Research map / Probabilistic and scientific modeling / Variational and generative sampling
QUEST: Quality-Aware Metropolis-Hastings Sampling for Machine Translation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes QUEST, a Metropolis-Hastings MCMC sampler that uses a sentence-level quality estimation metric (CometKiwi-XL) as the energy of a Gibbs distribution over LLM translations, yielding high-quality and diverse translation sets.
Problem
Generating a set of translations that is both high-quality and diverse, since likelihood correlates poorly with quality and reranking with quality metrics risks overfitting ('gaming the metric') and is limited by the candidate pool.
Approach
Treat a QE metric as the energy of a Gibbs distribution over translations and sample from it with Metropolis-Hastings, using a suffix-resampling proposal (pick an index, regenerate the continuation from the LM) that only needs one LM generation and one metric evaluation per step and keeps the chain irreducible and aperiodic.
Key results
- For English-Russian, QUEST improves xComet-XL by up to 2 points over ancestral sampling for both ALMA-7B and Tower-7B.
- QUEST outperforms ancestral sampling in six of eight quality-diversity settings reported on WMT23.
- QUEST produces hypotheses not found in a 4x larger independent ancestral pool, indicating exploration of less-likely high-quality regions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use QUEST sampling instead of ancestral sampling | Alma-7B and Tower-7B · WMT23 English↔German/Russian | xComet-XL: up to 2 points | Helped |
| Use QUEST sampling instead of ancestral sampling | Alma-7B and Tower-7B · WMT23 English↔German/Russian; eight settings | quality-diversity trade-off | Helped |
| Use QUEST sampling instead of ancestral sampling | Alma-7B and Tower-7B · WMT23 German→English | xComet-XL | Hurt |
| Use QUEST sampling instead of ancestral sampling | Alma-7B · WMT23 English→Chinese, English→Czech; WMT22 English↔Icelandic | quality-diversity trade-off | Helped |
| Use QUEST sampling instead of ancestral sampling | QUEST · WMT23 English↔German/Russian | CometKiwi-XL | Helped |
| Increase the number of decoding steps (ablation) | QUEST · WMT23 English→Russian | average reward | Helped |
| Use the sentence-level proposal instead of token-level modifications (ablation) | QUEST proposal · WMT23 example; 25k proposed hypotheses | reward difference | Helped |
| Use token-level modification with uniform probability instead of the sentence-level proposal (ablation) | QUEST proposal · WMT23 example; 25k proposed hypotheses | reward difference | Hurt |
| Use token-level modification with the full posterior instead of the sentence-level proposal (ablation) | QUEST proposal · WMT23 example; 25k proposed hypotheses | reward difference | Hurt |
| Increase the Gibbs temperature β (ablation) | QUEST · WMT23 English→Russian | accepted samples | Helped |
| Replace the QUEST target distribution with the RLHF target distribution (ablation) | QUEST (RLHF-QUEST) · WMT23 English→Russian, Tower-7B | translation quality | No clear effect |
| Use QUEST instead of RLHF-QUEST (ablation) | QUEST · WMT23 English→Russian, Tower-7B | diversity | Helped |
Surprising findings
- Ancestral sampling achieves better xComet-XL than QUEST on German-English, which the authors attribute partly to the QE metric and short passages.
- Quality of QUEST outputs degrades for longer source sentences, despite continued improvement of the reward.
Limitations
- Computationally expensive: one hour for ancestral versus six hours for QUEST on 2000 sources with a single GPU.
- Suffix-only proposal leads to slow mixing once the chain reaches high-quality regions, and many repeated accepted samples.
- Performance on longer sentences lags ancestral sampling.
Experimental setup
Datasets: WMT23
Benchmarks: WMT23 English-German WMT23 English-Russian WMT23 German-English WMT23 Russian-English
Metrics: xComet-XL pairwise BLEU CometKiwi-XL
Baselines: ancestral sampling temperature sampling
Builds on: Metropolis-Hastings quality-aware decoding ancestral sampling
Closest papers by meaning
- Calibrating Translation Decoding with Quality Estimation on LLMs (NeurIPS 2025 · similarity 0.91)
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.87)
- MBR and QE Finetuning: Training-time Distillation of the Best and Most Expensive Decoding Methods (ICLR 2024 · similarity 0.87)
- MT-Ranker: Reference-free machine translation evaluation by inter-system ranking (ICLR 2024 · similarity 0.87)
- Contrastive Preference Optimization: Pushing the Boundaries of LLM Performance in Machine Translation (ICML 2024 · similarity 0.87)
- Constrained Sampling for Language Models Should Be Easy: An MCMC Perspective (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.