Paper deep dive
SemReward-VL: Semantic Reward-Guided Video-Language Adaptation for Developmental Behavior Assessment
De Jiang, Shuo Zhang, Kehong Yuan, Hongen Liao
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Developmental screening videos show how children perform specific behaviors, but clinical records usually contain outcomes rather than descriptions of what happened. We present SemReward-VL, which learns to describe item-specific behavior from these outcomes. A vision-language model generates a description, and a frozen language model scores its agreement with the clinical outcome, relevance to the item, abstention on unrelated video-item pairs, and clarity. Group relative policy optimization (GRPO) updates LoRA adapters using this semantic reward. On 13,379 videos covering 41 items, the method improves aggregate accuracy and the number of items with recall above 0.5. Errors remain in temporal direction, duration, and age-specific interpretations of behavior.
Tags
Links
- Source: https://arxiv.org/abs/2609.33082v1
- Canonical: https://arxiv.org/abs/2609.33082v1
Trouble viewing inline? Open PDF directly →
Full Text
26,033 characters extracted from source content.
Expand or collapse full text
SEMREWARD-VL: SEMANTIC REWARD-GUIDED VIDEO-LANGUAGE ADAPTATION FOR DEVELOPMENTAL BEHAVIOR ASSESSMENT De Jiang 1 , Shuo Zhang 2 , Kehong Yuan 1 , Hongen Liao 2,∗ 1 Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China; 2 School of Biomedical Engineering, Tsinghua University, Beijing, China. ABSTRACT Developmental screening videos show how children perform specific behaviors, but clinical records usually contain outcomes rather than descriptions of what happened. We present SemReward-VL, which learns to describe item-specific behavior from these outcomes. A vision-language model generates a description, and a frozen language model scores its agreement with the clinical outcome, relevance to the item, abstention on unrelated video–item pairs, and clarity. Group relative policy optimization (GRPO) updates LoRA adapters using this semantic reward. On 13,379 videos covering 41 items, the method improves aggregate accuracy and the number of items with recall above 0.5. Errors remain in temporal direction, duration, and age-specific interpretations of behavior. Index Terms— video-language model, developmental behavior assessment, outcome supervision, GRPO, parameter-efficient adapta- tion 1. INTRODUCTION Naturalistic developmental videos capture gross-motor, fine-motor, language, adaptive, and social behaviors. Each clinical record in- cludes an age-specific item, filming guidance, a pass criterion, a clinician-reviewed outcome, and an uploaded video. Most records lack a description of the behavior visible in the video. The outcome states the clinical decision but leaves its visual evidence implicit. An item decision often depends on a small detail within a longer action. Walking backward requires the direction of movement, stand- ing on one foot requires a sustained posture, and giving an object on request requires both the action and its social context. A broad account of the clip may describe the child accurately yet omit the evidence needed for the item. The item criterion must therefore guide both what the model describes and how the description is interpreted. The same pass/fail outcome can arise from different visible be- haviors. A failed item may reflect an absent or incomplete action, occlusion, or a behavior that was never elicited. A direct classifier p(y | x,c)compresses these cases into one label. A single caption target is also unsuitable when several descriptions of a clip are valid. We instead generate an item-specific description before making the decision. When the video lacks the requested evidence, the model can respond “unable to assess.” Vision-language learning has progressed from transferable image- text representations such as CLIP and ALIGN [1,2] to generative mul- timodal models including Flamingo and BLIP-2 [3,4]. Instruction- tuned systems such as InstructBLIP, LLaVA, and MiniGPT-4 made free-form visual description a practical interface [5,6,7]. VideoChat, Video-ChatGPT, and Video-LLaMA extended this interface to tem- poral inputs [8,9,10], while LLaMA-VID, Video-LLaVA, and Vide- ∗ Corresponding author: Hongen Liao (liao@tsinghua.edu.cn). oLLaMA 2 improved long-context compression, alignment, and spa- tiotemporal modeling [11,12,13]. General-purpose VLMs such as Qwen2-VL, InternLM-XComposer2.5, and LLaVA-OneVision fur- ther broaden resolution, context length, and task transfer [14,15,16]. Video understanding still depends on temporal evidence. MVBench measures video reasoning skills [17]. MovieChat and MA- LMM use memory to retain evidence [18,19], VideoLLM-online processes streams incrementally [20], and video distillation improves video-language representations [21]. TimeChat and TempCompass document difficulties with temporal order, direction, and duration [22,23]. In biomedical applications, LLaVA-Med and Med-Flamingo adapt multimodal assistants to medical data [24,25]. BiomedCLIP learns biomedical representations [26], and CARE-VL explores autism-related screening [27]. Video-R1 and R1-VL apply group- relative reinforcement learning to multimodal reasoning [28,29]. We use clinician outcomes to train a description layer for developmental assessment. We model assessment as two coupled semantic transformations: behavior description and behavior decision. A VLM first converts the video and item criterion into an inspectable description; a frozen LLM then interprets that description under the criterion. The same LLM also scores sampled descriptions for outcome consistency, item relevance, unrelated-pair abstention, and generation quality. These semantic rewards allow GRPO to adapt the VLM with rank-16 LoRA while leaving the visual backbone and judge frozen. Figure 1 summa- rizes the training and inference paths. The description is the point at which visual recognition and clinical interpretation meet. It can state that an action occurred while also recording its direction, duration, or relation to an object. A clinician can then inspect whether the described behavior supports the item decision. During training, the outcome guides this intermediate description through the semantic reward. At inference, the decision model reads the description rather than the raw video, so errors in observation can be distinguished from errors in applying the criterion. Our contributions are: 1) We formulate item-level developmental behavior assessment as an outcome-supervised description–decision problem, introducing an inspectable natural-language evidence layer without requiring a unique expert-written caption for every video. 2) We construct a frozen-LLM semantic reward that combines out- come agreement, item relevance, unrelated-pair abstention, and gen- eration quality, and optimize the descriptor with GRPO and rank-16 LoRA for parameter-efficient adaptation. 3) We evaluate the framework on 13,379 videos from 4,670 children, covering 41 developmental behaviors and two VLM generations. Accuracy and item-level recall coverage improve on both backbones. The remaining errors involve temporal direction, duration, and age- specific semantics. arXiv:2609.33082v1 [cs.CV] 27 Sep 2026 Overview of the proposed VLM–LLM collaborative training framework for autism behavior assessment Training: Reward-guided VLM Optimization Inference: Autism Behavior Assessment FrozenTrainable Autism Screening Data Short screening videos x Screening items / criteria c VLM Behavior Description Base VLM (frozen) + LoRA (trainable) Generate natural language behavior descriptions dᵢ LLM-as-Judge Reward Construction LLM (Frozen) Outcome Reward Relevance Reward Unrelated Reward Quality Reward GRPO + LoRA Optimization GRPO + LoRA Update LoRA with reward signals Improved VLM Base VLM (frozen) + LoRA (trained) Better behavior descriptions Higher assessment performance GRPO objective + LoRA update Video & Criterion Video x Screening item / criterion c Trained VLM Trained VLM Behavior description for the given video d Semantic Decision Model Assessment Model Autism screening item prediction Assessment Result Pass (low risk) Fail (high risk) ? Unable to assess Fig. 1. Overview of SemReward-VL. The VLM generates item-conditioned descriptions, a frozen LLM supplies semantic rewards, and GRPO updates rank-16 LoRA adapters; inference reuses the adapted descriptor before the final item decision. 2. METHOD 2.1. Problem formulation and two-stage inference Letxdenote a child video,can age-specific assessment item and its pass criterion, andy ∈pass, failthe clinician-reviewed outcome. The VLM descriptor π θ samples a behavior description d i ∼ π θ (·| x,c), i = 1,...,G.(1) A frozen semantic decision modelhmaps(d i ,c)to pass, fail, or unable-to-assess. We use explicit item criteria in the judge prompt, following established LLM-as-a-Judge practice [30]. Inference first produces a behavior description and then applies the semantic deci- sion model, separating visual evidence from item interpretation. For example, a criterion may ask whether a child pours an object out of a bottle. A useful description records the bottle, the child’s manipulation of it, and whether an object visibly leaves the container. Details such as clothing or the bottle’s decoration do not determine the outcome. The decision model compares the reported action with the criterion. If the relevant part of the action is outside the clip, it can return unable-to-assess instead of treating the missing observation as a failed attempt. Several descriptions can accurately represent the same clip. A failed item also has no unique visual form: the target action may be absent, incomplete, replaced by another action, or not elicited. We optimize descriptions with outcome-level rewards. Each reward mea- sures whether a description supports the clinician-reviewed outcome without prescribing its wording. 2.2. Item-conditioned prompting and abstention The descriptor receives the exact item and pass criterion. For “stand on one foot for 2 seconds,” the prompt asks about the support leg and duration. For “expresses no,” it asks about visible refusal behavior. If the evidence is absent, occluded, or outside the recorded interval, the descriptor briefly states what is visible and says that the item cannot be assessed. Missing evidence therefore remains distinct from observed failure. Item conditioning narrows the requested evidence without as- suming that the child will express it in a single fixed way. A refusal, for instance, may appear as turning away or pushing an object. The description should name the visible movement before the decision model relates it to the age-specific criterion. This separation also makes an incorrect mapping visible: the model may observe the motion correctly but draw the wrong item conclusion. Training also includes mismatched video–item pairs(x,c ′ ), wherec ′ names a behavior unrelated to the video. The descriptor earns a reward for abstaining on these pairs. This teaches it to distinguish missing evidence from evidence of failure. 2.3. LLM-as-a-Judge semantic reward For each sampled descriptiond i , a frozen Qwen3-8B judge evaluates four binary properties. Outcome reward checks both evidence suffi- ciency and whether the item decision implied byd i agrees with the clinician outcomey. A candidate receivesr out i = 1only when both conditions hold. Unrelated-pair reward is active on mismatched pairs and assignsr unrel i = 1only when the descriptor correctly abstains. Relevance rewardr rel i checks whether the text actually addresses the requested item, while quality rewardr qual i rejects garbled or internally contradictory generations. The total semantic reward is r i = r out i + r unrel i + r rel i + r qual i .(2) Outcome agreement alone produced irrelevant or malformed re- sponses, particularly under the strong pass-label imbalance. Matched records receive an outcome reward, and mismatched pairs receive an abstention reward. Relevance and quality are scored for both. The Table 1. Training configuration. SettingValue Group size G16 LoRA targetVLM language layers LoRA rank16 Optimizer / learning rateAdamW / 10 −5 Batch size / epochs8 / 2 Training iterations5 Train/test protocolper-item 90/10 four binary terms allow varied wording while keeping descriptions tied to the item. The judge remains frozen throughout training. 2.4. GRPO optimization For every input, the descriptor samplesG = 16candidate descrip- tions. Their rewards are standardized within the group, ˆ A i = r i − μ r σ r + δ ,(3) whereμ r andσ r are the group reward mean and standard deviation. The clipped objective is J (θ) = 1 G X i min ρ i ˆ A i , clip(ρ i , 1− ε, 1 + ε) ˆ A i − βD KL , (4) Hereρ i is the current-to-old policy ratio, and the KL term anchors the reference policy. GRPO uses the sampled group as its baseline without a separate critic [31]. Similar rewards across all candidates give the update little directional signal. We observe this pattern in repeated temporal-direction errors (Sec. 3). 2.5. Parameter-efficient adaptation Only the language-model component of the VLM is adapted with LoRA [32]. For a frozen weight matrixW, LoRA represents the update as W ′ = W + BA,rank(BA) = r ≪ min(d in ,d out ),(5) with rankr = 16. The visual backbone, pretrained VLM weights, and judge remain frozen. Training uses AdamW (10 −5 ), batch size 8, two epochs, five iterations, and LLM-diversified item prompts. Preliminary full-parameter tuning was unstable on these data. Rank- 16 LoRA trained stably and is used in all reported models. 3. EXPERIMENTS 3.1. Dataset and evaluation protocol The dataset contains 13,379 naturalistic developmental videos from 4,670 children, collected at child-health and developmental clinics after guardian consent. Ages range from 2 to 42 months, with a mean of 12.73 months. The cohort includes 2,313 boys and 2,357 girls. We evaluate 41 items at ages 3, 6, 8, 10, 12, 18, 24, and 30 months. Each example includes the item, filming guidance, pass criterion, clinician-reviewed outcome, and a video uploaded by a caregiver or clinician. The videos have no dense behavior descriptions. The 41 items cover gross-motor, fine-motor, social, and adap- tive behaviors. They include rolling from supine to prone, walking backward, standing on one foot, transferring blocks, giving an object Table 2. Dataset summary. StatisticValue Children4,670 Videos13,379 Age range / mean2–42 / 12.73 months Male / female2,313 / 2,357 Key developmental ages8 Evaluated behavior items41 Dense behavior captionsNot available on request, and expressing refusal. A shared description layer can represent the different actions and criteria across items. Class prevalence is markedly imbalanced. For the six-month “reach for an object” item, 369 of 370 videos have a pass label. The same motion can also carry different meanings across ages and item criteria. We report item-wise precision and recall alongside aggregate accuracy. Each item uses 90% of its videos for training and 10% for testing. All prompt and adaptation comparisons use this split protocol. 3.2. Evaluation protocol We evaluate descriptions with a frozen LLM-as-a-Judge. To assess relevance, it reads the description and item prompt and returns a binary decision. To assess behavior accuracy, it first compares the described evidence with the criterion, then returns a constrained yes/no decision. An unrelated video–item pair is correct when the system responds “unable to assess.” Item-wise recall and precision measure evidence recovery; aggregate accuracy measures agreement with clinician-reviewed outcomes. The evaluation uses the same item text and pass criterion across the base and adapted models. The judge’s first turn considers whether the description contains enough behavior evidence to apply the cri- terion. Its second turn provides a constrained answer that can be parsed consistently. This protocol measures whether the generated text supports an item decision; it does not require the text to match a particular sentence. The unrelated-pair test separately checks whether the model avoids asserting a behavior that the video does not show. 3.3. Prompt specificity before adaptation We first test general-purpose VLMs with broad prompts. The study includes InternLM-XComposer2.5 [15], Qwen2.5-VL-Instruct-7B [33], and Qwen3-VL-Instruct-8B. Each prompt requests a summary of one developmental domain: gross motor, fine motor, language, personal-social, or adaptive behavior. The resulting descriptions often include fluent but irrelevant observations and omit details needed for an individual item. For Qwen3-VL, 5 of 41 items reach recall above 0.5: 124 (0.513), 31 (0.520), 187 (0.570), 244 (0.667), and 36 (0.713). The item-specific prompt states the exact item and pass criterion and asks the model to abstain when evidence is absent. With this prompt, 30 of 41 items reach recall above 0.5. Seven exceed 0.9: 81 (0.944), 123 (0.944), 34 (0.950), 186 (0.950), 243 (0.950), 63 (1.000), and 187 (1.000). The descriptions are also shorter and more relevant. The assessment items differ in the detail needed for a decision. Sitting independently centers on posture, turning pages requires an or- dered hand action, and drinking from a cup requires object use. Broad domain prompts can mention these scenes without resolving the re- quired detail. Item-specific prompts make the relevant action explicit, (a) Qwen2.5-VL(b) Qwen3-VL Fig. 2. Item-level recall before and after adaptation. Bars denote the base VLM and lines the GRPO-LoRA model over 41 age–item codes; adapted Qwen3-VL reaches recall > 0.5 on 38/41 items. Table 3. Aggregate accuracy under the per-item 90/10 split. BackboneBase+GRPO-LoRA Abs. gain Qwen2.5-VL 0.6690.817+0.148 Qwen3-VL0.8080.890+0.082 Table 4. Behavior coverage counted by items with recall > 0.5. SettingItems with recall > 0.5 Generic domain prompt5 / 41 Item-specific prompt30 / 41 Item-specific prompt + SemReward-VL38 / 41 which explains the large coverage increase before any parameter up- date. The remaining low-recall items show where prompt specificity alone does not recover sufficient visual or temporal evidence. 3.4. GRPO-LoRA adaptation Table 3 shows gains on both backbones. Qwen2.5-VL accuracy rises from 0.669 to 0.817 (+22.12% relative), and Qwen3-VL rises from 0.808 to 0.890 (+10.15% relative). The adapted Qwen2.5-VL approaches the unadapted Qwen3-VL. Adapted Qwen3-VL reaches the highest accuracy. Prompting and adaptation each improve item coverage. The number of items with recall above 0.5 rises from 5/41 with broad prompts to 30/41 with item-specific prompts, then to 38/41 with SemReward-VL (Table 4). Item-conditioned prompts produced shorter and more relevant descriptions than broad prompts. With outcome reward alone, mis- matched pairs often produced irrelevant or malformed responses. Full-parameter tuning was also less stable than rank-16 LoRA. The final design combines item-specific prompting, four semantic reward terms, and LoRA adaptation. 3.5. Item-level behavior and failure modes Figure 2 shows item-level gains for both backbones. Qwen2.5-VL starts with lower coverage and has the larger absolute accuracy gain. Qwen3-VL starts higher and reaches the best final accuracy and widest coverage (Table 3). Three items remain below 0.5 recall: walking backward (184, 0.000), expressing “no” (105, 0.053), and standing on one foot for 2 s (301, 0.471). Walking backward and standing on one foot require temporal grounding. The model may recognize walking but confuse its direction, or miss how long a posture lasts. When all 16 sampled descriptions repeat an error, GRPO has little signal to favor a correction. Frame reversal, order perturbation, and duration examples may help address this pattern. Expressing refusal requires a different interpretation: a young child may turn away or push an object instead of shaking their head. The descriptions reveal three error types. The model may miss a temporal attribute, misinterpret a visible action under the item criterion, or lack the requested evidence. These cases call for tem- poral examples, criterion-specific examples, and abstention training, respectively. 3.6. Discussion and stratified analysis Item conditioning directs the model toward the action, direction, duration, object interaction, or response specified by the criterion. GRPO-LoRA improves how consistently it describes that evidence. Coverage rises from 5/41 items with broad prompts to 30/41 with item-specific prompts and 38/41 after adaptation. Qwen3-VL im- proves despite its stronger starting point. The description layer makes visual-grounding and semantic- decision errors easier to distinguish. It also retains “unable to assess” for missing evidence. Each error type points to a specific training need: temporally difficult clips, criterion-specific examples, or mis- matched pairs. The abstention test addresses a separate failure. Most clinical outcomes in this collection are passes, so a model can appear accurate by predicting that a requested behavior occurred. Pairing a video with an unrelated item tests whether the description is grounded in the clip. A correct response briefly reports the visible content and withholds the item decision. Sex-stratified precision is broadly similar. For items 105, 184, and 187, the test subsets have no valid female examples, so those gaps cannot be interpreted as performance differences. The per-item split measures performance within the collection distribution. Child- disjoint and cross-institution splits would test transfer to new children and recording settings. 4. CONCLUSION Routine developmental records provide clinician outcomes but rarely describe the behavior visible in each video. SemReward-VL ad- dresses this gap by separating behavior description from item deci- sion. Frozen-LLM rewards guide GRPO-LoRA adaptation without caption labels. On 13,379 videos covering 41 items, item-specific prompts raise the number of items with recall above 0.5 from 5 to 30, and adaptation raises it to 38. Accuracy also improves on both VLM backbones. The descriptions make decisions easier to inspect and expose remaining errors in temporal direction, duration, and age-specific interpretation. 5. REFERENCES [1]A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” in Proc. ICML, 2021, p. 8748– 8763. [2] C. Jia, Y. Yang, Y. Xia, Y.-T. Chen, Z. Parekh, H. Pham, Q. Le, Y.-H. Sung, Z. Li, and T. Duerig, “Scaling up visual and vision- language representation learning with noisy text supervision,” in Proc. ICML, 2021, p. 4904–4916. [3] J.-B. Alayrac, J. Donahue, P. Luc, et al., “Flamingo: A visual language model for few-shot learning,” in Advances in Neural Information Processing Systems, vol. 35, 2022. [4] J. Li, D. Li, S. Savarese, and S. Hoi, “BLIP-2: Bootstrap- ping language-image pre-training with frozen image encoders and large language models,” in Proc. ICML, 2023, p. 19730– 19742. [5]W. Dai, J. Li, D. Li, A. M. H. Tiong, J. Zhao, W. Wang, B. Li, P. Fung, and S. Hoi, “InstructBLIP: Towards general-purpose vision-language models with instruction tuning,” in Advances in Neural Information Processing Systems, vol. 36, 2023. [6]H. Liu, C. Li, Q. Wu, and Y. J. Lee, “Visual instruction tuning,” in Advances in Neural Information Processing Systems, vol. 36, 2023. [7]D. Zhu, J. Chen, X. Shen, X. Li, and M. Elhoseiny, “MiniGPT-4: Enhancing vision-language understanding with advanced large language models,” arXiv:2304.10592, 2023. [8]K. Li, Y. He, Y. Wang, Y. Li, W. Wang, P. Luo, Y. Wang, L. Wang, and Y. Qiao, “VideoChat: Chat-centric video understand- ing,” arXiv:2305.06355, 2023. [9]M. Maaz, H. Rasheed, S. Khan, and F. S. Khan, “Video- ChatGPT: Towards detailed video understanding via large vision and language models,” arXiv:2306.05424, 2023. [10]H. Zhang, X. Li, and L. Bing, “Video-LLaMA: An instruction- tuned audio-visual language model for video understanding,” arXiv:2306.02858, 2023. [11]Y. Li, C. Wang, and J. Jia, “LLaMA-VID: An image is worth 2 tokens in large language models,” arXiv:2311.17043, 2023. [12]B. Lin, Y. Ye, B. Zhu, J. Cui, M. Ning, P. Jin, and L. Yuan, “Video-LLaVA: Learning united visual representation by align- ment before projection,” arXiv:2311.10122, 2023. [13]Z. Cheng et al., “VideoLLaMA 2:Advancing spatial- temporal modeling and audio understanding in video-LLMs,” arXiv:2406.07476, 2024. [14]P. Wang et al., “Qwen2-VL: Enhancing vision-language model’s perception of the world at any resolution,” arXiv:2409.12191, 2024. [15]P. Zhang et al., “InternLM-XComposer-2.5: A versatile large vision language model supporting long-contextual input and output,” arXiv:2407.03320, 2024. [16]B. Li et al., “LLaVA-OneVision: Easy visual task transfer,” arXiv:2408.03326, 2024. [17]K. Li et al., “MVBench: A comprehensive multi-modal video understanding benchmark,” in Proc. CVPR, 2024, p. 22195– 22206. [18]E. Song et al., “MovieChat: From dense token to sparse memory for long video understanding,” in Proc. CVPR, 2024, p. 18221– 18232. [19] B. He et al., “MA-LMM: Memory-augmented large multimodal model for long-term video understanding,” in Proc. CVPR, 2024, p. 13504–13514. [20]J. Chen et al., “VideoLLM-online: Online video large language model for streaming video,” in Proc. CVPR, 2024, p. 18407– 18418. [21]Y. Zhao et al., “Distilling vision-language models on millions of videos,” in Proc. CVPR, 2024, p. 13106–13116. [22]S. Ren, L. Yao, S. Li, X. Sun, and L. Hou, “TimeChat: A time-sensitive multimodal large language model for long video understanding,” in Proc. CVPR, 2024. [23]Y. Liu et al., “TempCompass: Do video LLMs really understand videos?” arXiv:2403.00476, 2024. [24] C. Li et al., “LLaVA-Med: Training a large language-and-vision assistant for biomedicine in one day,” arXiv:2306.00890, 2023. [25]M. Moor et al., “Med-Flamingo: A multimodal medical few- shot learner,” arXiv:2307.15189, 2023. [26]S. Zhang et al., “BiomedCLIP: A multimodal biomedical foun- dation model pretrained from fifteen million scientific image- text pairs,” arXiv:2303.00915, 2023. [27] C.-H. Yoo, J.-H. Yoo, and J. Jang, “CARE-VL: A domain- specialized vision-language model for early ASD screening,” in MICCAI, 2025. [28]K. Feng et al., “Video-R1: Reinforcing video reasoning in MLLMs,” in Advances in Neural Information Processing Sys- tems, 2025. [29] J. Zhang et al., “R1-VL: Learning to reason with multimodal large language models via step-wise group relative policy opti- mization,” in Proc. ICCV, 2025. [30]L. Zheng et al., “Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena,” arXiv:2306.05685, 2023. [31]Z. Shao et al., “DeepSeekMath: Pushing the limits of mathe- matical reasoning in open language models,” arXiv:2402.03300, 2024. [32] E. J. Hu et al., “LoRA: Low-rank adaptation of large language models,” in ICLR, 2022. [33]QwenTeam,“Qwen2.5-VLTechnicalReport,” arXiv:2502.13923, 2025.