Token-weighted Direct Preference Optimization with Attention

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Huang, Chengyu, Li, Zhuohang, Chou, Sheng-Yen, Cardie, Claire
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911717653479424
author Huang, Chengyu
Li, Zhuohang
Chou, Sheng-Yen
Cardie, Claire
author_facet Huang, Chengyu
Li, Zhuohang
Chou, Sheng-Yen
Cardie, Claire
contents Direct Preference Optimization (DPO) aligns Large Language Models with human preferences without the need for a separate reward model. However, DPO treats all tokens in responses equally, neglecting the differing importance of individual tokens. Existing token-level PO methods compute the token weights using either token-position-based heuristic functions or probability estimates given by a separately trained model, which lacks robustness and incurs extra training cost. In contrast, we propose Token-weighted DPO (TwDPO) -- a novel training objective grounded on token-weighted RL -- and AttentionPO -- an instantiation of TwDPO that uses attention from the LLM itself to estimate token weights. AttentionPO prompts the LLM to serve as a pairwise judge and check where the model attends when comparing the responses. This design makes AttentionPO content-aware, adjusting weights based on response content, and efficient, incurring only two extra forward passes per example. Experiment results show that AttentionPO significantly improves performance on AlpacaEval, MT-Bench, and ArenaHard, surpassing existing Preference Optimization methods.
format Preprint
id arxiv_https___arxiv_org_abs_2605_21883
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Token-weighted Direct Preference Optimization with Attention
Huang, Chengyu
Li, Zhuohang
Chou, Sheng-Yen
Cardie, Claire
Computation and Language
Direct Preference Optimization (DPO) aligns Large Language Models with human preferences without the need for a separate reward model. However, DPO treats all tokens in responses equally, neglecting the differing importance of individual tokens. Existing token-level PO methods compute the token weights using either token-position-based heuristic functions or probability estimates given by a separately trained model, which lacks robustness and incurs extra training cost. In contrast, we propose Token-weighted DPO (TwDPO) -- a novel training objective grounded on token-weighted RL -- and AttentionPO -- an instantiation of TwDPO that uses attention from the LLM itself to estimate token weights. AttentionPO prompts the LLM to serve as a pairwise judge and check where the model attends when comparing the responses. This design makes AttentionPO content-aware, adjusting weights based on response content, and efficient, incurring only two extra forward passes per example. Experiment results show that AttentionPO significantly improves performance on AlpacaEval, MT-Bench, and ArenaHard, surpassing existing Preference Optimization methods.
title Token-weighted Direct Preference Optimization with Attention
topic Computation and Language
url https://arxiv.org/abs/2605.21883