Technical Summary: Unlocking Fine-Grained Translation Quality Estimation in LRMs through Synergistically Evolving Implicit and Explicit Reasoning

1. Problem Statement

Large Reasoning Models (LRMs) have demonstrated strong multilingual capabilities and achieve state-of-the-art (SOTA) performance in sentence-level Translation Quality Estimation (QE). However, they struggle significantly with fine-grained QE, which requires identifying precise error spans and their severity (e.g., MINOR vs. MAJOR) rather than just assigning scalar scores.

The authors argue that the core challenge is not a lack of multilingual knowledge within LRMs, but rather the intrinsic difficulty of learning the fine-grained discrimination task. Existing approaches often fail because:

• Fluency Bias: LRMs may overlook quality issues in fluent sentences.
• Reasoning Faithfulness: LRMs may post-hoc rationalize predetermined answers or struggle to generate meaningful reasoning trajectories for error localization without explicit guidance.
• Task Complexity: Directly applying Reinforcement Learning with Verifiable Reward (RLVR) to complex, fine-grained tasks often fails if the model cannot initially "roll out" meaningful reasoning paths.

2. Methodology: RIEQE Framework

The paper proposes RIEQE (Reasoning both Implicitly and Explicitly for QE), a two-stage training framework designed to enable the synergistic co-evolution of implicit reasoning (layer-wise internal computations without explicit tokens) and explicit reasoning (token-level chain-of-thought).

2.1 Task Decomposition

To make implicit reasoning feasible, the complex span-level QE task is decomposed into three sequential, straightforward subtasks:

1. UnitSegmentation: Segmenting the translated sentence into semantically coherent units to reduce the search space for errors.
2. ErrorDetection: Examining each unit to determine if it contains an error span.
3. SeverityClassification: Predicting the severity (MINOR/MAJOR) for detected error spans.

This decomposition allows the model to learn from simpler, more direct input-output mappings, avoiding the need for complex reasoning chains during the initial learning phase.

2.2 Two-Stage Training Pipeline

The framework proceeds in two distinct stages:

Stage 1: NonThinking-SFT (Supervised Fine-Tuning)

• Objective: Strengthen the model's implicit reasoning tendency and capability.
• Process: The model is fine-tuned on the decomposed subtasks using ground-truth answers without generating reasoning chains (i.e., no "Thinking" tokens).
• Mechanism: This forces the model to compress reasoning capabilities into its internal hidden states (layer-wise computations) to directly predict outputs. The authors hypothesize that this builds a robust foundation for subsequent explicit reasoning.

Stage 2: Thinking-RLVR (Reinforcement Learning with Verifiable Reward)

• Objective: Strengthen explicit reasoning while further enhancing implicit capabilities.
• Process: Using a smaller dataset (1k random samples), the model undergoes RLVR (specifically using the GRPO algorithm) on the ErrorDetection and SeverityClassification subtasks.
• Reward Design:
  • For ErrorDetection: The reward is the character-level Matthews Correlation Coefficient (MCC).
  • For SeverityClassification: The reward is the proportion of correctly predicted severity labels.
• Outcome: The model learns to spontaneously organize coherent reasoning chains on top of the strengthened implicit reasoning foundation.

3. Key Contributions

• RIEQE Framework: A novel two-stage training paradigm that decouples implicit and explicit reasoning training to address the failure of LRMs in fine-grained QE.
• Synergistic Co-evolution: The paper demonstrates that implicit and explicit reasoning are not mutually exclusive but can be trained to benefit each other. NonThinking-SFT provides the necessary internal intuition, while Thinking-RLVR refines the explicit output.
• Task Decomposition Strategy: Proving that breaking down complex discriminative tasks into simpler subtasks is critical for enabling reliable implicit reasoning in LLMs.
• SOTA Performance: Achieving state-of-the-art results on WMT fine-grained QE benchmarks across multiple language pairs (high-resource: zh-en, en-de, en-ru; low-resource: en-mr).

4. Experimental Results

Evaluated on WMT 2022/2023 test sets using the Qwen3-4B-Thinking-2507 backbone:

• Performance Gains: RIEQE significantly outperforms strong baselines (including encoder-based models like xCOMET and DCSQE, and other LLMs like GPT-5.5) in both word-level and span-level metrics (MCC, F1, Precision, Recall).
  • On zh-en and en-de, RIEQE achieves the highest precision among all baselines, indicating accurate error localization rather than just recall-driven trade-offs.
  • On en-mr (low-resource), the method dramatically improves performance, surpassing models specifically trained on large-scale parallel corpora for this language pair.
• Implicit Reasoning Capability: The RIEQE-NonThinking variant (skipping explicit reasoning at inference) remains competitive with the best encoder-based models, demonstrating that the model has successfully internalized the reasoning capability.
• Synergy Evidence:
  • Hidden State Separability: Probing experiments show that NonThinking-SFT increases the separability of hidden states regarding error information, and Thinking-RLVR further enhances this.
  • Consistency: There is a high consistency between the implicit predictions (from hidden states) and the explicit reasoning paths in the final RLVR-trained model, suggesting the explicit reasoning is guided by the internalized implicit knowledge.

5. Significance and Claims

The paper claims that the potential of LLMs in fine-grained discriminative tasks has not been fully elicited due to the difficulty of the task rather than a lack of knowledge. By proposing RIEQE, the authors demonstrate that:

1. Implicit reasoning is foundational: Improving a model's ability to perform internal, layer-wise discrimination (via NonThinking-SFT) is a prerequisite for successful explicit reasoning in complex tasks.
2. Co-evolution is possible: Implicit and explicit reasoning can be trained to evolve synergistically, where explicit optimization further strengthens implicit tendencies.
3. Efficiency: The approach offers a computationally efficient option for fine-grained prediction (via the NonThinking variant) while maintaining high performance.

The authors conclude that this work offers a new research perspective on integrating implicit and explicit thinking in modern LRMs, suggesting that the two modes are deeply intertwined.

6. Limitations

The paper acknowledges several limitations:

• The study does not systematically investigate the optimal data composition or the impact of all training hyperparameters.
• The analysis of the co-evolution mechanism is primarily focused on final task performance and hidden state separability, rather than directly probing the dynamic interaction of the two reasoning modes during training.
• The framework is currently designed for fine-grained QE, and its applicability to other reasoning tasks remains unexplored.