Technical Summary: LegalOne

Problem Statement

While Large Language Models (LLMs) demonstrate impressive general capabilities, their direct application in the legal domain is hindered by two primary limitations:

1. Insufficient Domain Knowledge: General models lack precise grounding in authoritative sources (statutes, judicial interpretations, precedents) and struggle with the complexity of rigorous multi-step judicial reasoning.
2. Misaligned Reasoning Patterns: When applied to legal tasks, general LLMs often produce fluent but legally invalid outputs. Their reasoning patterns frequently fail to align with real-world legal practice, leading to hallucinations and a lack of source consistency.

Existing legal-specific models often lag behind leading general-purpose models in overall capability or fail to exhibit the deep reasoning required for complex statutory language and legal logic.

Methodology

The authors present LegalOne, a family of foundation models tailored for the Chinese legal domain. The models are developed through a comprehensive three-phase training pipeline designed to master legal reasoning while retaining general capabilities.

Phase 1: Mid-Training with Plasticity-Adjusted Sampling (PAS)

To bridge the gap between general pre-training and domain-specific instruction tuning, LegalOne introduces a mid-training stage.

• Data Strategy: The corpus combines high-quality general data (e.g., FinWeb-Edu, MathPile) to prevent catastrophic forgetting with a curated legal corpus (judgments, statutes, academic papers) and synthetic data. Synthetic data includes reasoning trajectories generated via the Feynman Technique and rephrased texts to enhance linguistic diversity.
• Plasticity-Adjusted Sampling (PAS): To address the challenge of learning new legal knowledge without forgetting original capabilities, the authors propose a perplexity-based data scheduler.
  • Mechanism: PAS treats sampling probability as a function of the learning rate. During the "re-warmup" phase (where the learning rate increases), the model is biased toward low-perplexity (low-PPL) "anchor" samples that align with its original distribution. This stabilizes optimization and prevents catastrophic forgetting. As the learning rate peaks, the distribution smoothly transitions to include high-PPL legal data.
  • Theoretical Basis: The method is grounded in numerical stability analysis, showing that low-PPL data acts as a "numerical damper" against the instability caused by increasing learning rates on out-of-distribution data.

Phase 2: Supervised Fine-Tuning (SFT) with Legal Agentic CoT Distillation (LEAD)

Standard distillation from general-purpose LLMs often propagates errors and lacks the rigorous structure of legal practice. LegalOne employs LEAD to distill explicit reasoning from raw legal texts.

• Agentic Workflow: Instead of naive distillation, LEAD utilizes an agentic system that emulates professional legal workflows. It decomposes complex legal problems into a structured sequence: Fact Finding, Issue Identification, Rule Retrieval, Rule Deduction, and Conclusion Derivation.
• Cognitive Scaffolding: The workflow integrates Expert Protocols, an external Legal Knowledge Base (for statute retrieval), and Theoretical Templates (e.g., judicial syllogism) to guide the generation of reasoning trajectories.
• Refinement: The raw traces undergo Knowledge Internalization (rewriting to remove explicit external references, forcing the model to internalize the logic) and Reasoning Convergence (merging fragmented sub-steps into a single, coherent Chain-of-Thought).
• Quality Control: A two-stage filtering process (heuristic and LLM-as-Judge) ensures the final dataset contains only high-quality, logically consistent, and factually grounded reasoning samples.

Phase 3: Multi-Stage Reinforcement Learning (RL)

To foster an autonomous "Legal Mentality," the authors implement a Curriculum Reinforcement Learning strategy using the DAPO algorithm with a novel Token-Level Baseline.

• Curriculum Design: The RL process spans five stages of increasing difficulty, targeting three core skills:
  1. Statute Memorization: Reciting exact statutory text (Reward: ROUGE).
  2. Statute Application: Mapping facts to applicable provisions (Reward: F1 score).
  3. Case Analysis: Answering complex multiple-choice questions (Reward: Accuracy).
  4. Judgment Prediction: Predicting outcomes based on case facts (Reward: LLM-as-Judge with tolerance for sentencing variance).
  5. Legal Document Generation: Drafting indictments or judgments (Reward: Rubric-based evaluation).
• Token-Level Baseline: To reduce variance in policy gradient estimation, the authors propose a token-level baseline ($b_2$) that weights rewards by sequence length, correcting the bias introduced by standard sample-level baselines ($b_1$) in long-sequence legal tasks.

Key Contributions

1. LegalOne Model Family: A series of models (1.7B, 4B, 8B) that achieve state-of-the-art performance in Chinese legal reasoning, outperforming significantly larger general-purpose models (e.g., Qwen3-235B, GPT-5) in specific legal tasks.
2. Plasticity-Adjusted Sampling (PAS): A novel data scheduling mechanism that effectively balances knowledge acquisition and retention during mid-training, mitigating catastrophic forgetting without requiring access to original pre-training data.
3. Legal Agentic CoT Distillation (LEAD): A framework for converting implicit judicial logic into explicit, structured reasoning trajectories using agentic workflows, resulting in higher factual grounding and logical rigor than naive distillation.
4. Curriculum RL with Token-Level Baseline: A multi-stage RL strategy that evolves the model from pattern matching to autonomous reasoning, coupled with a variance-reducing baseline optimization technique.
5. Open Resources: Public release of the LegalOne model weights and LegalKit, a comprehensive evaluation toolkit for legal-domain LLMs.

Experimental Results

The models were evaluated on JEC-QA (National Unified Legal Professional Qualification Examination) and LexEval (Legal Cognitive Ability Taxonomy).

• Performance: LegalOne-8B achieves superior performance across legal tasks, surpassing general-purpose models with vastly larger parameter counts (e.g., DeepSeek-v3.2, Qwen3-Max). For instance, LegalOne-8B scores 62.36 on LexEval Average and 62.36 on JEC-QA Average, outperforming Qwen3-8B (44.61 / 44.61) and competing with much larger models.
• Parameter Efficiency: LegalOne demonstrates high "capability density," where the 1.7B model significantly outperforms the base Qwen3-1.7B and approaches the performance of the 8B general model, proving that curated domain data can bridge the gap for compact models.
• Ablation Studies:
  • Mid-Training: Essential for smaller models (1.7B) to prevent performance degradation during SFT; it acts as a stabilizer for the loss landscape.
  • RL: Provides significant gains, particularly for larger models (8B), suggesting a scaling law specific to legal reasoning within the RL stage.
• General Capabilities: The models retain strong general capabilities (GSM8K, Math500, C-Eval). Notably, the RL phase showed positive transfer, improving general logical reasoning despite being trained exclusively on legal data.

Significance

The paper claims that LegalOne represents a significant step forward in developing trustworthy and interpretable AI for high-stakes judicial applications. By moving beyond surface-level pattern matching to autonomous, structured legal reasoning, LegalOne addresses the critical gap between general linguistic fluency and professional legal standards. The authors posit that their methodology—specifically the combination of PAS, LEAD, and Curriculum RL—provides a blueprint for building robust, domain-aligned foundation models in other specialized fields. The release of LegalKit and the model weights aims to lower barriers for future research and promote reproducibility in Legal AI.