Distillation and Refinement of Reasoning in Small Language Models for Document Re-ranking

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Samarinas, Chris, Zamani, Hamed
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866908426322313216
author Samarinas, Chris
Zamani, Hamed
author_facet Samarinas, Chris
Zamani, Hamed
contents We present a novel approach for training small language models for reasoning-intensive document ranking that combines knowledge distillation with reinforcement learning optimization. While existing methods often rely on expensive human annotations or large black-box language models, our methodology leverages web data and a teacher LLM to automatically generate high-quality training examples with relevance explanations. By framing document ranking as a reinforcement learning problem and incentivizing explicit reasoning capabilities, we train a compact 3B parameter language model that achieves state-of-the-art performance on the BRIGHT benchmark. Our model ranks third on the leaderboard while using substantially fewer parameters than other approaches, outperforming models that are over 20 times larger. Through extensive experiments, we demonstrate that generating explanations during inference, rather than directly predicting relevance scores, enables more effective reasoning with smaller language models. The self-supervised nature of our method offers a scalable and interpretable solution for modern information retrieval systems.
format Preprint
id arxiv_https___arxiv_org_abs_2504_03947
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Distillation and Refinement of Reasoning in Small Language Models for Document Re-ranking
Samarinas, Chris
Zamani, Hamed
Information Retrieval
Computation and Language
We present a novel approach for training small language models for reasoning-intensive document ranking that combines knowledge distillation with reinforcement learning optimization. While existing methods often rely on expensive human annotations or large black-box language models, our methodology leverages web data and a teacher LLM to automatically generate high-quality training examples with relevance explanations. By framing document ranking as a reinforcement learning problem and incentivizing explicit reasoning capabilities, we train a compact 3B parameter language model that achieves state-of-the-art performance on the BRIGHT benchmark. Our model ranks third on the leaderboard while using substantially fewer parameters than other approaches, outperforming models that are over 20 times larger. Through extensive experiments, we demonstrate that generating explanations during inference, rather than directly predicting relevance scores, enables more effective reasoning with smaller language models. The self-supervised nature of our method offers a scalable and interpretable solution for modern information retrieval systems.
title Distillation and Refinement of Reasoning in Small Language Models for Document Re-ranking
topic Information Retrieval
Computation and Language
url https://arxiv.org/abs/2504.03947