SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2506.06292 · ACL · 2025

Mutual-Taught for Co-adapting Policy and Reward Models

Ziyi Yang, Ming Yan, Xiaojun Quan, Fanqi Wan, Weizhou Shen, Longguang Zhong, Tianyuan Shi, Canbin Huang

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 7 functions out of this paper's own repositories and ran 1 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
Stycoo/Mutual-Taught canonical 1 of 7
FunctionStatusWhere it lives
maybe_torch Ran Stycoo/Mutual-Taught/reward_model_merge/rm_merge_custom.py
code served (permissive licence) · get_code("10ff293597eab810")
build_dpo_dataset Not yet run Stycoo/Mutual-Taught/completion_inference/build_dataset.py
code served (permissive licence) · get_code("f7d5662f1399084a")
create_node Not yet run Stycoo/Mutual-Taught/completion_inference/completion_infer.py
code served (permissive licence) · get_code("11ae53d561cabab0")
lerp Not yet run Stycoo/Mutual-Taught/reward_model_merge/rm_merge_custom.py
code served (permissive licence) · get_code("ed054b3a2f7dde35")
read_json_file Not yet run Stycoo/Mutual-Taught/completion_inference/build_dataset.py
code served (permissive licence) · get_code("1acc67cef16f4df5")
slerp Not yet run Stycoo/Mutual-Taught/reward_model_merge/rm_merge_custom.py
code served (permissive licence) · get_code("343787d4eb1da770")
split_by_newlines Not yet run Stycoo/Mutual-Taught/completion_inference/completion_infer.py
code served (permissive licence) · get_code("615a32d374e639ea")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

During the preference optimization of large language models (LLMs), distribution shifts may arise between newly generated model samples and the data used to train the reward model (RM). This shift reduces the efficacy of the RM, which in turn negatively impacts the performance of the policy model (PM). To address this challenge, we propose Mutual-Taught, a self-training method that iteratively improves both the PM and RM without requiring additional human annotation. Our approach mirrors the expectation-maximization (EM) algorithm. In the E-step, the PM is updated using feedback from the current RM, guiding the PM toward a better approximation of the latent optimal preference distribution. In the M-step, we update the RM by constructing training data from the outputs of the PM before and after the E-step update. This process ensures that the RM adapts to the evolving policy distribution. Experimental results demonstrate that this iterative approach leads to consistent improvements in both models. Specifically, our 8B policy model, Llama-3-8B-Instruct-MT, achieves a length-controlled win rate of 54.1% on AlpacaEval-2, while our 8B reward model, FsfairX-Llama3-RM-MT, performs on par with GPT-4o-2024-08-06 on Re-wardBench. Our code is available at https: //github.com/Stycoo/Mutual-Taught.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2506.06292")
get_code_for_paper("2506.06292")
have("2506.06292")

Connect an agent — have() is free.