Generating Adversarial Texts for Machine Translation via GRPO
Florian Zogaj, Jakob Hütteneder, Giovanni De Muri, Federico Villa, Aryan Sood, Vilém Zouhar
cs.CL
Quick verdict
GRPO fine-tuning rewrites English texts to cut MT quality (avg COMET 0.63→0.48, WMT25 En→It) while keeping grammar; caveat: eval metric = reward metric.
Summary
The paper fine-tunes Llama-3.1-8B-Instruct (LoRA) with GRPO to rewrite English source sentences so that MT systems translate them worse, turning existing benchmarks into challenge sets. The reward combines a difficulty term — the CometKiwi score drop between original and rewrite, computed on English→Italian translations by NLLB-200-600M and OPUS-MT (en-it) — with constraints on length, semantic similarity (MiniLM ≥0.70), CoLA acceptability, and LanguageTool errors. Four variants are trained (Sentinel, COMET-NLLB, COMET-Helsinki, COMET-Average). On WMT25, COMET-Average lowers the average CometKiwi score from 0.63 (original) to 0.48 versus 0.64 for the base model, with fewer grammatical errors than the original data (1.36 vs 3.45) and preserved diversity; the Sentinel-reward model fails to transfer to COMET metrics. Gains transfer to unseen WMT19–WMT24 (WMT24: 0.74→0.54), and a human evaluation with 10 German-native annotators on English→German NLLB translations shows translation quality dropping from 56.51 to 41.29, with naturalness down 7.1 and grammaticality down 2.6 points. Main limitations: difficulty is measured with the same QE metric family used as the reward (no held-out or reference-based metrics), only two weak MT systems are evaluated (no frontier systems), meaning preservation is never human-verified, and the human study lacks significance testing. Code is released.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- GRPO fine-tuning of Llama-3.1-8B with QE-difficulty rewards produces rewrites that cut English→Italian average CometKiwi on WMT25 from 0.63 to 0.48 while the base model remains at 0.64.
- The difficulty gains transfer to unseen WMT19–WMT24 benchmarks (e.g., WMT24 average COMET 0.74→0.54) without increasing grammatical errors.
- Averaging COMET rewards over two MT systems (COMET-Average) yields a more robust generator than any single estimator, whereas a Sentinel-only reward fails to transfer to COMET metrics.
- Human evaluation on unseen English→German confirms substantially lower translation quality (56.51→41.29) with moderate naturalness loss (−7.1) and small grammaticality change (−2.6).
- The effect is not explained by noise: rewrites contain fewer LanguageTool grammatical errors than the original benchmark data (1.36 vs 3.45 per instance on WMT25).
Strengths
Reward-source ablation evaluated cross-metric, with an honest negative result
Table 2 evaluates all four fine-tuned variants under all four difficulty metrics; the Sentinel-reward model wins only on Sentinel and is slightly easier than the original data under COMET (0.66 vs 0.63), while COMET-based rewards transfer — a candid Goodhart demonstration the authors report rather than hide.
Difficulty is not bought with noise
All fine-tuned models produce fewer LanguageTool errors than the original data (WMT25: 1.36 vs 3.45; WMT24: 0.78 vs 1.05), and COMET-Average keeps Self-chrF (21.65 vs 21.61), unique topics (400 vs 403) and output length (116.17 vs 119.95) near original on WMT25 (Table 2).
Cross-target-language transfer tested with humans
Transparent collapse and checkpoint selection
Above-average reproducibility posture
Weaknesses
Reward-evaluation circularity
Difficulty defined only against weak MT systems
Human evaluation lacks inferential statistics
Meaning preservation never human-verified
LLM judge shares its identity with the policy
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper trains a rewriter that converts existing MT benchmarks into challenge sets: Llama-3.1-8B-Instruct with LoRA, fine-tuned via GRPO (TRL GRPOTrainer + unsloth), where the reward couples a translation-difficulty term, rtd = clip(5·(QE(x)−QE(y)), −2, 2), computed from wmt22-cometkiwi-da over English→Italian translations by NLLB-200-distilled-600M and OPUS-MT-en-it, with constraint rewards for length (0.4), semantic similarity via MiniLM cosine (0.8), CoLA acceptability (0.5) and LanguageTool errors (0.2) (Sec 3.3). Four variants are trained: Sentinel-reward, COMET-NLLB, COMET-Helsinki and COMET-Average. Central claims: (1) WMT25 average COMET falls 0.63→0.48 vs 0.64 for the base model; (2) gains transfer to unseen WMT19–24; (3) human evaluation on unseen En→German confirms the effect (56.51→41.29) with moderate naturalness cost; (4) gains do not come from ungrammaticality. Within the paper's evaluation frame the claims are supported; the frame itself is the main caveat.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper do?
It fine-tunes Llama-3.1-8B with GRPO (reinforcement learning) so the model rewrites ordinary English sentences into versions that machine translation systems translate worse, while keeping the text grammatical and roughly meaning-preserving. The goal is to turn saturated benchmarks like WMT into harder challenge sets for exposing MT weaknesses.
How much harder do the rewrites make translation?
On WMT25 English→Italian, the average CometKiwi quality score falls from 0.63 (original sentences) to 0.48 (rewrites); the un-fine-tuned base model stays at 0.64. On unseen WMT24 the score drops from 0.74 to 0.54. In a human study, raters scored NLLB English→German translations of the rewrites at 41.29 on a 1–100 scale versus 56.51 for originals.
Does the method just make text noisier or ungrammatical?
No. By LanguageTool counts the rewrites contain fewer grammatical errors than the original benchmark data (1.36 vs 3.45 per instance on WMT25), and diversity stays near original levels. Human raters did note a moderate naturalness drop (−7.1 points on a 1–100 scale) and a small grammaticality drop (−2.6).
Is code or data available?
The authors release code at github.com/FlorianZogaj/Breaking-MT, with prompts and a full model/version list in the appendices. All datasets (WMT19–25) and models (Llama-3.1-8B, CometKiwi, NLLB, OPUS-MT, Sentinel) are public; trained checkpoints are not explicitly stated as released.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1, 3 and 5 are plot images; only their captions and in-text numbers are extractable, so training-curve shapes and scatter details are taken from the authors' descriptions rather than verifiable plotted values.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.