acceptodds
Under review as a conference paper at ICLR 2027

CrossLJP: An Asymmetric Cross-View Framework for Legal Judgment Prediction

Abstract

Legal judgment prediction (LJP) uses case facts to predict applicable provisions or judicial outcomes. Prior work has improved LJP through legal pretraining, statutory knowledge, precedents, and label dependencies, yet prediction often relies on a single task-adapted representation. This representation must capture case context and distinguish applicable provisions, but gives the predictor no separate access to pretrained features unchanged by legal supervision. We introduce CrossLJP, an asymmetric cross-view framework for multi-label legal provision prediction that preserves a frozen semantic view alongside an adapted decision view. Cross-attention uses the semantic view to query the decision view, allowing pretrained case features to guide the weighting of adapted features. Across three benchmarks derived from real-world court cases in Chinese, Korean, and English, CrossLJP achieves the highest mean Exact Match (EM), Micro-F1, and Macro-F1 among the evaluated models. These results support preserving pretrained case features and using them to guide prediction from task-adapted features.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
legal judgment prediction, legal natural language processing, multi-label classification, multi-view representation learning, cross-attention

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_reodOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CrossLJP, a framework that leverages cross-attention to construct an additional interaction context, which is combined with the original representations to produce more informative case representations for legal judgment prediction. The authors conduct extensive experiments, including comparisons with various baselines, ablation studies, and evaluations across datasets in different languages, to demonstrate the effectiveness of the proposed approach. However, the methodological novelty appears limited, as CrossLJP mainly integrates existing techniques, including frozen pretrained encoders, task-adapted encoders, and cross-attention mechanisms. Furthermore, the rationale for selecting different backbone encoders for the semantic and decision representations remains unclear. The paper would also benefit from a stronger justification for using relatively older embedding models, such as Nomic and BGE-M3, rather than more recent alternatives, as well as additional experiments comparing different embedding models to validate the encoder selection.

Critical strengths
  1. The proposed CrossLJP module utilised cross-attention to construct an additional interaction context. As a result, the interaction context is retained together with original representation to form an informative case representation.
  2. The authors have done extensive experiments for baselines comparison, ablation study, different language datasets and so on.
Critical weaknesses
  1. The novelty of the proposed CrossLJP module appears limited, as it mainly combines three existing techniques: frozen pretrained encoders, task-adapted encoders, and cross-attention mechanisms for representation aggregation. The authors should further clarify the methodological contributions beyond the integration of these existing techniques.
  2. The authors should clarify the rationale for using different backbone encoders for the semantic encoder and decision encoder, rather than adopting the same backbone for both components.
  3. The proposed CrossLJP module relies on embedding models such as Nomic and BGE-M3, both of which were introduced approximately two years ago. More recent embedding models, such as Qwen3-Embedding, have demonstrated strong performance on legal retrieval tasks in the MTEB benchmark. The authors should justify their choice of these relatively older embedding models. Additional experiments comparing different embedding models as backbone encoders are recommended to further investigate the impact of encoder selection on the performance of CrossLJP.
Questions for authors

Why the semantic encoder and decision decoder using different backbones? Why not use more recently and more powerful encoders?

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

This paper introduces CrossLJP, an asymmetric cross-view framework for multi-label legal provision prediction. It addresses the tension between preserving pretrained case semantics and adapting representations for provision prediction. CrossLJP combines a frozen semantic encoder with a LoRA-adapted decision encoder and uses cross-attention, with the semantic view providing queries and the decision view providing keys and values, to guide the aggregation of adapted features. Experiments on CAIL2018, LBOX, and ECtHR-A show improvements over single-encoder baselines and direct concatenation. Ablation studies further examine the effects of integration strategies and attention directions.

Critical strengths
  1. The paper clearly formulates the representation-learning tension between preserving pretrained case semantics and specializing for provision prediction. It focuses on multi-label legal provision prediction, where models must identify provisions supported by case facts while avoiding missing and unsupported predictions.
  2. CrossLJP retains both frozen and task-adapted representations and uses cross-attention to connect them before prediction. The paper also compares cross-view interaction with direct concatenation of the same two encoder views, providing a controlled evaluation of the proposed integration strategy.
  3. The paper evaluates CrossLJP on three legal benchmarks spanning Chinese, Korean, and English. The reported results show improvements over BGE-LoRA and Dual-View Concat, while ablation studies examine the effects of integration strategy and attention direction.
Critical weaknesses
  1. The motivation is not sufficiently supported by the experiments. Appendix A.4 analyzes only one selected CAIL2018 case. Moreover, it compares a frozen Nomic encoder with a LoRA-adapted BGE-M3 encoder, making it difficult to attribute the observed differences to fine-tuning. The analysis does not establish that the frozen view consistently provides complementary information.
  2. The evaluation relies on relatively old datasets. The paper evaluates CrossLJP on CAIL2018, LBOX, and ECtHR-A, without testing on newer LJP datasets, including subsequent CAIL-related benchmarks, or more recent legal cases. Its generalization to newer data remains unclear.
  3. The methodological novelty is limited. CrossLJP combines existing encoders, LoRA adaptation, and cross-attention. Its main contribution lies in the asymmetric integration of the two views, while the overall methodological novelty is limited.
  4. The evaluation lacks controlled comparisons with open-source LLMs and legal-domain LLMs. Although Table 1 reports results from prior work, these results use their original evaluation settings. Direct comparisons under a unified protocol would provide a clearer assessment of CrossLJP against these models.
Questions for authors
  1. In Appendix A.4, can the authors provide a controlled comparison using the same encoder before and after fine-tuning, together with perturbation analysis across multiple test cases? This would help determine whether the observed differences are attributable to fine-tuning and whether the two views consistently provide complementary information.
  2. Can the authors provide comparisons with representative general-purpose LLMs, such as GPT, Claude, DeepSeek, and Qwen, as well as legal-domain LLMs, under the same evaluation protocol? This would provide a clearer assessment of CrossLJP against these models under comparable settings.
Reviewer_FdBLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

CrossLJP addresses multilabel legal provision prediction from case facts. Its motivating observation is that adapting one pretrained encoder to provision labels may reduce access to broader semantic features that remain useful for distinguishing cases. The model consequently maintains two representations of the same document: a frozen semantic encoder and a decision encoder adapted through LoRA. In the principal configuration these are Nomic Embed and BGE-M3, respectively. Both encode document chunks, and asymmetric cross-attention uses frozen-view queries with adapted-view keys and values. The classifier receives independently pooled summaries from both encoders together with the pooled interaction context. Training uses weighted binary cross-entropy, and provision-specific decision thresholds are selected on validation predictions.

Critical strengths

[1]. The study includes a particularly relevant baseline: direct concatenation of exactly the frozen and adapted encoder views used by CrossLJP. This comparison is more informative than a single-encoder comparison alone because it tests whether the interaction module adds value after the classifier already has access to both representations. The CAIL2018 improvements in exact match and Micro-F1 are substantial enough to make this a credible architectural finding. Reporting exact match alongside aggregate and per-label F1 is appropriate for provision prediction, where an otherwise strong prediction can still contain an unsupported article or omit a necessary one. [2]. The ablations also address several plausible alternative explanations. A frozen copy of BGE-M3 remains beneficial alongside its adapted counterpart, indicating that the effect is not entirely attributable to combining unrelated encoder families. The reverse and bidirectional attention variants retain the same context width, making the attention-direction comparison reasonably controlled. Encoder substitution supplies additional evidence that the framework is not confined to the original Nomic/BGE-M3 pairing, although those experiments use one seed.

Critical weaknesses

[1]. The evidence is strongest for the usefulness of a dual-view architecture, but weaker for the proposed explanation in terms of preserving semantics. Both representations contain factual and label-relevant information, and the comparison in which both encoders are adapted trails CrossLJP by only 0.35 Micro-F1 points on CAIL2018, against reported deviations of roughly 0.6–0.8 points. The illustrative perturbation case was selected because CrossLJP succeeded where the decision-only model failed. Representation sensitivity in that selected example does not establish that fine-tuning systematically loses legally relevant semantics or that semantic preservation explains the aggregate gains. [2]. The benefits also vary considerably across datasets. On ECtHR-A, CrossLJP obtains 74.93 Micro-F1 versus 74.91 for BGE-LoRA; the improvement over concatenation partly recovers the latter's decline. LBOX's gain over concatenation is 0.39 points, while its test set contains only 276 cases. Highest mean performance should therefore be distinguished from a reliable advantage over the strongest simpler model. Paired uncertainty estimates would be useful. [3]. Architectural attribution remains incomplete. CrossLJP adds projections and a context pathway beyond concatenation; a parameter-matched nonlinear or gated pooling baseline is absent. Because the final context averages query-conditioned outputs, it also warrants comparison with simpler learned chunk attention. Two full encoders increase inference cost, yet latency and memory are not reported. Finally, the label-specific threshold search is potentially unstable for LBOX's 169 labels and small validation set, and the claimed multilingual breadth would benefit from documenting how the frozen encoder's language coverage affects each benchmark.

Questions for authors
  1. Can you quantify semantic preservation over a representative, predefined sample rather than a case selected after observing the models' predictions? For example, how do the two views respond to legally decisive fact changes, irrelevant contextual changes, and paraphrases, and do those response differences predict which cases CrossLJP improves?
  2. Are the improvements over the strongest single encoder statistically reliable on LBOX and ECtHR-A? Please provide paired per-seed differences or confidence intervals, particularly for the ECtHR-A Micro-F1 difference of 0.02 points. Which labels and case types account for the exact-match and Macro-F1 gains on these datasets?
  3. How does the model compare with parameter-matched gated fusion, an expanded concatenation classifier, and decision-only learned chunk attention? Since Equation 11 reduces the context to one weighted summary of decision values, these controls would clarify whether frozen-view conditioning contributes beyond additional pooling capacity and a larger prediction pathway.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.