learning post-training framework for video-based depression assessment, aiming to improve the consistency between generated reasoning and observable visual evidence. The authors use OpenFace to extract facial Action Unit (AU) dynamics and construct four discrete behavioral indicators as auxiliary verifiable rewards. They also introduce GroundFace-CoT and GroundVideo-CoT datasets for progressive image-level and video-level supervised fine-tuning. During reinforcement learning, the proposed Dynamics-Grounded Sign-GRPO (DGS-GRPO) replaces group-relative classification advantages with fixed-sign advantages based on prediction correctness, while using AU dynamics agreement to adjust their magnitudes. Experiments on AVEC 2014 and a curated Vlog dataset demonstrate modest improvements in binary classification and structured facial-dynamics agreement over the supervised baseline.

Critical strengths

The paper studies the difference between getting the final answer correct and using the correct visual evidence in medical video reasoning. Using facial changes as extra signals to check the reasoning is a reasonable idea.

The proposed method combines image-level facial understanding, video-level supervised fine-tuning, and reinforcement learning in a clear training pipeline.

The paper evaluates both classification accuracy and how well the model results match facial changes extracted by an independent method. This provides more evidence than task accuracy alone.

Critical weaknesses
  1. Limited novelty of DGS-GRPO. The zero-advantage problem in homogeneous rollout groups has already been studied. For example, the ICLR 2026 paper No Prompt Left Behind (https://arxiv.org/abs/2509.21880) addresses this problem through advantage shaping. The main difference of DGS-GRPO is using AU dynamics rewards to adjust the advantage values. Besides, the paper does not compare with these existing methods.

  2. The AU-based reward is not well justified. The reward uses only four facial AUs and their changes over time, but these AUs are not clearly validated as indicators of depression. Matching the facial dynamics extracted by OpenFace does not necessarily mean better depression detection. The paper also provides limited analysis of why these four AUs are selected. Since I am not an expert in depression detection, I would leave this concern to reviewers with more domain knowledge.

  3. Better dynamics agreement does not mean better CoT reasoning. The RL reward only checks the four categories in the <dynamics> block, but does not directly evaluate whether the reasoning in <think> is based on the correct video evidence. Therefore, better dynamics agreement cannot fully support the claim that the CoT reasoning becomes more faithful to the video.

  4. Limited improvements and missing ablation studies. Compared with SFT, VeriRL only improves Accuracy and Macro-F1 from 0.61 to 0.63 on Vlog Full, while the four-class AVEC Macro-F1 remains at 0.28. Each test set also contains only 100 videos. In addition, the paper does not include an ablation that removes the dynamics reward while keeping DGS-GRPO. Therefore, it is unclear how much each component contributes to the final performance.

  5. The experiments mainly use Qwen2.5-VL-7B, which is relatively outdated compared with newer models like Qwen3-VL. Although the paper includes newer models as baselines, it does not test the proposed training method on these stronger models.

Questions for authors

see weakness

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

This paper introduces a method called VeriRL for depression assessment in videos. It follows the SFT+RL paradigm. For SFT, it curates two datasets GroundFace-CoT and GroundVideo-CoT, while for RL, it further introduces a dynamics-grounded Sign-GRPO, which combines a fixed-sign task classification advantage with a bounded dynamics agreement term. The method is valudated on AVEC 2014 and Vlog, and results are promising.

Critical strengths
  • The paper studies an important problem in medical video reasoning to encourage plausible reasoning rationals rather than only the final answer.
  • The CoT datasets could be valuable resources to relevant research.
Critical weaknesses
  • While the SFT model achieves notable improvements, the full DGS-Depression-7B yields only marginal gains over the SFT counterpart. This suggests that the RL components provides limited additional benefits. Particularly, the model shows no meaningful improvement on four-class AVEC severity prediction, and exposes a significant limitation in scaling the method beyond binary tasks to multi-class clinical gradations.
  • The verification reward relies on OpenFace's ability to reliably detect and track faces and extract AUs. Consequently, the pipeline may be vulnerable to pose variations, poor lighting, or occlusions. This limits the model's robustness in real-world settings.
Questions for authors
  • The proposed method is evaluated exclusively using Qwen2.5. It is essential to study more latest models such as Qwen3 to check whether its benefits generalize across base models.
  • Does it matter to separate local facial SFT and video SFT into two distinct stages? Would joint training on both types of data be feasible, and how would its performance compare with the proposed two-stage training strategy?
Reviewer_dwzuOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VeriRL for video-based depression assessment and structured facial-behavior reporting. It constructs synthetic image- and video-level chain-of-thought (CoT) data, progressively fine-tunes the Qwen2.5-VL-7B vision-language model (VLM), and applies reinforcement learning using classification correctness and agreement with facial Action Unit (AU) dynamics extracted by OpenFace. The proposed DGS-GRPO uses fixed-reference classification advantages with a bounded dynamics adjustment. Experiments on AVEC 2014 and a curated Vlog dataset show modest improvements in binary classification over supervised fine-tuning (SFT), together with improved agreement between structured behavioral reports and AU references, while four-class depression-severity prediction remains largely unchanged.

Critical strengths

1. Using measurable behavioral attributes to supervise report accuracy is a reasonable idea with some empirical support.

Distinguishing final-answer correctness from the accuracy of accompanying behavioral descriptions addresses a relevant problem. AU measurements provide computable auxiliary targets for checking specific behavioral claims. Table 3 reports held-out improvements in dynamics agreement from 0.64 to 0.72 and exact channel agreement from 0.50 to 0.60. These results show that the model can learn to predict the specified AU references more accurately and provide initial support for this auxiliary-supervision approach.

Critical weaknesses

1. The motivation for the VLM formulation is insufficiently established, and a direct AU-detection-plus-VLM/LLM baseline is missing. A natural alternative is to extract AU dynamics with OpenFace and provide these measurements to a large language model (LLM), or provide both the video and the measurements to a VLM, for classification and report generation. This approach would also use observable behavioral evidence and produce natural-language explanations, but the paper does not evaluate it. This omission is particularly consequential because OpenFace already supplies the reference targets, and the evaluation preprocessing itself uses OpenFace (Appendix C.5). The additional value of training a VLM to predict these attributes and then aligning its outputs with the detector-derived references is therefore unclear. Comparisons with general-purpose multimodal models and the authors’ own SFT initialization demonstrate the benefit of domain adaptation, but do not establish an advantage over directly using detected AU information. The paper also does not demonstrate a specific benefit of its generative reasoning formulation in classification accuracy, report quality, or computational cost.

2. The post-training method primarily modifies rewards and advantages for this particular task, and its methodological novelty and independent contribution are limited. The method uses classification correctness to determine the sign of the advantage and a bounded AU-consistency term to adjust its magnitude. For binary classification with a valid dynamics report, Eqs. (3), (5), and (6) reduce to

Ai=1.8ci+0.2Rdyn,i−1,A_i = 1.8c_i + 0.2R_{\mathrm{dyn},i} - 1, where cic_i indicates classification correctness and Rdyn,iR_{\mathrm{dyn},i} is the dynamics-agreement reward. The outer clipping does not change this expression under the stated coefficients. The central modification is thus a fixed-reference combination of task and auxiliary rewards. This may be a reasonable adaptation, but the paper does not establish a substantial algorithmic contribution beyond this task-specific design. Preserving correctness-dependent advantage signs is a direct consequence of the construction; the corresponding optimization benefit requires independent evidence. Table 2 does not clearly specify the advantage used by “GRPO w/o Dynamics,” and the reported comparisons do not separately establish the effects of the advantage modification and the dynamics reward. In particular, standard GRPO with the same combined reward is missing.

3. The paper provides limited research insight beyond improved agreement with the specified auxiliary targets. The clearest finding is that adding AU-consistency supervision improves prediction of AU references on held-out videos. Although useful, this finding offers limited additional understanding of how AU information should be incorporated, when reinforcement learning is preferable to direct supervision, or under what conditions fixed-reference advantages are beneficial. The experiments do not distinguish improved video reasoning from distillation of OpenFace attributes or gains specific to the chosen reward weights and training configuration. Consequently, readers cannot readily derive reusable design principles from the results. Even within video-based depression assessment, the paper does not sufficiently establish the conditions under which this pipeline should be preferred or explain the source of its advantage.

4. The classification gains are small, and the experimental evidence is insufficient to establish their robustness or the broader value of the method. The Vlog test set contains only 100 videos, with accuracy increasing from 0.61 to 0.63. Figure 3 shows nine corrected predictions and seven newly incorrect predictions, yielding only two additional correct predictions overall. AVEC binary accuracy increases from 0.57 to 0.60, while four-class Macro-F1 remains at 0.28 (Table 1). Results across multiple training seeds and uncertainty estimates are not reported, leaving the reliability of these differences unclear. The larger improvements concern agreement with OpenFace-derived references. Table 3 validly measures generalization to the specified auxiliary targets, but training and evaluation use the same reference-construction procedure. These results do not independently establish more accurate descriptions of the original videos or more reliable use of behavioral evidence in the final predictions. Table 4 also lacks an SFT comparison, so it cannot establish how much rationale quality improves because of reinforcement learning. The practical benefit of the additional training procedure therefore remains insufficiently supported.

5. The validity of the CoT annotations and the independent value of the reasoning text are not established. GroundFace-CoT and GroundVideo-CoT are central components of the framework, but the paper does not adequately validate the quality or independent contribution of these synthetic rationales. In GroundVideo-CoT, the teacher receives both the AU reference and the task label and is instructed to generate an explanation consistent with them (Appendix C.8). Checks on formatting, label agreement, and descriptive direction constrain the outputs, but do not independently validate the visual descriptions or the inferences connecting observed behavior to the task conclusion. This construction allows the possibility of explanations organized around a known answer, which the current evaluation does not rule out. The paper does not report an independent human audit of the training CoTs against the original images or videos. Manual inspection during Vlog screening does not substitute for such an audit. The human evaluation in Table 4 concerns rationales generated by the trained model, and the raters receive AU references without viewing the original videos, so it does not validate the training annotations. Moreover, removing GroundFace-CoT in Table 2 removes the entire image-supervision stage, preventing attribution of the gain specifically to natural-language reasoning text. A comparison that retains the same training examples, AU supervision, and task labels while removing only the reasoning text is absent.

Questions for authors

1. Motivation and direct alternatives. On the same data splits, how does directly extracting AU dynamics with OpenFace and supplying them to an LLM, or supplying both the video and AU measurements to a VLM, perform in classification and report quality? What is the main motivation for training the VLM to predict these attributes through the proposed pipeline, and which results demonstrate an additional capability or practical benefit over these alternatives?

2. Methodological contribution and attribution. Does “GRPO w/o Dynamics” in Table 2 use the group-relative advantage in Eq. (4), or the fixed-reference construction in Eq. (5) with the dynamics term removed? Can you provide controlled comparisons with standard GRPO using the same combined reward and with fixed-reference advantages without the dynamics reward? How do these comparisons establish a contribution beyond task-specific reward adjustment?

3. Research insight. Beyond improved agreement with AU references after adding AU supervision, what is the main reusable insight established by this work? Which controlled experiment supports a concrete conclusion about how AU information should be incorporated, the need for reinforcement learning, or the conditions favoring fixed-reference advantages, while distinguishing improved video reasoning from distillation of automatically detected attributes?

4. Robustness and behavioral-report evaluation. Are the binary-classification gains consistent across training seeds? Can you provide paired comparisons and uncertainty estimates appropriate for the test-set size? Is there an evaluation comparing SFT and VeriRL on the same videos in which blinded human raters inspect the original videos, to establish improvements in behavioral-report quality and their relationship to reinforcement learning?

5. CoT quality and independent value. Have the teacher-generated training CoTs undergone an independent human audit against the original images or videos? If so, please report the audit size, evaluation criteria, and findings concerning both behavioral-description accuracy and the validity of inferences connecting behavior to task conclusions. Is there also a controlled comparison that retains the same training examples, AU supervision, and task labels while removing only the natural-language reasoning text? Such evidence is needed to distinguish the contribution of CoT text from that of additional supervision.

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

Related papers

Discussion (0)

Sign in to comment.