PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yanan Wang, Tingsong Li, Kaixun Jiang, Chongyang Zhong, Chenwei Xoe, Zhaohe Liao

arXiv:2609.36628v1cs.CV

TL;DR

VLM captions can miss which person and limb perform an action, and existing preference labels and evaluation do not adequately distinguish omissions from incorrect assertions or systematically assess each person across shots. The paper introduces FlexBench and GPA for graded, context-specific caption evaluation, and GM-DPO for severity-aware preference training. Across three backbones, GM-DPO achieves the highest substantive-action and GPA scores among the evaluated preference objectives while sustaining long-form output.

  • Problem

    VLMs can generate rich captions but may misidentify the person or limb performing an action, while systematic limb-level assessment of every identifiable person across shots remains insufficiently addressed.

    Source: §1 · pp. 1–2

  • Method

    FlexBench uses human-verified identities and reference-derived, person- and shot-specific checklists, while GPA grades omissions and incorrect claims and GM-DPO weights preference updates by action-error severity.

    Source: §1 · p. 2

  • Results

    Across three backbones, GM-DPO achieves the best substantive-action and GPA scores among tested preference objectives while sustaining long-form output.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper’s evaluation and preference objective focus on limb-motion fidelity within complete, richly detailed video captions.

    Source: §1 · p. 2

Abstract

from arXiv · show

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 in their person and shot contexts. Our Graded Physical Alignment score (GPA) awards credit for correct content and deducts points for incorrect or fabricated actions, making these errors explicit in the aggregate score. Building on this rubric, we propose Graded Margin Direct Preference Optimization (GM-DPO), which assigns stronger 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.

1 INTRODUCTION

The introduction frames limb-motion captioning as a problem of grounding actions in the correct person, limb, and shot, while distinguishing missing details from false assertions. It presents FlexBench and GPA for evaluation and GM-DPO for severity-aware preference training.

Motivation

  • Limb-motion captions must keep action details attached to the correct person and shot, including which hand acts and how another supports the action.The introduction gives opening a box as an example of an action whose limb-specific details remain unresolved by recognizing the event alone.Source: §1 · pp. 1–2
  • Binary preference labels do not distinguish omissions from limb confusions or fabricated actions, although missing details and false assertions differ as training annotations.Missing information leaves supervision incomplete, whereas false assertions introduce incorrect video–text associations.Source: §1 · p. 2

Contributions

  • FlexBench uses human-verified identities and reference-derived checklists to cover each person’s actions in corresponding shots, including stationary limbs and limb visibility.Its GPA awards full or partial credit, assigns zero credit to omissions, and deducts points for incorrect or fabricated actions.Source: §1 · p. 2
  • GM-DPO uses action-error severity in its preference margins and loss weights, and achieves the best substantive-action and GPA scores among tested objectives across three backbones while sustaining long-form output.FlexBench spans 3,105 shots and 18,161 queries; GPA distinguishes incomplete coverage from incorrect motion claims.Source: §1 · p. 2

2 RELATED WORK

The related work situates limb-motion captioning alongside video understanding, preference optimization, and motion-caption evaluation. It distinguishes this paper’s focus on dense, multi-shot captions with per-person limb assessment and severity-aware grading.

2.1 FROM VIDEO UNDERSTANDING TO LIMB-MOTION CAPTIONING

  • FlexBench targets limb-level accuracy in dense captions covering every identifiable person across shots, including stationary limbs and visibility, complementing prior fine-grained motion-captioning work.KPM-Bench advances fine-grained motion captioning with multi-level motion representations, SFT, and GRPO.Source: §2.1 · p. 2

2.2 PREFERENCE OPTIMIZATION FOR VIDEO CAPTIONING

  • GM-DPO uses a limb-motion error rubric to set both preference margins and loss weights, distinguishing incomplete action content from incorrect or fabricated actions.Prior video-captioning preference approaches include DPO after supervised training, metric-based preference selection, and losses over increasingly disrupted captions.Source: §2.2 · pp. 2–3

2.3 MOTION BENCHMARKS AND CAPTION EVALUATION

  • FlexBench assesses limb claims per person and shot, including stationary and hidden states, while GPA grades credits and penalties and a temporal suite assesses consecutive-action coverage.Prior motion evaluations include multiple-choice queries, QA-caption tasks, and free-form caption scoring or sequence matching.Source: §2.3 · p. 3

3 FLEXBENCH: FINE-GRAINED LIMB PHYSICAL BENCHMARKING

FlexBench evaluates fine-grained limb-motion fidelity in complete multi-shot captions using verified references and person- and shot-specific checks. Its graded evaluation covers individual action claims, temporal sequences, omissions, and incorrect or fabricated actions.

3 FLEXBENCH: FINE-GRAINED LIMB PHYSICAL BENCHMARKING

  • FlexBench pairs verified references with detailed checklists to audit each person’s active limbs and exact shot locations within complete multi-shot captions.The benchmark’s sequences are curated from untrimmed cinematic scenes and CC-licensed YouTube streams; manual shot-boundary calibration yields 660 sequences and 3,105 shots.Source: §3 · p. 3, §3.1 · p. 3

3.2 PER-PERSON COVERAGE AND CHECKLIST EVALUATION

  • Human-verified captions structure each individual’s actions by shot across sequences spanning 1–10 shots and 0–23 actors, while accounting for stationary or hidden limb states.Checklist items cover actions, limb states, object interactions, and temporal order, tied to a specific person and shot; full credit requires joint alignment across action, person, and shot.Source: §3.2 · pp. 3–4

3.3 GRADED EVALUATION

  • FlexBench spans 18,161 queries: 7,752 substantive-action queries, 5,758 non-substantive-posture queries, and 4,651 temporal-sequence queries.The suites measure individual action-claim fidelity and joint coverage across sequences.Source: §3.3 · p. 4
  • GPA grades action items from −1 to 1 to distinguish correct, partial, omitted, contradictory, or fabricated actions, and reports scores from −100 to 100.The clipped variant credits content without negative penalties; weighted hallucination rate combines the proportions of items graded −0.5 and −1. Temporal queries assess consecutive substantive actions, with TSAw weighting two-action, three-action, and full-chain queries.Source: §3.3 · p. 4

4 GM-DPO: GRADED MARGIN PREFERENCE OPTIMIZATION

GM-DPO addresses uniform binary preference treatment by grading physical-error severity and using that severity to shape optimization. The section describes the baseline limitation, perturbation labels, severity-modulated updates, and construction of preference pairs.

4.1 PRELIMINARIES AND LIMITATIONS OF STANDARD DPO

  • Standard DPO scales updates by prediction error without accounting for defect severity, applying identical penalties to benign omissions and fatal chirality inversions.The paper characterizes this as failing to penalize critical physical violations without over-correcting harmless variations.Source: §4.1 · p. 5

4.2 PERTURBATION SEVERITY FROM THE GRADING RUBRIC

  • Each preference pair applies one error category consistently to targeted action descriptions in a selected shot while retaining the rest of the caption.The perturbation grade is local severity, not the GPA difference between complete captions; it uses the unscaled rubric.Source: §4.2 · p. 5

4.3 GM-DPO OBJECTIVE AND OPTIMIZATION DYNAMICS

  • GM-DPO modulates updates with a severity-dependent margin and gradient weight, sustaining gradients until the margin is cleared and magnifying updates for more severe physical flaws.The weight is ω(∆S) = 1 + α∆S; the margin shift is γ∆S. The passage describes scaling up to (1 + 2α)× for confabulations.Source: §4.3 · p. 5

4.4 PREFERENCE PAIR CONSTRUCTION

  • The preference-pair pipeline perturbs ground-truth dense captions across graded physical-error regimes using rule-based spatial editing and schema-constrained LLM rewriting.It retains untargeted descriptions and constrains rewrites to preserve the original style; reported error types include laterality neglect, chirality inversion, motion omission, subaction drift, and action confabulation.Source: Figure 4 · p. 6, §4.4 · p. 6

5 EXPERIMENTS

The experiments evaluate fine-grained limb-motion fidelity, shot structure, and hallucination-aware alignment across models and benchmarks. GM-DPO improves substantive-action and penalty-aware scores across three backbones, with gains in sequence coverage and maintained long-form output.

5.1 EXPERIMENTAL SETUP

  • Training compares GM-DPO with four preference objectives on three backbones, using three seeds; evaluation reports GPA, TSAw, caption length, and shot-structure match rate.The Qwen3-8B SFT baseline provides a caption-imitation reference.Source: §5.1 · p. 6, §5 · p. 6

5.3 MAIN ALIGNMENT RESULTS

  • Across three backbones, GM-DPO leads preference methods in substantive-action and GPA scores, improving GPA over DPO by 2.02, 3.40, and 2.71 points.The main alignment comparison is reported in Table 2.Source: Table 2 · p. 7, §5.3 · p. 8

5.2 EVALUATING EXISTING MODELS ON FLEXBENCH

  • Zero-shot substantive-action scores trail non-substantive scores across evaluated models; Qwen-3.8-Max scores 33.50 versus 46.88, and Gemini-3.1-Pro scores 17.80 versus 42.21.Table 1 reports zero-shot performance on FlexBench.Source: §5.2 · p. 7, Table 1 · p. 7

5.4 DIAGNOSTIC ANALYSIS AND EVALUATOR FIDELITY

  • On Qwen3-8B, GM-DPO reduces weighted hallucination rate from 18.8% to 14.8% while increasing GPA by 3.40 points.Across backbones it has the largest GPA increase among compared methods while reducing weighted hallucination rate; the advantage is not exclusive to it occupying the upper-left figure region.Source: Figure 5 · p. 8, §5.4 · p. 9
  • Across five suites, automated evaluation achieves 95.7%–98.9% exact agreement with human consensus and quadratic weighted κw values of 0.929–0.988.Figure 6 shows row-normalized agreement matrices between human consensus and the LLM judge.Source: Figure 6 · p. 8, §5.4 · p. 9

5.3 MAIN ALIGNMENT RESULTS

  • GM-DPO improves TSAw over DPO by 1.45, 2.56, and 3.68 points, while leading Seqchain on all three backbones.Sequence queries assess consecutive actions in their shot contexts, but do not isolate temporal reasoning from action accuracy.Source: §5.3 · p. 8
  • On Qwen3-8B, GM-DPO matches full-SFT in caption length (1610 vs. 1634) and shot-structure match rate (89.8% vs. 89.7%).Preference-tuned variants preserve descriptive length while improving structure match over base models.Source: §5.3 · p. 8

5.5 ABLATION ON GM-DPO LOSS FORMULATION

  • Combining severity-dependent margins and weights yields the highest GPA, GPAclip, and TSAw in every ablation group; on Qwen3, joint optimization reaches 23.06 GPA.Table 3 reports the ablation across backbones.Source: §5.5 · p. 9, Table 3 · p. 9

6 CONCLUSION

The paper concludes that fine-grained evaluation and severity-aware preference learning improve limb-motion fidelity in dense captions, while noting unresolved monocular ambiguity and unvalidated caption dimensions. The appendix outlines benchmark curation and annotation verification.

6 CONCLUSION

  • The framework combines fine-grained evaluation with severity-aware preference learning to improve limb-motion fidelity, but does not establish accuracy for camera angles, camera motion, or sound.Monocular ambiguity remains a challenge.Source: §6 · p. 10

APPENDIX OVERVIEW

  • The appendix describes benchmark selection and annotation quality assurance, training configurations and baselines, further experiments, and limitations including monocular ambiguity.It also identifies extensions to the grading rubric and broader dense-caption evaluation as future directions.Source: APPENDIX OVERVIEW · p. 12

A.1 VIDEO FILTERING AND SEQUENCE CURATION

  • Curation filtered an initial pool of over 56,000 candidate sequences to a final benchmark of 660 verified videos spanning 3,105 continuous shots.Experts manually selected sequences for scenario diversity, actor density, and rich limb-level interactions.Source: §A.1 · p. 12

A.2 HUMAN-IN-THE-LOOP ANNOTATION AND MULTI-ROUND VERIFICATION

  • Eighteen trained annotators refined captions under a standardized rubric, with three rounds of blind cross-verification and 94.2% raw inter-annotator agreement.Edge-case chirality and action states were finalized by panel consensus; positive QA probes received two rounds of independent human inspection.Source: §A.2 · p. 12

A.3 LINGUISTIC NATURALNESS AND NON-SPURIOUSNESS OF PREFERENCE PAIRS

Preference-pair edits were confined to local limb kinematics while surrounding wording and style were frozen; blind human judgments could not distinguish the pairs from chance.

  • Perturbations changed local limb kinematics while keeping sentence structure, appearance, and stylistic register fixed.Laterality changes used deterministic token substitutions; temporal omissions, subaction variations, and phantom actions were synthesized under schema constraints.Source: §A.3 · p. 12
  • Ten independent annotators evaluated 500 randomly sampled blind pairs, producing 5,000 total judgments.Source: §A.3 · p. 12
  • 49.60% identification accuracy matched chance (binomial p = 0.5813 > 0.05), with sample-level and evaluator-level results close to the Binomial(10, 0.5) baseline.The study reports no detectable identification of human-refined versus LLM-modified descriptions.Source: Figure 7 · p. 13, §A.3 · p. 13

B EXPERIMENTAL SETUP AND BASELINE OBJECTIVES

Across three backbones and seed configurations, GM-DPO leads preference methods on key FlexBench metrics, and paired GPA gains remain positive across runs.

B EXPERIMENTAL SETUP AND BASELINE OBJECTIVES

  • Experiments use three joint training and data-order seed configurations, and reported model results are arithmetic means over the runs.Alignment uses full-parameter instruction tuning with vision encoders and cross-modal projectors frozen, plus an auxiliary next-token prediction loss.Source: B EXPERIMENTAL SETUP AND BASELINE… · p. 13, §C.1 · p. 13

C.1 ROBUSTNESS ACROSS RANDOM SEEDS

  • GM-DPO leads preference methods in substantive-action accuracy, GPA, and Seqchain on every backbone in all three runs.It also consistently leads GPAclip and TSAw on Qwen3-8B and LLaVA-2-8B; on Qwen2.5-7B, IPO retains higher mean GPAclip and TSAw.Source: §C.1 · p. 14, Table 6 · p. 14
  • All paired GPA gains over DPO are positive, with pointwise 95% confidence intervals above zero.With only three runs, the paired-difference assumption cannot be reliably checked, so the intervals are supplementary evidence.Source: §C.1 · p. 14, Table 7 · p. 14

C.2 FULL FINE-GRAINED KINEMATIC AND SEQUENTIAL ABLATION

Ablations show that severity-dependent margins and weights contribute differently, while combining them yields the strongest GPA, credited-content, and sequence-coverage results.

C.2 FULL FINE-GRAINED KINEMATIC AND SEQUENTIAL ABLATION

  • Disabling both severity-dependent components recovers the DPO baseline while retaining the auxiliary NLL term.Source: §C.2 · p. 14, Table 8 · p. 15
  • Margin alone improves GPA over DPO by 1.16, 0.41, and 1.86 points on Qwen2.5-7B, Qwen3-8B, and LLaVA-2-8B, respectively.It also improves GPAclip and TSAw on all three backbones; substantive-action scores rise from 3.92 to 5.66, 7.16 to 8.76, and 10.66 to 11.88, respectively.Source: §C.2 · p. 14, Table 8 · p. 15
  • Weighting alone raises GPA by 0.25, 0.27, and 1.08 points across the three backbones, but does not consistently improve credited content or sequence coverage.On Qwen2.5-7B, TSAw falls from 17.01 to 16.52; on LLaVA-2-8B, GPAclip falls from 29.15 to 28.97.Source: §C.2 · pp. 14–15, Table 8 · p. 15
  • 23.06 GPA on Qwen3-8B is higher with both components than with weighting alone (19.93) or margin alone (20.07).The combined objective achieves the highest GPA, GPAclip, and TSAw in every ablation group; on LLaVA-2-8B it raises TSAw from 25.27 to 26.34 over margin alone.Source: §C.2 · p. 15, Table 8 · p. 15

C.3 EXTENDED DOWNSTREAM GENERALIZATION BENCHMARKS

GM-DPO remains competitive across three additional multimodal benchmarks, while a qualitative example illustrates improved aggregate fidelity without eliminating laterality errors.

C.3 EXTENDED DOWNSTREAM GENERALIZATION BENCHMARKS

  • 60.62%, 62.72%, and 65.39% are GM-DPO’s Video-MME overall accuracies, improving over the corresponding base models by 1.25, 1.02, and 0.69 percentage points.It also exceeds DPO on all three duration splits for every backbone.Source: §C.3 · p. 15, Table 9 · p. 16
  • 87.56, 89.40, and 88.97 are GM-DPO’s POPE Overall Acc scores on Qwen2.5-7B, Qwen3-8B, and LLaVA-2-8B, respectively, the highest among aligned variants for each backbone.It also attains top balanced F1 scores of 86.15, 89.02, and 88.41, respectively.Source: §C.3 · p. 15, Table 10 · p. 16
  • 2.51, 2.75, and 2.57 are GM-DPO’s Video-ChatGPT Overall scores, exceeding DPO by 0.01, 0.05, and 0.06, respectively.GM-DPO leads Consistency on all three backbones and has the best or tied-best Temporal Reasoning scores, though it does not lead every dimension.Source: §C.3 · pp. 15–16, Table 11 · p. 17

C.4 QUALITATIVE VISUALIZATIONS AND CASE STUDIES

  • In the illustrated sequence, GM-DPO scores +1.5 versus -2.5 for the base model and -1.5 for DPO, but still makes two laterality errors.The example demonstrates improved aggregate action fidelity, not error-free limb grounding.Source: §C.4 · p. 16, Figure 8 · p. 17

D LIMITATIONS AND FUTURE WORK

The paper identifies limits in its rubric, monocular-video setting, and dense-caption evaluation, despite consistent gains across three VLM backbones. It calls for further investigation of these boundaries and richer grading schemes.

Extending the Grading Rubric to Broader Motion Domains

  • The five-level rubric targets identifiable limb actions; continuous, multi-phase activities need criteria for whole-body transitions, movement speed, and deformable-object interactions.Such activities may involve overlapping errors that a single perturbation category cannot capture.Source: D LIMITATIONS AND FUTURE WORK · pp. 17–18
  • Future work could investigate richer grading schemes and their agreement with human judgments while preserving the distinction between incomplete coverage and incorrect assertions.Source: D LIMITATIONS AND FUTURE WORK · p. 18

Ambiguity in Monocular Video

  • Monocular inputs lack depth measurements, and self-occlusion, extreme viewpoints, and camera cuts can obscure limb identity and motion.These ambiguities may contribute to residual errors, but experiments do not isolate their effects from model-perception or caption-generation limits.Source: D LIMITATIONS AND FUTURE WORK · p. 18

Scope of Dense Caption Evaluation

  • Dense-caption evaluation covers limb-motion fidelity, per-person coverage, and shot structure, but not camera, scene, or sound descriptions.Caption length and structural compliance do not establish the accuracy or completeness of other caption content.Source: D LIMITATIONS AND FUTURE WORK · p. 18

Related papers

Loading 2609.36628v1…