acceptodds
Under review as a conference paper at ICLR 2027

Beyond Binary Preferences: Graded Preference Optimization for Limb-Motion Captioning

Abstract

Vision-Language Models (VLMs) can generate rich video captions, yet often misidentify which person performs an action or which limb is involved, particularly across camera cuts. Improving these details requires evaluation and training that distinguish missing information from incorrect assertions. We introduce FlexBench, a benchmark spanning 3,105 shots and 18,161 evaluation queries, with human-verified identities and systematic per-person coverage of fine-grained limb actions and states. Its reference-derived checklists support automated assessment of complete captions against person- and shot-specific annotations. Our Graded Physical Alignment score (GPA) awards credit for correct content and deducts points for incorrect or fabricated actions, explicitly accounting for errors in the aggregate score. Building on this rubric, we propose Graded Margin Direct Preference Optimization (GM-DPO), which assigns larger preference margins and greater training weight to more severe action errors. Across three VLM backbones, GM-DPO achieves the highest substantive-action and GPA scores among the evaluated preference objectives, improving GPA over DPO by 2.02–3.40 points. On Qwen3-8B, it reduces the weighted hallucination rate by 21.3% relative to DPO. These gains accompany sustained long-form output, improved shot structure, and competitive performance on three additional multimodal benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Captioning, Fine-Grained Limb Motion, Preference Optimization, Hallucination Mitigation, Multi-Shot Video Benchmarking

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_kCuWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies fine-grained limb-motion fidelity in dense, multi-shot video captioning, emphasizing the distinction between omitted details and incorrect assertions. It introduces FlexBench, comprising 660 video sequences, 3,105 shots, and 18,161 evaluation queries, with human-verified person identities and shot-specific annotations of actions, limb states, and visibility. Its Graded Physical Alignment (GPA) metric combines positive credit with explicit penalties for incorrect motion claims. The proposed Graded Margin DPO (GM-DPO) uses synthetic caption perturbations to assign error-severity-dependent preference margins and loss weights. Experiments compare four preference-learning baselines across three VLM backbones and three training seeds. GM-DPO improves GPA over DPO by 2.02–3.40 points and reduces weighted hallucination rate by 21.3% on Qwen3-8B, while retaining long-form captions and competitive performance on three additional multimodal benchmarks.

Critical strengths
  1. The benchmark addresses an important and insufficiently measured failure mode of video captioning. FlexBench evaluates freely generated captions at the person, limb, and shot levels, including stationary and hidden limbs rather than only salient actions. Distinguishing omissions from false assertions is particularly consequential when captions serve as downstream training supervision. The multi-round human verification and reported evaluator–human agreement provide substantial support for the benchmark’s usefulness. I regard this targeted evaluation resource as the paper’s strongest contribution.

  2. The empirical improvements are consistent across architectures and training seeds. Under a shared preference dataset and auxiliary NLL objective, GM-DPO achieves the highest substantive-action and GPA scores among the evaluated preference methods on all three backbones. The positive paired GPA gains across seeds make the results more convincing than a single-model demonstration. Component ablations also support combining margins and weights, while caption-length statistics indicate that the gains are not simply obtained by producing shorter, less informative outputs. This establishes practical value for the proposed training recipe.

Critical weaknesses
  1. The experiments do not isolate the benefit of severity conditioning from stronger generic preference optimization. The weight-only and margin-only ablations establish that these components help, but both remain severity-dependent. The paper lacks controls using a constant margin and constant weight matched to their average magnitudes, shuffled severity assignments, or comparable tuning of DPO’s optimization strength.
Questions for authors

Have you compared GM-DPO against controls using (a) constant margins and weights matched to the averages in GM-DPO, and (b) shuffled severity labels? How does the evaluator handle an unsupported action claim that does not correspond to any reference-derived checklist item?

4 comments

Reviewer comments (verbatim)

"Have you compared GM-DPO against controls using (a) constant margins and weights matched to the averages in GM-DPO, and (b) shuffled severity labels? How does the evaluator handle an unsupported action claim that does not correspond to any reference-derived checklist item?"

Response

Thank you for the helpful suggestions. We have added both controls and evaluations on four external benchmarks, and clarified the scope of our evaluation. All changes in the revised manuscript are highlighted in blue.

1) Severity-conditioning controls

Thank you for suggesting these controls. We ran both with the same seeds and evaluation protocol as in the paper:

  • constant_mean: fix severity to its training-set mean (1.10029924), giving a constant margin of 0.22006 and weight of 1.11003.
  • shuffled: shuffle severity labels across pairs while preserving their distribution and margin–weight relationship. The preferred and rejected captions remain unchanged.

Both use α=0.1\alpha=0.1, γ=0.2\gamma=0.2, β=0.1\beta=0.1, and 1.0LNLL1.0\mathcal{L}_{\mathrm{NLL}}, with the paper's multi-judge evaluation.

MethodGPA ↑GPA_clip ↑Sub. ↑Non-sub. ↑TSA_w ↑WHR ↓
DPO19.6628.867.1638.4019.7818.80%
constant_mean20.4629.259.1337.4521.1718.46%
shuffled20.1230.649.1836.5420.8421.82%
GM-DPO23.0630.529.3043.6922.3414.80%

As the table shows, both controls improve GPA over DPO, so part of the gain can come from changing the margin and weight. GM-DPO nevertheless improves GPA over these controls by 2.60 and 2.94 points, while reducing WHR by 3.66 and 7.02 percentage points. Shuffling gives slightly higher GPA_clip (30.64 vs. 30.52), but also higher WHR (21.82% vs. 14.80%). This supports using each pair's actual error severity, beyond changing the overall optimization strength.

We have added these results and discussion to Sec. 5.5 (lines 474–477) and Appendix C.3 (lines 1004–1018; Table 9).

2) Unsupported actions and hallucination coverage

Thank you for raising this point. We should have explained the evaluator's inputs more clearly. Each question is scored using the full structured limb reference. The original Sec. 3.2 (lines 159–160) describes this setup, and the submitted FlexBench/data/FlexBench_en.jsonl and its Chinese counterpart already contain gt_extracted_actions_by_shot and questions. The annotations list each person's actions, laterality, order, and limb states by shot. For each QiQ_i, the evaluator receives the full reference, generated caption, and question list.

This allows questions about inactive or invisible limbs to catch invented actions as well. For example, if the reference states that both hands are invisible, a claim about table-tapping receives −1 under the non-substantive/no-hand rubric. We have clarified this in Sec. 3.2 (lines 180–183) and added a visual example in Appendix A.3, Fig. 8 (lines 731–745; caption, lines 803–808).

We agree, however, that an extra claim unrelated to all checklist items may still escape a penalty. Our detailed annotations reduce this risk. FlexBench focuses on limb-motion accuracy within dense captions; it does not evaluate all scene, camera, or audio descriptions. This limitation was already acknowledged in our original Conclusion. The revised Sec. 6 (lines 524–539) retains this scope and outlines future evaluation, with further discussion in Appendix D.

Reviewer comment (verbatim)

"How does the evaluator handle an unsupported action claim that does not correspond to any reference-derived checklist item?"

Response (continued)

To check whether improving limb-motion accuracy affects broader video understanding, we evaluated all five Qwen3-8B policies on four complementary benchmarks. These more directly probe the factual and temporal understanding needed for accurate captions than our original Video-MME, POPE, and Video-ChatGPT evaluations.

  • VideoHallucer [30] tests intrinsic and extrinsic hallucinations across five categories. Both answers in each yes/no pair must be correct; we report the macro-average across categories.
  • VidHalluc [14] tests action (ACH: binary pairs and MCQ) and temporal sequence (TSH) hallucinations, with 13,426 responses per model.
  • TempCompass [18] tests action, direction, speed, attribute change, and event order through MCQ and caption generation (CG). CG provides candidate information; Qwen3.8-Max answers a question from the generated caption to check its temporal content.
  • VinoGround [34] tests temporal matching on 500 video–caption groups. Text (video-to-caption) and Video (caption-to-video) require both choices in each direction to be correct; Group requires all four. We use local inference and the official scorer.

All scores are percentages. Results marked with † are cited from the original benchmark papers for comparison; the ACH binary references use Binary QA Pair. Bold marks the best score among our five policies.

The table shows VideoHallucer paired accuracy (VH), VidHalluc action-hallucination binary pairs (ACH pair) and temporal sequence hallucinations (TSH), TempCompass caption generation (TC-CG), and VinoGround group accuracy (VG-Group). Table 4 reports all nine metrics.

ModelVHACH pairTSHTC-CGVG-Group
DPO58.5678.6990.3366.6217.00
cDPO58.5878.3690.6766.5217.80
IPO58.4678.2792.0066.0217.20
SimPO58.6978.1191.3365.8217.20
GM-DPO58.9879.2092.8067.2318.60
GPT-4o †53.3066.7982.00—24.60
GPT-4o (CoT) †————35.00
Gemini-1.5-Pro †37.8059.1083.83—10.20
Qwen2-VL-72B †————17.40
Qwen2-VL-7B †————15.20
LLaVA-OneVision-Qwen2-7B †————14.60
PLLaVA-7B †38.10————
LLaVA-NeXT-Video-DPO-7B †32.00————
VideoLLaMA2-7B †—29.0926.17——
VideoChat2-7B †7.80——38.50—
Qwen-VL-Chat-7B †———37.30—
Video-LLaVA-7B †17.809.8727.1734.806.60

In the full results (Table 4), GM-DPO leads on eight of the nine metrics. Over DPO, it gains 0.51 and 2.47 percentage points on VidHalluc ACH binary pairs and TSH, 0.61 points on TempCompass caption generation, and 1.60 points on VinoGround Group. These modest improvements support our aim of improving limb-motion factuality while retaining broader video understanding, including temporal information conveyed in generated captions.

We have replaced the original evaluations in Sec. 5.6 (lines 479–492; Table 4) with these results and discussion, and moved the Video-MME, POPE, and Video-ChatGPT results to Appendix C.4.

References

[30] Wang et al. VideoHallucer: Evaluating Intrinsic and Extrinsic Hallucinations in Large Video-Language Models. 2024.

[14] Li, Im, and Fazli. VidHalluc: Evaluating Temporal Hallucinations in Multimodal Large Language Models for Video Understanding. CVPR 2025.

[18] Liu et al. TempCompass: Do Video LLMs Really Understand Videos?. Findings of ACL 2024.

[34] Zhang, Cai, and Lee. Vinoground: Scrutinizing LMMs over Dense Temporal Reasoning with Short Videos. 2024.

Dear Reviewer kCuW,

Thank you again for your careful review and thoughtful comments. I very much appreciate the time and effort you have devoted to reviewing this paper.

The authors have now replied to your initial comments. Could you please take a look at their responses and continue the discussion with them when you have a chance?

Thank you so much for your continued help and engagement!

Best wishes,
Your AC

Dear Reviewer kCuW,

Thank you once again for the time and effort you dedicated to reviewing our paper. Your insightful comments—especially regarding isolating the effect of severity conditioning—have helped us further clarify the contribution of GM-DPO and significantly improve our work.

We fully understand that you have a busy schedule, so we just wanted to gently follow up and see if our responses, the newly added control experiments (constant margins/weights and shuffled severity labels), and the results across four additional benchmarks have adequately addressed your concerns. If there are any remaining questions, or if we have misunderstood any of your points, please do let us know—we are more than happy to provide further clarification.

If you feel that your concerns have been satisfactorily resolved, we would be sincerely grateful if you would kindly consider raising your rating.

Best regards,

The Authors

Reviewer_n1UAOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • This paper claims that while VLMs can produce rich video captions they often misidentify fine-grained details such as which person performs an action or which limb is involved, particularly across camera cuts.
  • First, this paper introduces FlexBench (Fine-grained Limb-motion EXamination Benchmark), a benchmark for evaluating fine-grained limb actions and states in multi-shot video understanding tasks, which includes 3,105 shots and 18,161 evaluation queries.
  • Second, this work introduces Graded Margin DPO to strongly align the model preference against such action errors in multi-shot video understanding tasks.
Critical strengths
  • this paper addresses the problem of fine-grained action understanding in vieo llms -- which is a relevant area of broader interest, and current video llms are known to fail in such setups.
  • their benchmark covers diverse domains (various indoor and outdoor activities), contains diversity in terms of video duration, number of shots, etc.
  • their proposed loss formulation is well motivated for the problem.
  • the writing is fairly clear and easy to follow.
  • this paper evaluates 3 popular models and performs necessary ablations.
Critical weaknesses
  • GPA full form seems confusing, is it Graded Physical Alignment score (l.018) or graded penalty-aware metric (l.078).

  • instead of generating monolithic text for GT, does it make sense to actually structure the GT. for example, we can construct the GT as {"object1": "little girl", "action": "lifting", "object2": "a book"}. I think preparing the GT in benchmark in such structured manner makes the evaluation/verification more objective while using an automated judge and can reduce bias/inconsistency arise from AI-based evaluation. to clarify, the model you're evaluating can still produce open text. Can you try that out?

  • i am trying to get a clear distinction wrt KPM-Bench -- is the main difference b/w KPM-Bench and yours is that KPM-bench evaluates limb-motion fidelity in single shot setups and yours is in multi-shot setting?

  • the explanation of Graded Evaluation is not clear. some of the weights introduced in the metric seem random and without any scientific justification. such as 0.6 and 0.4 in eqn 1 of S_GPA formulation; assigning different weights (why did you choose 1 and 2 and not something else) to WHR formulation; similarly, what is the basis of weights 0.2, 0.3, 0.5 in TSA_w metric.

  • the GMDPO loss at its core is a modification of DPO by adding a margin, isn't ODPO loss (see Direct Preference Optimization with an Offset) doing the same thing? if not could you please clarify the distinction in terms of loss formulation? i think it's worth discussing their paper in the your work.

  • this paper lacks evaluation coverage. it would be interesting to see whether teaching the model to ground better on fine-grained action alignment reduce hallucination on video centric benchmarks such as videohallucer, vidhalluc, eventhallucion; and improve in finegrained temporal tasks like tempcompass, tvbench, vinoground, vcgbench, temporalbench. I would recommend authors to eval on these benchmarks to ground the effectiveness of GMDPO loss. POPE is an object hallucintion benchmark which is not at all suitable here. also the performance of your method on videoMME is weak.

Questions for authors

i summarized them under weakness, please see above.

7 comments

Thank you for the careful review and concrete suggestions. They have helped us clarify the benchmark and strengthen the experiments. Manuscript revisions are marked in blue.

1) GPA terminology:

GPA full form seems confusing, is it Graded Physical Alignment score (l.018) or graded penalty-aware metric (l.078)."

Response: Thank you for catching this ambiguity. GPA means Graded Physical Alignment score; “graded, penalty-aware metric” at original line 078 describes its properties. We apologize for the confusing wording and have made the full name explicit (revised lines 78–79).

2) Structured ground truth and evaluation:

instead of generating monolithic text for GT, does it make sense to actually structure the GT. for example, we can construct the GT as {"object1": "little girl", "action": "lifting", "object2": "a book"}. I think preparing the GT in benchmark in such structured manner makes the evaluation/verification more objective while using an automated judge and can reduce bias/inconsistency arise from AI-based evaluation. to clarify, the model you're evaluating can still produce open text. Can you try that out?

Response: Thank you for this thoughtful suggestion. Making evaluation explicit and consistent was also our intention. Our submitted evaluator already combines structured GT with questions. The original Sec. 3.2 (lines 159–160) states that captions are judged against both the checklist and GT reference. The submitted FlexBench/data/FlexBench_en.jsonl includes questions and gt_extracted_actions_by_shot, which organizes each person's actions and limb states by shot.

In our early experiments, GT-only evaluation was less reliable for complex descriptions with interleaved actions from multiple people. We therefore derived questions from the same GT, with substantive-action, non-substantive-state, and action-sequence categories, making individual checks and model weaknesses easier to identify. The judge uses both the questions and the full GT; the model still generates free-form captions.

We apologize that this was not clear enough. We have clarified Sec. 3.2 (lines 180–183) and Appendix A.3 (lines 731–745), and added Figure 8 to illustrate the annotation structure and evaluation flow. We hope this can help us present FlexBench more clearly.

3) Distinction from KPM-Bench:

"i am trying to get a clear distinction wrt KPM-Bench...?"

Response: Thank you for raising this. Our focus is limb-motion factuality within comprehensive video captions, beyond multi-shot identity matching and multi-person interactions.

Unfortunately, we have not found publicly downloadable KPM model weights or the KPM-Bench dataset, so a direct performance comparison is not yet possible. The distinction below follows their paper [17].

Coverage. For each identifiable person and shot, FlexBench checks substantive actions, accompanying movements, stationary states, visibility, and action order. A resting or hidden hand therefore remains relevant to evaluation. KPM's MoPE extracts dynamic actions and excludes static verbs (Sec. 4.2 of [17]).

Captioning objective. KPM describes its targets as “dense motion-centric captions,” organized through kinematic decomposition and PaMoR. We retain a natural narrative of scene, atmosphere, cinematography, dialogue, and interactions, and improve the limb facts within it. GM-DPO learns from complete caption pairs that differ in targeted limb errors while preserving surrounding content (Sec. 4.2).

We have add Figure 10 in Appendix A.5 to compare examples from FAVOR-Bench, ActivityNet-Captions used by VidChain, VideoComp, KPM-Bench, and FlexBench. An excerpt from FlexBench GT-calibrated English reference caption reads (Fig.10 (e)):

The video contains two shots from a domestic comedy, set in a warmly lit living room with... A man ... Meanwhile, he adjusts his coat with both hands ... A hard cut and a slight camera pan bring the woman into an over-the-shoulder medium close-up, ... The same man's jacketed arms remain... Looking back toward the man at the left...

The scene follows camera, invitation, food offering, and conversation; limb facts must stay attached to the correct person as the view changes. In Table 2, GM-DPO retains caption length and shot structure close to full SFT on Qwen3-8B (1,610 vs. 1,634 words; 89.8% vs. 89.7% SMR). The four benchmarks in our response to Question 6 further assess retained video understanding.

We clarified the distinction in Sec. 2.1 (lines 92–96), the dense-caption setting in Sec. 3 (lines 138–140), and GM-DPO's aim of preserving natural descriptions while improving limb accuracy in Sec. 4 (lines 214–215).

[17] Lin et al. KPM-Bench: A Kinematic Parsing Motion Benchmark for Fine-grained Motion-centric Video Understanding. 2026.

4) Metric weights

the explanation of Graded Evaluation is not clear. some of the weights introduced in the metric seem random and without any scientific justification. such as 0.6 and 0.4 in eqn 1 of S_GPA formulation; assigning different weights (why did you choose 1 and 2 and not something else) to WHR formulation; similarly, what is the basis of weights 0.2, 0.3, 0.5 in TSA_w metric.

Response: Thank you for asking us to explain these choices. The weights reflect the relative importance of the evaluated content:

  • GPA, 0.6/0.4: substantive actions define the main event and receive greater weight, while accompanying movements and stationary limb states still contribute.
  • WHR, 1:2: these follow the penalty magnitudes of grades −0.5 and −1, expressed in units of 0.5.
  • TSA, 0.2/0.3/0.5: longer chains require more actions to be correctly covered together, so full-chain coverage receives the greatest weight.

With equal weights (1/2 for each action category; 1/3 for each sequence component), the results are:

BackboneEqual-weight GPA: DPO → GM-DPOEqual-weight TSA: DPO → GM-DPO
Qwen2.5-7B14.07 → 16.0516.49 → 18.03
Qwen3-8B22.78 → 26.5019.38 → 22.00
LLaVA-2-8B25.70 → 28.2822.34 → 26.13

These values reaggregate the component means in Table 2, rounded to two decimals. GM-DPO remains ahead of DPO on both metrics across all three backbones. Each component improves over DPO, so any nonnegative normalized weights over these components preserve this pairwise advantage.

We have added the rationale to Sec. 3.3 (lines 201, 204–205, and 208).

5) Relation to ODPO and severity-conditioning controls

the GMDPO loss at its core is a modification of DPO by adding a margin, isn't ODPO loss (see Direct Preference Optimization with an Offset) doing the same thing? if not could you please clarify the distinction in terms of loss formulation? i think it's worth discussing their paper in the your work.

Response: Thank you for pointing this. You are right that a preference-dependent margin is already present in that work, and we should have made the connection explicit. Let z=hθ(x,yw)−hθ(x,yl)z=h_\theta(x,y_w)-h_\theta(x,y_l) denote the DPO logit. The two preference terms can be written as

ℓODPO=−log⁡σ(z−δ),ℓGM=−(1+αΔS)log⁡σ(z−γΔS).\ell_{\mathrm{ODPO}}=-\log\sigma(z-\delta),\qquad \ell_{\mathrm{GM}}=-(1+\alpha\Delta S)\log\sigma(z-\gamma\Delta S).

When α=0\alpha=0 and δ=γΔS\delta=\gamma\Delta S, our preference term has the same mathematical form as ODPO. In the component ablation (Table 3), the full GM-DPO objective consistently outperforms the margin-only variant using this severity-derived offset. GPA increases from 13.20 to 14.06 on Qwen2.5-7B, 20.07 to 23.06 on Qwen3-8B, and 24.55 to 25.40 on LLaVA-2-8B; GPA_clip and TSA_w also improve on all three backbones. These results support using local limb-error severity to control both the margin and the weight when learning from complete captions. We have cited ODPO [1] and clarified this relationship in Sec. 5.5 (lines 468–472).

To examine whether assigning the right severity to each pair matters, we added two controls using the same preference pairs, training seeds, auxiliary NLL term, and evaluation protocol as in the paper:

  • constant_mean: replace every severity with its training-set mean, keeping the margin and weight fixed at GM-DPO's averages (0.22006 and 1.11003).
  • shuffled: permute severity labels across pairs, preserving their distribution and the margin–weight relationship while breaking the link to the actual error.
MethodGPA ↑GPA_clip ↑TSA_w ↑WHR ↓
DPO19.6628.8619.7818.80%
constant_mean20.4629.2521.1718.46%
shuffled20.1230.6420.8421.82%
GM-DPO23.0630.5222.3414.80%

Both controls improve GPA over DPO, so some benefit comes from changing the overall optimization strength. GM-DPO gains a further 2.60 and 2.94 GPA points, while lowering WHR by 3.66 and 7.02 percentage points, respectively. The shuffled model obtains slightly more clipped credit than GM-DPO, yet makes substantially more weighted errors. This supports matching optimization strength to the actual limb error, beyond using a margin or larger weight alone.

The controls and discussion are included in Sec. 5.5 (lines 474–477) and Appendix C.3 (lines 1004–1018; Table 9).

[1] Amini, Vieira, and Cotterell. Direct Preference Optimization with an Offset. Findings of ACL 2024.

6) Broader video evaluation

this paper lacks evaluation coverage. it would be interesting to see whether teaching the model to ground better on fine-grained action alignment reduce hallucination on video centric benchmarks such as videohallucer, vidhalluc, eventhallucion; and improve in finegrained temporal tasks like tempcompass, tvbench, vinoground, vcgbench, temporalbench. I would recommend authors to eval on these benchmarks to ground the effectiveness of GMDPO loss. POPE is an object hallucintion benchmark which is not at all suitable here. also the performance of your method on videoMME is weak.

Response: Thank you for these recommendations. Following the two directions you suggested, we evaluated VideoHallucer [30] and VidHalluc [14] for video hallucination, and TempCompass [18] and VinoGround [34] for fine-grained temporal understanding. All five policies use Qwen3-8B, without benchmark-specific fine-tuning.

Scores are percentages; bold marks the best of our five policies. Results marked † come from the original benchmark papers. TempCompass CG uses candidate information.

The table shows VideoHallucer paired accuracy (VH), VidHalluc action-hallucination binary pairs (ACH pair) and temporal sequence hallucinations (TSH), TempCompass caption generation (TC-CG), and VinoGround group accuracy (VG-Group). Table 4 reports all nine metrics.

ModelVHACH pairTSHTC-CGVG-Group
DPO58.5678.6990.3366.6217.00
cDPO58.5878.3690.6766.5217.80
IPO58.4678.2792.0066.0217.20
SimPO58.6978.1191.3365.8217.20
GM-DPO58.9879.2092.8067.2318.60
GPT-4o †53.3066.7982.00—24.60
GPT-4o (CoT) †————35.00
Gemini-1.5-Pro †37.8059.1083.83—10.20
Qwen2-VL-72B †————17.40
Qwen2-VL-7B †————15.20
LLaVA-OneVision-Qwen2-7B †————14.60
PLLaVA-7B †38.10————
LLaVA-NeXT-Video-DPO-7B †32.00————
VideoLLaMA2-7B †—29.0926.17——
VideoChat2-7B †7.80——38.50—
Qwen-VL-Chat-7B †———37.30—
Video-LLaVA-7B †17.809.8727.1734.806.60

In the full results (Table 4), GM-DPO leads on eight of nine metrics, gaining 2.47 points on VidHalluc TSH, 0.61 on TempCompass CG, and 1.60 on VinoGround Group over DPO. These results support retained hallucination detection and temporal understanding alongside improved limb fidelity.

We also compared DPO and GM-DPO on two video-specialized VLMs using FlexBench (Appendix C.6, Table 14, lines 1143–1155):

BackboneMethodGPA ↑GPA_clip ↑Sub. ↑Non-sub. ↑TSA_w ↑WHR ↓
VideoLLaMA2-7BDPO10.3920.573.1521.259.6415.86%
VideoLLaMA2-7BGM-DPO12.1822.944.7023.4110.6113.03%
Tarsier2-Recap-7BDPO23.8230.4711.3942.4712.3810.67%
Tarsier2-Recap-7BGM-DPO25.5430.8913.2244.0113.348.02%

GM-DPO improves all six metrics on both backbones.

We have replaced the original evaluation in Sec. 5.6 with these four benchmarks (revised lines 479–492; Table 4) and moved the original Video-MME, POPE, and Video-ChatGPT results to Appendix C.4. Thank you again for the careful review and specific recommendations. They helped us make the evaluation more relevant to our claims and strengthen the paper.

References

[30] Wang et al. VideoHallucer: Evaluating Intrinsic and Extrinsic Hallucinations in Large Video-Language Models. 2024.

[14] Li, Im, and Fazli. VidHalluc: Evaluating Temporal Hallucinations in Multimodal Large Language Models for Video Understanding. CVPR 2025.

[18] Liu et al. TempCompass: Do Video LLMs Really Understand Videos?. Findings of ACL 2024.

[34] Zhang, Cai, and Lee. Vinoground: Scrutinizing LMMs over Dense Temporal Reasoning with Short Videos. 2024.

Reviewer_MCLLOpenReview ↗
Rating
3(+1)
Confidence
4(+2)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to improve the factual accuracy of fine-grained video captions generated by vision-language models (VLMs), focusing on limb laterality errors, incorrect person–action attribution, omissions, and fabricated actions. It introduces FlexBench, a benchmark comprising 660 video sequences, 3,105 shots, and 18,161 evaluation queries, with human-verified annotations supporting person- and shot-specific factual checks. The authors also propose Graded Physical Alignment (GPA), a metric that distinguishes correct and partially correct descriptions, omissions, and incorrect assertions through graded credits and penalties.

To improve caption factuality, the authors propose Graded Margin Direct Preference Optimization (GM-DPO), which uses predefined action-error severity levels to adjust both preference margins and loss weights. Using 125,000 constructed preference pairs, the authors compare GM-DPO with other preference optimization objectives across three VLM backbones, reporting GPA improvements of 2.02–3.40 points over standard DPO. The evaluation also includes component ablations, an agreement analysis between human annotators and the automated evaluator, and experiments on additional multimodal benchmarks.

Critical strengths
  1. FlexBench provides a potentially useful resource through its systematic attention to person identities, shot context, stationary limbs, and visibility, rather than only salient actions. Distinguishing omitted facts from incorrect assertions is valuable for analyzing the reliability of complete video descriptions. Human verification and the reported evaluator-agreement study provide additional support for the evaluation protocol.
  2. GM-DPO improves GPA and substantive-action scores across the three evaluated backbones. Comparisons with several preference objectives, component ablations, and repeated training runs provide evidence that the training strategy improves the evaluated captioning performance. My reservation concerns the incremental significance and interpretation of these improvements, rather than whether the reported gains exist.
Critical weaknesses

1. My primary concern is whether the proposed framework establishes a sufficiently distinct and consequential contribution beyond existing fine-grained motion-captioning approaches.

I recognize that the paper targets factuality in multi-person, multi-shot captions, rather than merely left/right limb classification. Captioning is a legitimate task in its own right. However, the paper does not sufficiently justify its reliance on dense captioning and preference optimization relative to approaches incorporating explicit pose or kinematic information. Prior work such as PeVL (CVPR 2024) demonstrates the value of pose-enhanced vision language representations for fine-grained action recognition, yet the paper provides neither a pose-assisted captioning baseline nor convincing evidence of its approach’s advantages. Pose estimation need not solve captioning alone for such a comparison to be informative.

Furthermore, the reported GPA improvements do not sufficiently distinguish better visual motion understanding from better caption-generation alignment and structural compliance. Targeted evaluations of laterality errors, person-action attribution, and cross-shot grounding are missing. While the results support improved captioning performance under the proposed protocol, they do not clearly establish which underlying capabilities improve or what significant new insight the framework provides beyond task-specific caption alignment. This substantially limits my assessment of the paper's scientific significance. Without clarifying the specific capability being improved and the advantages of the proposed formulation over more direct or structured alternatives, it is difficult to determine what fundamentally new understanding the work contributes beyond task-specific caption alignment.

2. My second concern is the limited novelty of the proposed contributions relative to existing work.

The methodological contribution of GM-DPO appears relatively incremental. Its central formulation combines a sample-dependent preference margin with severity-based loss weighting. However, previous work such as ODPO (ACL 2024), AlphaDPO (ICML 2025), and γ-PO (ACL 2025) has already explored preference-strength-dependent offsets or adaptive margins in DPO.

In particular, the margin component of GM-DPO can be interpreted as an instance of ODPO in which the offset is derived from a predefined physical-error severity score. The additional loss weighting constitutes a straightforward extension of this formulation.

While the proposed approach achieves improvements over standard DPO and several preference optimization baselines, the paper does not compare against these closely related methods or provide sufficient controls to establish the necessity of its specific severity assignments.

The training perturbations and primary evaluation also share the same error taxonomy. This does not imply data leakage or invalidate the results, but it makes independent validation particularly important. Video-MME, POPE, and Video-ChatGPT mainly support the retention of broader multimodal capabilities, while the human-agreement study evaluates the scoring protocol. Neither directly provides an independent comparison of DPO and GM-DPO captions under a separate motion-factuality protocol or a blinded, video-grounded human evaluation.

Consequently, it remains unclear whether GM-DPO represents a substantive methodological advance beyond applying established preference optimization mechanisms to a new task.

Questions for authors

Q1. Motivation and Necessity of the VLM-based Formulation The paper motivates its approach through limitations in fine-grained limb-motion understanding, particularly limb laterality confusion and incorrect person-action attribution. However, existing pose estimation methods provide explicit anatomical and kinematic representations, and pose-enhanced vision-language models have already been investigated for fine-grained action understanding.

What is the specific motivation for addressing these difficulties primarily through RGB-based VLM preference optimization rather than incorporating structured pose information?

More importantly, could the authors demonstrate the advantages of the proposed framework over a pose-guided VLM or a pose-enhanced captioning pipeline on the same captioning task?

Such a comparison would help establish whether the proposed approach offers meaningful advantages in factual accuracy, robustness, efficiency, or generality, rather than simply providing another way to generate descriptions of already identifiable movements.

This clarification would substantially influence my assessment of the necessity and significance of the proposed framework.

Q2. What capability actually improves? Could the authors report changes in laterality errors, person–action misattribution, cross-shot grounding errors, and fabricated actions, while accounting for output structure and descriptive coverage?

A controlled diagnostic on the same videos, or a blinded human assessment against the source videos, would help establish whether the gains reflect more reliable visual factual judgments or primarily changes in caption-generation behavior. An independently constructed motion-caption benchmark could provide complementary evidence.

Q3. Is meaningful severity assignment necessary? How should GM-DPO be positioned relative to ODPO-style preference objectives? Beyond the task-specific severity labels and additional reweighting, what is the independent methodological contribution?

Could the authors compare against appropriately tuned constant-margin/constant-weight controls and severity-permuted controls that preserve the severity distribution, using the same preference pairs, auxiliary NLL term, and training configuration? These comparisons would materially affect my assessment of the necessity of the graded formulation.

Q4. Backbone Selection and Generalization The alignment experiments are limited to Qwen2.5-7B, Qwen3-8B, and LLaVA-2-8B, covering two model families within a similar parameter range. Could the authors explain the rationale behind this selection and provide evidence that GM-DPO also benefits other model families? In particular, a comparison between DPO and GM-DPO on a video-specialized backbone, such as Tarsier2-Recap-7B or VideoLLaMA2-7B, both already evaluated in Table 1, would help establish whether the improvements generalize beyond the selected backbones.

Additional Clarification The scale comparison in Figure 3B should consistently distinguish video sequences from shots: FlexBench contains 660 sequences and 3,105 shots. The 215 videos attributed to KPM-Bench (HA) correctly refer to its evaluation subset and should not be treated as an incorrect count of the entire KPM resource. Training resources, evaluation resources, and shot counts should be explicitly separated.

17 comments

Thank you for the careful review. We have added human error analysis, severity controls, four external benchmarks, and experiments on both suggested video backbones. Since the five questions cover the two weaknesses, we reply in that order. Manuscript revisions are marked in blue.

1) Motivation, dense captions, and pose guidance

Q1. Motivation and Necessity of the VLM-based Formulation The paper motivates its approach through limitations in fine-grained limb-motion understanding, particularly limb laterality confusion and incorrect person-action attribution. However, existing pose estimation methods provide explicit anatomical and kinematic representations, and pose-enhanced vision-language models have already been investigated for fine-grained action understanding.

What is the specific motivation for addressing these difficulties primarily through RGB-based VLM preference optimization rather than incorporating structured pose information?

More importantly, could the authors demonstrate the advantages of the proposed framework over a pose-guided VLM or a pose-enhanced captioning pipeline on the same captioning task?

Such a comparison would help establish whether the proposed approach offers meaningful advantages in factual accuracy, robustness, efficiency, or generality, rather than simply providing another way to generate descriptions of already identifiable movements.

This clarification would substantially influence my assessment of the necessity and significance of the proposed framework.

Thank you for asking us to make the motivation and comparison more concrete.

(a) Our goal is end-to-end caption factuality. As stated in Sec. 3 and Sec. 4 (revised lines 138–140 and 214–215), we aim to improve limb-motion accuracy and reduce fabricated actions within complete video captions, while retaining the model's broader descriptive ability. Graded preferences distinguish omissions from false claims and directly improve the assertions made from RGB video. GM-DPO retains the original captioning architecture and requires no additional pose encoder, tracking module, or structured pose input at inference.

(b) What “dense” means in this setting. Here, a dense caption is a coherent account of the scene, people, dialogue, camera transitions, and unfolding events, with limb details embedded in that narrative. KPM-Bench describes its targets as “dense motion-centric captions” [1]; our aim is to preserve the surrounding description while correcting its limb facts. Our training pairs modify targeted limb facts while preserving surrounding content, so the model learns motion accuracy within a general video caption.

Appendix A.5, Figure 10 follows an invitation, seating, napkin handling, food offering, and conversation across two shots. As the view changes to the woman, the man's arms remain in the foreground; his actions must stay connected to the same person across shots. On Qwen3-8B, GM-DPO retains caption length and shot structure close to full SFT (1,610 vs. 1,634 words; 89.8% vs. 89.7% SMR), while increasing GPA from 21.08 to 23.06. Figure 7 illustrates fewer fabricated limb actions. We have clarified this task distinction in Sec. 2.1 (lines 92–96).

(c) Pose guidance: our exploration and next step. We agree that pose can provide useful complementary evidence, as PeVL [2] demonstrates for fine-grained action recognition. Following your suggestion, we tested automatic RTMW whole-body skeleton inputs on 100 sequences: 40 single-shot, 40 multi-shot with 1–2 people, and 20 multi-shot with 5–6 people. For each fixed checkpoint, we compared RGB+pose with RGB+duplicate-RGB, holding frames, resolution, and visual-token budget fixed. No GT boxes or identities were supplied. Differences below are pose minus control:

PolicyΔGPA [95% CI]ΔTSA_wΔWHR (percentage points)
DPO+1.41 [-2.29, 5.29]-1.15-1.20
GM-DPO-0.96 [-3.70, 1.89]+0.49+0.53

Intervals use 5,000 paired sequence resamples. The effects are mixed, with GPA intervals spanning zero. This exploration tests input augmentation of existing captioners; a jointly trained pose-fusion model remains an open comparison. Your suggestion motivates further work combining trained pose representations with graded alignment to improve grounding while preserving the broader narrative.

[1] Lin et al. KPM-Bench: A Kinematic Parsing Motion Benchmark for Fine-grained Motion-centric Video Understanding. 2026.

[2] Zhang et al. PeVL: Pose-Enhanced Vision-Language Model for Fine-Grained Human Action Recognition. CVPR 2024.

2) Factual errors, coverage, and independent evaluation

Q2. What capability actually improves? Could the authors report changes in laterality errors, person–action misattribution, cross-shot grounding errors, and fabricated actions, while accounting for output structure and descriptive coverage?

A controlled diagnostic on the same videos, or a blinded human assessment against the source videos, would help establish whether the gains reflect more reliable visual factual judgments or primarily changes in caption-generation behavior. An independently constructed motion-caption benchmark could provide complementary evidence.

Thank you for asking us to distinguish factual improvements from changes in caption structure.

(1) Human error analysis. We manually compared Qwen3-8B DPO and GM-DPO captions on 300 FlexBench sequences, comprising 1,436 constituent video clips:

Human-assessed categoryRelative change vs. DPO
Laterality errors↓ 7.05%
Explicit person–action attribution errors↓ 5.84%
Fabricated or incorrectly substituted actions/limb states↓ 11.85%
Cross-shot identity association errors↓ 3.06%

These are relative error reductions, calculated as (NDPO−NGM)/NDPO(N_{\mathrm{DPO}}-N_{\mathrm{GM}})/N_{\mathrm{DPO}}. Cross-shot errors concern person identity, e.g., describing the same woman in yellow as a new person after a cut. Given the short rebuttal period, we assessed 300 sequences; we will complete all 660 and update the manuscript. Current results are in Appendix C.5, Table 13 (lines 1122–1140), cited in Sec. 5.4 (lines 426–428).

(2) Coverage and structure. The full-benchmark results show both more correct content and fewer false assertions:

MethodGPA ↑GPA_clip ↑Substantive ↑TSA_w ↑WHR ↓
DPO19.6628.867.1619.7818.80%
GM-DPO23.0630.529.3022.3414.80%

GPA_clip increases by 1.66 points and the penalty deduction decreases from 9.20 to 7.46, together accounting for the 3.40-point GPA gain. WHR falls by 4.00 percentage points.

Checkpoints are selected by highest GPA on FlexBench among ten candidates near the training-loss minimum, not by SMR. Both methods use the same auxiliary NLL coefficient but reach their best GPA at different steps. We therefore evaluated DPO at GM-DPO's selected step:

Qwen3-8B method / checkpointGPA ↑SMR (%) ↑
DPO, best GPA19.6655.6
DPO, GM-DPO's selected step19.3481.6
GM-DPO, best GPA23.0689.8

At the matched step, DPO's SMR increases to 81.6%, while GPA falls by 0.32 points. We have clarified the selection rule in Appendix B (lines 850–851).

Full SFT has similar SMR to GM-DPO (89.7% vs. 89.8%), yet GM-DPO improves GPA from 21.08 to 23.06 and substantive-action score from 6.92 to 9.30 at similar lengths (1,634 vs. 1,610 words). This supports improved limb factuality beyond format compliance.

(3) Independent evaluation. We also evaluated all five Qwen3-8B policies on four external benchmarks without benchmark-specific fine-tuning. Scores are percentages; bold marks the best of our policies. Values marked † are cited from the original benchmark papers.

ModelVHACH pairTSHTC-CGVG-Group
DPO58.5678.6990.3366.6217.00
cDPO58.5878.3690.6766.5217.80
IPO58.4678.2792.0066.0217.20
SimPO58.6978.1191.3365.8217.20
GM-DPO58.9879.2092.8067.2318.60
GPT-4o †53.3066.7982.00—24.60
GPT-4o (CoT) †————35.00
Gemini-1.5-Pro †37.8059.1083.83—10.20
Qwen2-VL-72B †————17.40
Qwen2-VL-7B †————15.20
LLaVA-OneVision-Qwen2-7B †————14.60
PLLaVA-7B †38.10————
LLaVA-NeXT-Video-DPO-7B †32.00————
VideoLLaMA2-7B †—29.0926.17——
VideoChat2-7B †7.80——38.50—
Qwen-VL-Chat-7B †———37.30—
Video-LLaVA-7B †17.809.8727.1734.806.60

VideoHallucer (VH) and VidHalluc (ACH binary pairs/TSH) assess hallucinations; TempCompass (TC-CG) and VinoGround (VG-Group) assess temporal understanding. TC-CG uses candidate information and Qwen3.8-Max to answer questions from generated captions. The full nine-metric results appear in Sec. 5.6 (lines 479–492; Table 4), where GM-DPO leads eight of nine metrics.

3) Relation to ODPO and the role of severity

Q3. Is meaningful severity assignment necessary? How should GM-DPO be positioned relative to ODPO-style preference objectives? Beyond the task-specific severity labels and additional reweighting, what is the independent methodological contribution?

Could the authors compare against appropriately tuned constant-margin/constant-weight controls and severity-permuted controls that preserve the severity distribution, using the same preference pairs, auxiliary NLL term, and training configuration? These comparisons would materially affect my assessment of the necessity of the graded formulation.

Thank you for identifying the controls needed to test whether the severity assigned to each pair matters.

(1) Relationship to ODPO and adaptive margins. Let zz denote the DPO logit. The preference terms are

ℓODPO=−log⁡σ(z−δ),ℓGM=−(1+αΔS)log⁡σ(z−γΔS).\ell_{\mathrm{ODPO}}=-\log\sigma(z-\delta),\qquad \ell_{\mathrm{GM}}=-(1+\alpha\Delta S)\log\sigma(z-\gamma\Delta S).

When α=0\alpha=0 and δ=γΔS\delta=\gamma\Delta S, our preference term has the same mathematical form as ODPO [1]. The full GM-DPO objective improves over this severity-derived margin-only variant on all three backbones: GPA rises from 13.20 to 14.06 on Qwen2.5-7B, 20.07 to 23.06 on Qwen3-8B, and 24.55 to 25.40 on LLaVA-2-8B; GPA_clip and TSA_w also improve (Table 3).

AlphaDPO [2] adapts margins through reference-distribution reparameterization, while γ-PO [3] calibrates margins to favor confident preference pairs. Our contribution concerns how a known local factual error should affect learning from an otherwise complete caption: its severity controls both the required preference gap and the loss weight. The contribution is this correspondence and its empirical value for limb factuality. We have added these connections in Sec. 2.2 (lines 107–123) and clarified the conditional relationship to ODPO in Sec. 5.5 (lines 468–472).

(2) Does the severity assignment matter? We added two controls with the same preference pairs, training seeds, auxiliary NLL term, and evaluation protocol:

  • constant_mean: replace each severity with its training-set mean (1.10029924), yielding margin 0.22006 and weight 1.11003, matched to GM-DPO's averages.
  • shuffled: permute severities across pairs, retaining their distribution and margin–weight relationship while breaking the link to the actual error.

Both retain α=0.1\alpha=0.1, γ=0.2\gamma=0.2, β=0.1\beta=0.1, and NLL coefficient 1.0.

MethodGPA ↑GPA_clip ↑TSA_w ↑WHR ↓
DPO19.6628.8619.7818.80%
constant_mean20.4629.2521.1718.46%
shuffled20.1230.6420.8421.82%
GM-DPO23.0630.5222.3414.80%

Both controls improve GPA over DPO, indicating that changing the overall margin and weight contributes. GM-DPO gains a further 2.60/2.94 GPA points and reduces WHR by 3.66/7.02 percentage points, respectively. Shuffling obtains slightly higher clipped credit but incurs substantially more weighted errors. This supports assigning stronger optimization to the corresponding factual error: preserving average strength, or even the severity distribution, does not reproduce the full result. The constant control evaluates matched average strength; it is not a sweep over all fixed margins and weights.

The controls are reported in Sec. 5.5 (lines 474–477) and Appendix C.3 (lines 1004–1018; Table 9). The human analysis and independent benchmarks in our response to Q2 further test whether the gains extend beyond the shared rubric.

[1] Amini, Vieira, and Cotterell. Direct Preference Optimization with an Offset. Findings of ACL 2024.

[2] Wu et al. AlphaDPO: Adaptive Reward Margin for Direct Preference Optimization. ICML 2025.

[3] Sun et al. Robust Preference Optimization via Dynamic Target Margins. Findings of ACL 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.