acceptodds
Under review as a conference paper at ICLR 2027

Perspective-Conditioned Multimodal Reasoning for E-Commerce Disputes Verdict

Abstract

The growing volume of E-commerce disputes creates an increasing demand for automated verdict systems. Existing multi-agent methods introduce diverse role prompts and interaction contexts, aiming to elicit complementary reasoning and improve verdict accuracy. However, their effectiveness remains limited by the inherent biases of underlying models. To address this, we propose *PrismJudge*, a Perspective-Conditioned Multimodal Reasoning Framework for E-Commerce Dispute Verdict (EDV). Inspired by diverse human perspectives, *PrismJudge* constructs perspective branches, each associated with a learnable latent perspective condition. These conditions jointly guide evidence grounding and verdict reasoning to capture diverse responsibility attribution patterns. Specifically, *Perspective-Conditioned Visual Cue Grounding* (PCVCG) exploits perspective-conditioned visual scores to selectively ground informative visual tokens, enabling the discovery of pivotal cues from redundant inputs. Built upon this, *Perspective-Conditioned Verdict Rationale Generation* (PCVRG) injects perspective conditions into the verdict reasoning process through case-adaptive perspective states and perspective-conditioned LoRA, enabling branches to share common verdict knowledge while learning distinct reasoning viewpoints for the final verdicts. Subsequently, *Rationale-Aligned Joint Learning* (RAJL) aligns human rationales with perspective branches and jointly optimizes generation, rationale matching, and diversity objectives to encourage diverse cue interpretation and complementary verdict generation. Extensive experiments show that *PrismJudge* improves accuracy by and reduces the buyer-seller accuracy gap by compared with the strongest baseline, highlighting the effectiveness of perspective-conditioned reasoning for accurate and balanced EDV.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
E-commmer Dispute Verdicts, Multimodal large language models, Multi-Perspective

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ahVpOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PrismJudge, a perspective-conditioned multimodal reasoning framework for automated e-commerce dispute resolution. It introduces multiple learnable perspective branches that guide visual evidence selection and verdict rationale generation, aiming to overcome the limited diversity of role-prompted multi-agent systems. The framework jointly optimizes generation, alignment with human rationales, and perspective diversity to encourage complementary interpretations of evidence.

Critical strengths
  • This paper studies a very practical issue with clear motivation, and it is easy to follow.

  • Good representation: Figures in the main paper and Alg.1 in Appendix A clearly show the main idea.

  • The insight “role diversity alone does not guarantee complementary decision perspectives” is interesting.

Critical weaknesses
  • Despite the interesting use case, the paper can be categorized as a LoRA-based method. The novelty and technical contribution should be clearly highlighted.

  • The ablation study of using different backbones is not reported. For instance, will scaling up the MLLMs to 30B benefit the performance? I’m also interested in other backbones, e.g., the InternVL series.

Questions for authors
  • Is there any user study of applying your approach to a practical website?

  • What’s the efficiency comparison and analysis? In E-commerce transactions, latency plays a significant role.

Reviewer_WQYaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PrismJudge, a multimodal reasoning framework based on latent perspective conditions, to address redundant multimodal evidence and verdict bias in e-commerce dispute resolution. By learning multiple perspective conditions, the framework guides the model to extract evidence from different viewpoints and generate corresponding verdict rationales. Specifically, Perspective-Conditioned Visual Cue Grounding (PCVCG) uses perspective conditions to select dispute-relevant visual cues. Perspective-Conditioned Verdict Rationale Generation (PCVRG) incorporates different perspectives into verdict reasoning through case-adaptive perspective states and conditioned LoRA. Rationale-Aligned Joint Learning (RAJL) uses human verdict rationales to match and jointly train different branches, encouraging complementary reasoning patterns. Experiments on VerdictBench show that PrismJudge, using Qwen3-VL-4B as the backbone, achieves an accuracy of 81.50%, outperforming CyberJurors by 8.58 percentage points while reducing the buyer–seller accuracy gap from 22.98% to 15.74%. The authors further validate the effectiveness of the method through component ablations, multi-branch analyses, and cross-backbone experiments.

Critical strengths
  1. Section 3 connects visual evidence selection with verdict reasoning through shared latent perspective conditions. PCVCG jointly uses text relevance and perspective conditions to compute visual feature scores, while PCVRG further adjusts the subsequent verdict generation process through case-adaptive states and conditioned LoRA. This design considers the characteristics of e-commerce disputes, where evidence can be extensive and different pieces of evidence may admit different interpretations. It also provides a concrete implementation of multi-perspective reasoning beyond fixed role prompting.

  2. Table 2 provides relatively comprehensive component ablations, showing that model accuracy progressively improves from 73.82% with single-branch supervised fine-tuning to 81.50% with the complete method, with PCVCG, PCVRG, and multi-branch training each contributing to the improvement. The complete method also outperforms the HydraLoRA baseline (81.50% vs. 77.21%). Furthermore, the experiments with Gemma-3-4B-IT in the appendix show that PrismJudge improves accuracy from 75.66% to 78.50%, while the accuracy standard deviation across three random seeds is only 0.12 percentage points. These results provide support for the effectiveness of the method and its applicability across different backbones to some extent.

Critical weaknesses
  1. Section 4.1 highlights the 8.58-percentage-point accuracy improvement of PrismJudge over CyberJurors as a major experimental result, but the two methods differ substantially in their learning approaches and supervision information. PrismJudge is supervised using ground-truth verdict labels and human verdict rationales, whereas the multi-agent baselines in Table 1 mainly rely on role prompting and model interactions to make decisions. In fact, the single-branch Opinion LoRA SFT model in Table 2 already achieves an accuracy of 73.82%, exceeding the 72.92% of CyberJurors. When compared with multi-branch training baselines, PrismJudge improves over HydraLoRA and Shared A, B by 4.29 and 3.84 percentage points, respectively. Therefore, the overall performance advantage reported in Table 1 cannot be directly attributed to the proposed multi-perspective mechanism. Although the ablation experiments provide some supporting evidence, the lack of sufficiently controlled comparisons with conventional multi-branch fine-tuning methods under matched supervision data, training budgets, and model capacity makes it difficult to assess the actual incremental contribution of the proposed method.

  2. The Rationale Alignment mechanism in Section 3.3 only uses human rationales consistent with the ground-truth verdict label for branch matching, and constructs supervision sequences with the same ground-truth verdict label for all branches through Eq. (7). This training strategy can encourage different branches to learn different explanations supporting the same verdict, but it may exclude human perspectives that disagree with the final decision. Given that a central motivation of the paper is to interpret evidence from different perspectives and mitigate the inherent verdict bias of models, it remains questionable whether this prior filtering strategy can adequately preserve genuinely conflicting perspectives on responsibility attribution in human verdicts. Moreover, the current experiments do not directly verify whether the perspectives learned by different branches correspond to distinct principles of evidence interpretation. Consequently, it is difficult to determine whether the model learns substantively different verdict perspectives or merely diverse explanatory expressions under the same supervision label.

  3. Section 3.1 selects key visual cues using the similarity between textual and visual features together with perspective-conditioned residual scores, but there is a lack of direct validation that the selected features actually correspond to reliable evidence affecting the verdict. In e-commerce disputes, descriptions provided by buyers and sellers may contradict each other or even contain false statements. Therefore, visual regions highly relevant to one party's textual description are not necessarily reliable evidence supporting the final verdict. Table 2 shows that introducing PCVCG improves accuracy from 75.56% to 77.06%, but this result only demonstrates that the corresponding feature selection mechanism benefits final prediction. Appendix A.5 reports an accuracy decrease after randomly removing 25% of the selected visual tokens, but does not compare against random selection, conventional attention-based selection, or other selection strategies with matched sparsity. It also does not evaluate the agreement between selected regions and the actual decisive evidence. Therefore, the current results do not sufficiently support the claim that PCVCG can more accurately locate decisive visual cues.

  4. Section 3.2 adopts an autoregressive generation order in which the verdict is generated before the rationale, while the main evaluation metrics in Section 4 almost exclusively focus on the final verdict labels, without assessing the correctness or evidential faithfulness of the generated rationales. Although rationale supervision can influence model representations during training, the verdict has already been produced before the rationale is generated at inference time. Therefore, the generated explanation may simply be a post-hoc rationalization of an existing decision. The paper does not further verify whether the evidence cited in generated rationales actually exists, whether responsibility attribution is consistent with the original case materials, or whether the rationales accurately reflect the visual cues used by the model to reach its verdict. For a method that treats Verdict Rationale Generation as a core component, classification accuracy and rationale theme distributions alone are insufficient to demonstrate that the model has acquired more reliable evidence reasoning capabilities.

  5. Equation (4) in Section 3.2 constructs the case-adaptive perspective state by averaging representations across all evidence items, while the visual cue scoring in Section 3.1 is also primarily performed independently within each evidence item. However, critical judgments in e-commerce disputes often depend on relationships between different pieces of evidence, such as whether the seller's pre-transaction description matches the product condition reported by the buyer after receipt, or whether evidence submitted at different times contains contradictions regarding the same defect. Although the subsequent input sequence preserves the original evidence order, the aggregated representation used to modulate branch states and LoRA parameters does not explicitly model cross-evidence ordering or relationships. The paper does not analyze whether this design affects complex disputes requiring comparisons across multiple rounds of evidence, nor does it provide experiments specifically examining evidence order, cross-evidence contradictions, or temporal dependencies. Given that multi-round evidence submission is an important scenario emphasized in the introduction, the method design and experimental validation in this respect remain insufficient.

  6. The diversity objective in Section 3.3 encourages differences among branches by constraining the similarity of perspective representations, visual attention, and case states, but there is insufficient evidence that these differences contribute to more accurate responsibility attribution. Not every dispute requires attention to different visual regions. When multiple pieces of evidence point to the same critical fact, forcing branches to maintain differences may even introduce irrelevant cues. In Table 2, removing the Diversity Loss decreases accuracy from 81.50% to 78.45%, indicating that this loss contributes to overall prediction performance. However, WGA slightly increases from 71.62% to 72.16%, suggesting that its contribution to improving the worst-performing group remains unclear. Furthermore, Appendix A.5 reports verdict disagreement rates of only 5.44%–5.79% among different branches. These results support the existence of differences in branch representations, but do not establish that such differences lead to genuinely complementary reasoning paths or constitute the primary reason for reducing buyer–seller verdict bias.

  7. Section 4.1 primarily measures verdict bias using Buyer-win Accuracy, Seller-win Accuracy, and their Accuracy Gap, and concludes from the reduced gap that multi-perspective reasoning can mitigate the inherent bias of foundation models. However, differences in accuracy between the two categories may also be influenced by case complexity, evidence quality, and category distribution, and cannot be directly equated with systematic verdict bias. In Table 1, some LLMs and MLLMs achieve higher accuracy on Buyer-win cases, whereas AgentCourt, TED, and CyberJurors all show higher Seller-win Accuracy, indicating that the direction of the performance imbalance varies substantially across methods. Although PrismJudge improves accuracy for both categories and reduces the gap to 15.74%, this result is still insufficient to explain how the proposed perspective mechanism changes the model's verdict tendencies. In particular, the paper does not further examine buyer-win and seller-win performance across different product categories and difficulty levels, nor does it analyze the direct relationship between perspective conditions and changes in verdict bias. Therefore, the claims regarding mitigation of inherent model bias go beyond what the current experimental evidence can fully support.

Questions for authors

Please refer to Critical Weaknesses.

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

This paper introduces PrismJudge, a perspective-conditioned framework for predicting verdicts and generating rationales from multimodal evidence in e-commerce disputes. Within each branch, the same latent perspective condition guides visual cue grounding (PCVCG) and verdict and rationale generation (PCVRG). Rationale-Aligned Joint Learning (RAJL) aligns human rationales with perspective branches and jointly optimizes generation, semantic matching, and diversity objectives. Experiments on VerdictBench show that PrismJudge outperforms the compared methods on most reported metrics.

Critical strengths

1 Coherent perspective-conditioned design. PrismJudge uses the same latent perspective condition within each branch to connect visual cue grounding with verdict reasoning. Incorporating this condition into both evidence selection and model adaptation provides a coherent mechanism for learning different ways of interpreting the same dispute evidence. 2 Rationale-aligned multi-branch learning. RAJL uses human rationales to supervise pre-generation branch representations and assign distinct explanation targets to different branches. Combining this alignment with generation and diversity objectives provides a concrete mechanism for encouraging complementary perspectives. The pre-verdict readout and separate matching pass also guard against target-rationale leakage into the representations used for matching. 3 Empirical support for the framework. Comparisons with supervised baselines and component ablations provide evidence for the benefits of perspective-conditioned grounding, adaptation, and rationale alignment. Additional backbone experiments and repeated runs broaden the evaluation. The strong performance of a validation-selected individual branch also suggests practical value in combining multi-branch training with single-branch inference.

Critical weaknesses
  1. Supervision, rationale assignment, and training budgets for key baselines (L2, L12, and L13) are insufficiently specified, limiting attribution of the reported gains.
  2. RAJL lacks comparison with direct one-to-one assignment using cosine-distance costs, leaving the added value of Sinkhorn alignment and representative-rationale selection unclear.
  3. Branch diversity and keyword-based analyses provide limited evidence of semantic specialization, especially when diversity is explicitly encouraged during training. Rationale factuality and faithfulness remain insufficiently validated.
  4. Despite improved overall accuracy and class balance, buyer-win accuracy drops from 86.08% for the zero-shot backbone to 71.62%. This shift is insufficiently explained. Class-wise results and analysis across the supervised baseline and key ablations would help distinguish the effects of general fine-tuning from those of the proposed components.
Questions for authors

1 Limited human rationale availability. Appendix A.2 states that each case has at least K eligible rationales. Is this condition naturally satisfied, or does it require excluding cases? Please clarify whether the K = 2, 3, and 5 experiments use the same training cases, and how RAJL would handle cases with fewer eligible rationales. 2 Misleading or incomplete evidence. Since both PCVCG scoring terms depend on text, how robust is PrismJudge when textual claims conflict with images, omit decisive visual details, or accompany irrelevant screenshots? Could the authors test these conditions against L2 and L13 using controlled perturbations verified to preserve the correct verdict? 3 Unseen product categories. Could the authors provide a leave-one-category-out evaluation against supervised baselines using the same backbone, excluding the held-out category from task-specific training and model selection? Figure 3 reports category-wise performance, while Appendix A.6 evaluates different backbones; neither directly tests transfer to categories unseen during task-specific training.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.