Finetuning LLMs for Comparative Assessment Tasks

Vatsal Raina, Adian Liusie, Mark Gales
9/24/2024
cs.CL

No official code found. Best implementation: VatsalRaina/POE (1 star, last commit 2024-12-06).

Abstract

Automated assessment in natural language generation is a challenging task. Instruction-tuned large language models (LLMs) have shown promise in reference-free evaluation, particularly through comparative assessment. However, the quadratic computational complexity of pairwise comparisons limits its scalability. To address this, efficient comparative assessment has been explored by applying comparative strategies on zero-shot LLM probabilities. We propose a framework for finetuning LLMs for comparative assessment to align the model's output with the target distribution of comparative probabilities. By training on soft probabilities, our approach improves state-of-the-art performance while maintaining high performance with an efficient subset of comparisons.

View on arXivView PDF

Code Implementations(1)

VatsalRaina/POECommunity74%

Finetuning LLMs for Comparative Assessment Tasks

11PythonApr 9, 20241 years ago

Similar Papers

Who can we trust? LLM-as-a-jury for Comparative Assessment

88% match2026

Enhancing LLM Evaluations: The Garbling Trick

85% match2024

Language Models can Evaluate Themselves via Probability Discrepancy

85% match1 repo62024

A Judge-free LLM Open-ended Generation Benchmark Based on the Distributional Hypothesis

84% match1 repo1032025

Rating Roulette: Self-Inconsistency in LLM-As-A-Judge Frameworks

84% match2025

A Survey on LLM-as-a-Judge

84% match1 repo1752024

Cite this paper

@article{raina2024finetuning,
  title  = {Finetuning LLMs for Comparative Assessment Tasks},
  author = {Vatsal Raina and Adian Liusie and Mark Gales},
  year   = {2024},
  eprint = {2409.15979},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2409.15979v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2409.15979.svg)](https://codeofpaper.com/paper/2409.15979)

Discussion