Decomposition-Enhanced Training for Post-Hoc Attributions In Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Balasubramanian, Sriram, Basu, Samyadeep, Goswami, Koustava, Rossi, Ryan, Manjunatha, Varun, Santhosh, Roshan, Zhang, Ruiyi, Feizi, Soheil, Lipka, Nedim
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911250869387264
author Balasubramanian, Sriram
Basu, Samyadeep
Goswami, Koustava
Rossi, Ryan
Manjunatha, Varun
Santhosh, Roshan
Zhang, Ruiyi
Feizi, Soheil
Lipka, Nedim
author_facet Balasubramanian, Sriram
Basu, Samyadeep
Goswami, Koustava
Rossi, Ryan
Manjunatha, Varun
Santhosh, Roshan
Zhang, Ruiyi
Feizi, Soheil
Lipka, Nedim
contents Large language models (LLMs) are increasingly used for long-document question answering, where reliable attribution to sources is critical for trust. Existing post-hoc attribution methods work well for extractive QA but struggle in multi-hop, abstractive, and semi-extractive settings, where answers synthesize information across passages. To address these challenges, we argue that post-hoc attribution can be reframed as a reasoning problem, where answers are decomposed into constituent units, each tied to specific context. We first show that prompting models to generate such decompositions alongside attributions improves performance. Building on this, we introduce DecompTune, a post-training method that teaches models to produce answer decompositions as intermediate reasoning steps. We curate a diverse dataset of complex QA tasks, annotated with decompositions by a strong LLM, and post-train Qwen-2.5 (7B and 14B) using a two-stage SFT + GRPO pipeline with task-specific curated rewards. Across extensive experiments and ablations, DecompTune substantially improves attribution quality, outperforming prior methods and matching or exceeding state-of-the-art frontier models.
format Preprint
id arxiv_https___arxiv_org_abs_2510_25766
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Decomposition-Enhanced Training for Post-Hoc Attributions In Language Models
Balasubramanian, Sriram
Basu, Samyadeep
Goswami, Koustava
Rossi, Ryan
Manjunatha, Varun
Santhosh, Roshan
Zhang, Ruiyi
Feizi, Soheil
Lipka, Nedim
Computation and Language
Large language models (LLMs) are increasingly used for long-document question answering, where reliable attribution to sources is critical for trust. Existing post-hoc attribution methods work well for extractive QA but struggle in multi-hop, abstractive, and semi-extractive settings, where answers synthesize information across passages. To address these challenges, we argue that post-hoc attribution can be reframed as a reasoning problem, where answers are decomposed into constituent units, each tied to specific context. We first show that prompting models to generate such decompositions alongside attributions improves performance. Building on this, we introduce DecompTune, a post-training method that teaches models to produce answer decompositions as intermediate reasoning steps. We curate a diverse dataset of complex QA tasks, annotated with decompositions by a strong LLM, and post-train Qwen-2.5 (7B and 14B) using a two-stage SFT + GRPO pipeline with task-specific curated rewards. Across extensive experiments and ablations, DecompTune substantially improves attribution quality, outperforming prior methods and matching or exceeding state-of-the-art frontier models.
title Decomposition-Enhanced Training for Post-Hoc Attributions In Language Models
topic Computation and Language
url https://arxiv.org/abs/2510.25766