Remedy-R Agent: Generative MT Evaluation

• Remedy-R Agent is a reasoning-driven framework that integrates explicit error analysis for transparent machine translation evaluation and refinement.
• It utilizes a three-step evaluate–revise pipeline where structured chain-of-thought feedback guides zero-shot translation improvements.
• Empirical results show consistent quality gains and increased explanation faithfulness on various MT benchmarks without supervised post-editing data.

The Remedy-R Agent refers to a reasoning-driven, generative approach to machine translation (MT) evaluation and refinement that leverages explicit error analysis to guide post-editing, yielding improved translation quality without supervised post-editing data or closed-source LLM guidance. The concept appears in "Remedy-R: Generative Reasoning for Machine Translation Evaluation without Error Annotations" (Tan et al., 21 Dec 2025), where the agent encapsulates a simple, interpretable evaluate–revise loop built around the Remedy-R evaluation metric.

1. Design Motivation and Foundations

Remedy-R was introduced to address critical deficiencies in prevailing neural MT metrics, such as COMET and MetricX, which, despite strong benchmark performance, reduce translation output to a scalar score and provide little interpretability or actionable insight. Remedy-R instead outputs a structured, step-by-step analysis assessing three dimensions: accuracy, fluency, completeness, and only then produces a final quality score. The Remedy-R Agent leverages these factored rationales in a pipeline designed to demonstrate the faithful and practical usefulness of open-ended, reasoning-driven evaluation for automated translation refinement (Tan et al., 21 Dec 2025).

The central hypothesis motivating the Remedy-R Agent is that if a metric's explanations truly reflect the salient flaws in translation, then conditioning a generative model on these explanations as feedback should improve the translation, even if the generator was not explicitly trained for post-editing.

2. Evaluate–Revise Pipeline Architecture

The Remedy-R Agent consists of three components:

• The initial translation system ($M_\mathrm{base}$)
• The Remedy-R evaluator ($M_\mathrm{feedback}$), which produces a structured chain-of-thought (COT) analysis and score given (source, translation)
• A refinement generator ($M_\mathrm{refinement}$), which is prompted to revise the translation based on (source, initial translation, COT analysis)

The workflow executes as follows:

1. $mt_0 \leftarrow M_\mathrm{base}(\mathrm{src})$
2. $$(A, s) \leftarrow M_\mathrm{feedback}(\mathrm{src}, mt_0)$$, where $A$ is the multi-dimensional COT analysis, and $s$ is the scalar score
3. $$mt_1 \leftarrow M_\mathrm{refinement}(\mathrm{src}, mt_0, A)$$

This loop can be iterated if further refinement is desired. The entire pipeline operates without explicit supervision for post-editing (Tan et al., 21 Dec 2025).

3. Formal Objectives and Learning Paradigm

3.1 Evaluation Function and Reasoning Structure

Given input $x = (\mathrm{src}, mt)$, the evaluator $M_\mathrm{feedback}$ defines a probability distribution $M_\mathrm{feedback}$0 over output sequences $M_\mathrm{feedback}$1. Each output consists of:

• $M_\mathrm{feedback}$2 (the chain-of-thought blocks for accuracy, fluency, completeness)
• $M_\mathrm{feedback}$3 (final scalar score)

The evaluation score is $M_\mathrm{feedback}$4. While the model does not regress explicit separate subscores, the instruction template ensures all three dimensions are analyzed; the overall score is implicitly a weighted sum of latent subscores.

3.2 Reinforcement Learning for Evaluator Training

Remedy-R is trained on pairwise human preference data using Proximal Policy Optimization (PPO), incorporating a two-part reward:

• Rank accuracy: $M_\mathrm{feedback}$5 if model's ranking matches reference, $M_\mathrm{feedback}$6 otherwise
• Huber-shaped calibration: $M_\mathrm{feedback}$7 with error shaping term $M_\mathrm{feedback}$8 based on score deviation from human judgment

The PPO objective is

$M_\mathrm{feedback}$9

3.3 Refinement Objective

The refinement generator is a LLM with parameter $M_\mathrm{refinement}$0, run in conditional generation mode (not fine-tuned for post-editing). The loss is the negative log-likelihood of the refined translation conditioned on (source, initial translation, analysis):

$M_\mathrm{refinement}$1

In practice, $M_\mathrm{refinement}$2 simply generates a new translation conditioned on the analysis.

4. Empirical Results and Evaluations

Experiments on WMT24++ and similar MT meta-evaluation benchmarks show that the Remedy-R Agent pipeline delivers consistent improvements across diverse base translators:

Base MT Model	chrF / BLEU Improvement	XCOMET Improvement
Qwen2.5-7B	+0.8 chrF	+1.8
Qwen2.5-14B	+2.2 chrF	+2.8
Qwen2.5-32B	+1.1 chrF	+2.3
ALMA-R-7B	+1.5 BLEU	+1.4
ALMA-R-13B	+3.1 BLEU	+1.4
GPT-4o-mini	+5.2 XCOMET
Gemini-2.0-Flash	+2.2 XCOMET

The gains are observed even in zero-shot post-editing, confirming that Remedy-R's explicit rationales capture actionable translation flaws. Furthermore, COT explanation faithfulness, as measured by GPT-4o-mini judgment, increases with model scale (76.9→79.5 for 7B→32B on 900 WMT22 segments), suggesting the model's explanations are supported by observable evidence (Tan et al., 21 Dec 2025).

Remedy-R also demonstrates strong performance on out-of-distribution stress tests (MSLC24), avoiding major pitfalls of scalar metrics (e.g., high scores for source copies), and benefits from reward shaping in RL training (+1.7% average accuracy on WMT22 with Huber loss).

5. System Implementation and Architectural Aspects

Remedy-R and the Agent pipeline use Qwen2.5 decoder-only models, evaluated at the 7B, 14B, and 32B parameter scales. Training occurs on 60k pairwise preferences (English–German, Chinese–English MQM); RL uses the VeRL + PPO framework with standard hyperparameters (Adam, 5e-6, batch size 2048 tokens, PPO ε=0.2, etc.), and typically converges within 27 hours on current hardware (H100/H200 GPUs). The same checkpoint is used for both evaluation and refinement; no further fine-tuning is applied to the reviser. Refinement uses prompting with either beam or top-p sampling.

The explicit pseudocode for the pipeline appears as:

$M_\mathrm{refinement}$3 (Tan et al., 21 Dec 2025)

6. Limitations, Ablations, and Open Questions

Major limitations stem from lightweight revision mechanisms and supervision constraints:

• The refinement model is not explicitly trained for post-editing, operating purely in zero-shot mode with chain-of-thought feedback.
• Remedy-R's explanations, while faithful and actionable, do not offer granular, referential error highlighting, and cannot substitute for high granularity MQM annotation.
• OOD robustness is strong for gross pathologies (e.g., source copying), but subtle cross-linguistic issues may pass undetected.
• Iterative refinement is possible but not exhaustively explored; further improvements might be achieved by adaptive loop scheduling.

Ablation demonstrates that RL reward shaping is beneficial: adding Huber-shaped error signals to pairwise ranking yields up to +1.7% accuracy on WMT22. On commercial LLM systems, Remedy-R Agent surpasses or closely matches self-refinement baselines, sometimes at much smaller scale (32B vs. 180B). This suggests that explanation-driven revision generalizes beyond the Remedy-R family and remains useful in practical scenarios.

Open directions include integrating more sophisticated post-editing models, joint reasoning-score learning, and extending explainable evaluation to more diverse MT styles and languages.

7. Significance and Implications

The Remedy-R Agent demonstrates that transparent, stepwise evaluators can serve as both reliable metrics and as interactive agents for translation improvement, with gains in both score and human-aligned faithfulness. Unlike traditional scalar metrics, Remedy-R's explicit decompositions of accuracy, fluency, and completeness enable direct interpretability, guide targeted revision, and provide quality assurance in post-editing workflows—even in the absence of error span annotation or closed-source distillation.

This architecture provides empirical evidence for the practical impact of reasoning-centric evaluation in real MT pipelines and highlights the evolving intersection between evaluation metrics, explanation, and automatic system improvement (Tan et al., 21 Dec 2025).