Research map / Large language models / Instruction and fine-tuning

Teaching Large Language Models to Translate with Comparison

AAAI 2024 · Large language models › Instruction and fine-tuning · approach: Calibration and knowledge in LLM tuning · 13 citations (OpenAlex)

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 TIM, which instruction-tunes open LLMs (BLOOMZ-7b-mt, LLaMA-2) for translation using output comparison (order-, dictionary-, error-guided examples) and a token-level preference loss over correct vs. incorrect translations.

Problem

Open-sourced LLMs fine-tuned with standard instruction tuning (next-token loss) underperform on translation, especially smaller models trained on limited high-quality data, and tend to hallucinate.

Approach

Teach translation by comparison: include contrastive examples (reordered, dictionary-aligned, and error-annotated translations) plus a token-level preference loss that pushes the hidden-state reward head to prefer correct translations over noisy or LM-generated bad outputs, added to the LM loss without a separate reward model or RLHF.

Key results

• TIM-FixEmb on BLOOMZ-7b-mt with LM-based bad outputs reaches 24.51 BLEU on Zh-En and 37.83 on En-Zh, beating MT-FixEmb (23.08 / 37.09) and the NLLB-3.3b baseline.
• On FLORES-200 with LLaMA-2-7b, TIM-FixEmb (LM-based) reaches 42.91 BLEU on De-En versus 42.14 for MT-FixEmb.
• TIM-LLaMA-13b is reported as top-ranked reference-free metric on De-En system-level accuracy tie-ranked, with Pearson 50.37 on De-En, outperforming COMET-QE and COMETKiwi.

Measured effects reported in the paper

Surprising findings

• Adding dictionary alignment notes to inputs harmed BLEU in inference-strategy experiments, which the authors attribute to dictionary words being mostly common words.
• Preference loss alone (without output comparison data) reduced hallucination relative to plain MT tuning, and the 'no error' instruction gave no significant gain.

Limitations

• Gains over MT baselines are modest (roughly 0.3-1.5 BLEU) and the method is evaluated on a small set of language pairs.
• Dictionary-guided notes hurt performance in some settings, and better dictionary use is left for future work.
• Zero-shot results rely on reported numbers from other papers for several baselines.

Experimental setup

Datasets: WMT22 test sets FLORES-200 dev-test WMT newstest2017-2021 validation data Alpaca MQM newstest2020 (en-de, zh-en) MUSE bilingual dictionaries

Benchmarks: WMT22 test sets (De-En, En-De, Zh-En, En-Zh) FLORES-200 dev-test WMT22 multilingual-to-English zero-shot (cs, ja, ru, uk to en) WMT22 Metrics Shared Task MQM (De-En, En-De)

Metrics: BLEU COMET hallucination rate (sentence and token level) system-level accuracy Pearson correlation

Baselines: Alpaca-LoRA Alpaca-Full MT-LoRA MT-FixEmb MT-Full WMT22 winners NLLB-3.3B Alpaca-7b Vicuna-13b BayLing-7b BayLing-13b ChatGPT

Builds on: instruction tuning LoRA RLHF reward modeling Alpaca

Closest papers by meaning

• A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.91)
• When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method (ICLR 2024 · similarity 0.90)
• Instruction Tuning With Loss Over Instructions (NeurIPS 2024 · similarity 0.89)
• InstructZero: Efficient Instruction Optimization for Black-Box Large Language Models (ICML 2024 · similarity 0.89)
• Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.89)
• Two-stage LLM Fine-tuning with Less Specialization and More Generalization (ICLR 2024 · similarity 0.88)