Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ALMol: Aligned Language-Molecule Translation LLMs through Offline Preference Contrastive Optimisation

Dimitris Gkoumas

TL;DR

The paper tackles training efficacy and out-of-distribution generalization in language–molecule translation by introducing contrastive preference optimisation (CTO), an offline, closed-form RLHF-like objective that leverages offline preference data $\mathcal{D}$ to discourage merely adequate translations. By training on only 10% of the L+M-24 dataset and using a BC-regularized CPO loss, the approach achieves up to a 32% improvement over strong baselines, even in cross-modal and agnostic initialization scenarios. A fine-grained, domain-agnostic evaluation framework is proposed to measure hallucinations and factual consistency, including QAFactEval for language outputs and Chr-F for molecule generation, plus bias analyses. The work demonstrates that CTO can yield robust, generalizable language–molecule translation with improved factual alignment and reduced length bias, offering practical benefits for chemistry-focused AI applications and responsible deployment.

Abstract

The field of chemistry and Artificial Intelligence (AI) intersection is an area of active research that aims to accelerate scientific discovery. The integration of large language models (LLMs) with scientific modalities has shown significant promise in this endeavour. However, challenges persist in effectively addressing training efficacy and the out-of-distribution problem, particularly as existing approaches rely on larger models and datasets. In this context, we focus on machine language-molecule translation and deploy a novel training approach called contrastive preference optimisation, which avoids generating translations that are merely adequate but not perfect. To ensure generalisability and mitigate memorisation effects, we conduct experiments using only 10% of the data. Our results demonstrate that our models achieve up to a 32% improvement compared to counterpart models. Finally, we introduce a fine-grained, domain-agnostic evaluation method to assess hallucination in LLMs and promote responsible use.

ALMol: Aligned Language-Molecule Translation LLMs through Offline Preference Contrastive Optimisation

TL;DR

The paper tackles training efficacy and out-of-distribution generalization in language–molecule translation by introducing contrastive preference optimisation (CTO), an offline, closed-form RLHF-like objective that leverages offline preference data to discourage merely adequate translations. By training on only 10% of the L+M-24 dataset and using a BC-regularized CPO loss, the approach achieves up to a 32% improvement over strong baselines, even in cross-modal and agnostic initialization scenarios. A fine-grained, domain-agnostic evaluation framework is proposed to measure hallucinations and factual consistency, including QAFactEval for language outputs and Chr-F for molecule generation, plus bias analyses. The work demonstrates that CTO can yield robust, generalizable language–molecule translation with improved factual alignment and reduced length bias, offering practical benefits for chemistry-focused AI applications and responsible deployment.

Abstract

The field of chemistry and Artificial Intelligence (AI) intersection is an area of active research that aims to accelerate scientific discovery. The integration of large language models (LLMs) with scientific modalities has shown significant promise in this endeavour. However, challenges persist in effectively addressing training efficacy and the out-of-distribution problem, particularly as existing approaches rely on larger models and datasets. In this context, we focus on machine language-molecule translation and deploy a novel training approach called contrastive preference optimisation, which avoids generating translations that are merely adequate but not perfect. To ensure generalisability and mitigate memorisation effects, we conduct experiments using only 10% of the data. Our results demonstrate that our models achieve up to a 32% improvement compared to counterpart models. Finally, we introduce a fine-grained, domain-agnostic evaluation method to assess hallucination in LLMs and promote responsible use.
Paper Structure (18 sections, 6 equations, 9 figures, 2 tables)

This paper contains 18 sections, 6 equations, 9 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Methodology
  4. Task Formulation
  5. Contrastive Preference Optimisation
  6. Proposed Evaluation Methodology
  7. Language Evaluation:
  8. Molecule Evaluation:
  9. Bias Evaluation:
  10. Experiments
  11. Data
  12. Bechmark Models
  13. Experimental Settings
  14. Experiment Results
  15. Evaluation Results
  16. ...and 3 more sections

Figures (9)

  • Figure 1: A toy example illustrating a factual inconsistency between a generated and a reference caption. The QAFactEval metric selects a noun-phrase answer from the generated caption. A QG model then generates an associated question that a QA model answers based on the reference caption. The scores measure the semantic overlap between the QA model's answer and the selected answer from the generated caption
  • Figure 2: Factual consistency in generated captions against references, assessed through (A) semantic overlap, (B) F1 accuracy, and (C) answerability using QAFactEval ($\S$\ref{['sec:eval_method']}) across various LLMs.
  • Figure 3: Overlaps of n-gram matches between generated and reference molecules as captured by the char-F ($\S$\ref{['sec:eval_method']}) score across various LLMs.
  • Figure 4: Length-bias across different LLMs.
  • Figure 5: Instruction for molecule to language translation, i.e., $M \rightarrow L$
  • ...and 4 more figures