id
string
sources
list
title
string
abstract
string
authors
list
categories
list
fields_of_study
list
published_date
timestamp[s]
url
string
pdf_url
string
arxiv_id
string
doi
string
citation_count
int64
influential_citation_count
int64
has_code
bool
code_url
string
venue
string
quality_score
float64
5c0c37b90e9c180a75fa7d09218f3930b3b33de642aee8c39728246017d3d554
[ "arxiv", "semantic_scholar" ]
SPARK: Synergistic Policy And Reward Co-Evolving Framework
Recent Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) increasingly use Reinforcement Learning (RL) for post-pretraining, such as RL with Verifiable Rewards (RLVR) for objective tasks and RL from Human Feedback (RLHF) for subjective tasks. However, RLHF incurs high costs and potential reward-polic...
[ "Ziyu Liu", "Yuhang Zang", "Shengyuan Ding", "Yuhang Cao", "Xiaoyi Dong", "Haodong Duan", "Dahua Lin", "Jiaqi Wang" ]
[ "cs.CV", "cs.LG" ]
[ "Computer Science" ]
2025-09-26T00:00:00
https://arxiv.org/abs/2509.22624
https://arxiv.org/pdf/2509.22624v1
2509.22624
10.48550/arXiv.2509.22624
7
0
true
https://github.com/InternLM/Spark
arXiv.org
0.4852
d131ab1547596287013b6a4eab582f3b9a4045e49e21e5bb8284225ae0393785
[ "arxiv", "semantic_scholar" ]
GRPO is Secretly a Process Reward Model
Process reward models (PRMs) allow for fine-grained credit assignment in reinforcement learning (RL), and seemingly contrast with outcome reward models (ORMs), which assign a single reward to an entire trajectory. However, we provide theoretical proof in this work that the Group Relative Policy Optimization (GRPO) RL a...
[ "Michael Sullivan", "Alexander Koller" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-09-25T00:00:00
https://arxiv.org/abs/2509.21154
https://arxiv.org/pdf/2509.21154v4
2509.21154
10.48550/arXiv.2509.21154
4
0
false
null
arXiv.org
0.3128
01c59db9ecbe6fa44caf0f4469c0124517a18284f1618a8f5f2798d6a2f86fd3
[ "arxiv", "semantic_scholar" ]
What Fundamental Structure in Reward Functions Enables Efficient Sparse-Reward Learning?
Sparse-reward reinforcement learning (RL) remains fundamentally hard: without structure, any agent needs $Ω(|\mathcal{S}||\mathcal{A}|/p)$ samples to recover rewards. We introduce Policy-Aware Matrix Completion (PAMC) as a first concrete step toward a structural reward learning framework. Our key idea is to exploit app...
[ "Ibne Farabi Shihab", "Sanjeda Akter", "Anuj Sharma" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-09-04T00:00:00
https://arxiv.org/abs/2509.03790
https://arxiv.org/pdf/2509.03790v2
2509.03790
10.48550/arXiv.2509.03790
1
0
false
null
arXiv.org
0.2888
3abdc739bc0b7070e2158b7650f636230fb2a86a0307fb2fb4b9ad7f12889c0a
[ "arxiv", "semantic_scholar" ]
SharedRep-RLHF: A Shared Representation Approach to RLHF with Diverse Preferences
Uniform-reward reinforcement learning from human feedback (RLHF), which trains a single reward model to represent the preferences of all annotators, fails to capture the diversity of opinions across sub-populations, inadvertently favoring dominant groups. The state-of-the-art, MaxMin-RLHF, addresses this by learning gr...
[ "Arpan Mukherjee", "Marcello Bullo", "Deniz Gündüz" ]
[ "cs.LG", "stat.ML" ]
[ "Computer Science", "Mathematics" ]
2025-09-03T00:00:00
https://arxiv.org/abs/2509.03672
https://arxiv.org/pdf/2509.03672v1
2509.03672
10.48550/arXiv.2509.03672
0
0
false
null
AAAI Conference on Artificial Intelligence
0.2876
e5702c497b56606ff5504820dcad2171002f137dc8b7ee5166f9080f182eebd5
[ "arxiv", "semantic_scholar" ]
GRAM-R$^2$: Self-Training Generative Foundation Reward Models for Reward Reasoning
Significant progress in reward modeling over recent years has been driven by a paradigm shift from task-specific designs towards generalist reward models. Despite this trend, developing effective reward models remains a fundamental challenge: the heavy reliance on large-scale labeled preference data. Pre-training on ab...
[ "Chenglong Wang", "Yongyu Mu", "Hang Zhou", "Yifu Huo", "Ziming Zhu", "Jiali Zeng", "Murun Yang", "Bei Li", "Xiaoyang Hao", "Chunliang Zhang", "Fandong Meng", "Jingbo Zhu", "Tong Xiao" ]
[ "cs.CL", "cs.LG" ]
[ "Computer Science" ]
2025-09-02T00:00:00
https://arxiv.org/abs/2509.02492
https://arxiv.org/pdf/2509.02492v3
2509.02492
10.48550/arXiv.2509.02492
5
1
false
null
AAAI Conference on Artificial Intelligence
0.2865
31f06e60ac7f248fc0e4e2c2f898fda43c13977dd2bc09119a679df15e8d2556
[ "arxiv", "semantic_scholar" ]
Counterfactual Reward Model Training for Bias Mitigation in Multimodal Reinforcement Learning
In reinforcement learning with human feedback (RLHF), reward models can efficiently learn and amplify latent biases within multimodal datasets, which can lead to imperfect policy optimization through flawed reward signals and decreased fairness. Bias mitigation studies have often applied passive constraints, which can ...
[ "Sheryl Mathew", "N Harshit" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-08-27T00:00:00
https://arxiv.org/abs/2508.19567
https://arxiv.org/pdf/2508.19567v1
2508.19567
10.48550/arXiv.2508.19567
0
0
false
null
arXiv.org
0.2796
c9b19869f8cfdbe093f3650876eb40396c1fc8bced8787cd323eafa178fbf66b
[ "arxiv", "semantic_scholar" ]
Fusing Rewards and Preferences in Reinforcement Learning
We present Dual-Feedback Actor (DFA), a reinforcement learning algorithm that fuses both individual rewards and pairwise preferences (if available) into a single update rule. DFA uses the policy's log-probabilities directly to model the preference probability, avoiding a separate reward-modeling step. Preferences can b...
[ "Sadegh Khorasani", "Saber Salehkaleybar", "Negar Kiyavash", "Matthias Grossglauser" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-08-15T00:00:00
https://arxiv.org/abs/2508.11363
https://arxiv.org/pdf/2508.11363v1
2508.11363
10.48550/arXiv.2508.11363
2
0
false
null
arXiv.org
0.2658
663b5ef7858511b7bc7a8fdaa54efef7ef36cde7c7cbdfa663908cc941b4bb2a
[ "arxiv", "semantic_scholar" ]
Interpretable Reward Model via Sparse Autoencoder
Large language models (LLMs) have been widely deployed across numerous fields. Reinforcement Learning from Human Feedback (RLHF) leverages reward models (RMs) as proxies for human preferences to align LLM behaviors with human values, making the accuracy, reliability, and interpretability of RMs critical for effective a...
[ "Shuyi Zhang", "Wei Shi", "Sihang Li", "Jiayi Liao", "Hengxing Cai", "Xiang Wang" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-08-12T00:00:00
https://arxiv.org/abs/2508.08746
https://arxiv.org/pdf/2508.08746v5
2508.08746
10.48550/arXiv.2508.08746
14
1
true
https://github.com/schrieffer-z/sarm
AAAI Conference on Artificial Intelligence
0.4055
b799dd705b4715541bc6e7d2cf3a69a580dec74f033ad0c213b6877e76fb430a
[ "arxiv", "semantic_scholar" ]
ReCode: Reinforcing Code Generation with Reasoning-Process Rewards
In practice, rigorous reasoning is often a key driver of correct code, while Reinforcement Learning (RL) for code generation often neglects optimizing reasoning quality. Bringing process-level supervision into RL is appealing, but it faces two challenges. First, training reliable reward models to assess reasoning quali...
[ "Lishui Fan", "Yu Zhang", "Mouxiang Chen", "Zhongxin Liu" ]
[ "cs.SE", "cs.AI", "cs.CL", "cs.LG" ]
[ "Computer Science" ]
2025-08-07T00:00:00
https://arxiv.org/abs/2508.05170
https://arxiv.org/pdf/2508.05170v3
2508.05170
null
21
3
false
null
null
0.3356
79c008908f6d3ea0df9c8d715d23a747f4a412f187eccbe80eee972cdeb29bf4
[ "arxiv", "semantic_scholar" ]
Difficulty-Based Preference Data Selection by DPO Implicit Reward Gap
Aligning large language models (LLMs) with human preferences is a critical challenge in AI research. While methods like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) are widely used, they often rely on large, costly preference datasets. The current work lacks methods for hig...
[ "Xuan Qi", "Rongwu Xu", "Zhijing Jin" ]
[ "cs.CL", "cs.AI", "cs.LG" ]
[ "Computer Science" ]
2025-08-06T00:00:00
https://arxiv.org/abs/2508.04149
https://arxiv.org/pdf/2508.04149v2
2508.04149
10.48550/arXiv.2508.04149
7
0
true
https://github.com/Difficulty-Based-Preference-Data-Select/Difficulty-Based-Preference-Data-Select
arXiv.org
0.3949
97da6da7b3c1d7f1733121c49ef497c79a0355e043107cc37e555d528e7c87fd
[ "arxiv", "semantic_scholar" ]
Inferring Reward Machines and Transition Machines from Partially Observable Markov Decision Processes
Partially Observable Markov Decision Processes (POMDPs) are fundamental to many real-world applications. Although reinforcement learning (RL) has shown success in fully observable domains, learning policies from traces in partially observable environments remains challenging due to non-Markovian observations. Inferring...
[ "Yuly Wu", "Jiamou Liu", "Libo Zhang" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-08-03T00:00:00
https://arxiv.org/abs/2508.01947
https://arxiv.org/pdf/2508.01947v1
2508.01947
10.48550/arXiv.2508.01947
0
0
true
https://github.com/sousoura/Inferring-Reward-Machines-and-Transition-Machines-from-POMDP.git
arXiv.org
0.3896
7491d5683c99bfeefe018fabf618c77bffcbc329edfcb0acaade8612acea86df
[ "arxiv", "semantic_scholar" ]
The Bidirectional Process Reward Model
Process Reward Models (PRMs), which assign fine-grained scores to intermediate reasoning steps within a solution trajectory, have emerged as a promising approach to enhance the reasoning quality of Large Language Models (LLMs). However, most existing PRMs rely on a unidirectional left-to-right (L2R) evaluation scheme, ...
[ "Lingyin Zhang", "Jun Gao", "Xiaoxue Ren", "Ziqiang Cao" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-08-03T00:00:00
https://arxiv.org/abs/2508.01682
https://arxiv.org/pdf/2508.01682v2
2508.01682
10.48550/arXiv.2508.01682
2
0
false
null
arXiv.org
0.2521
9587517d1a250d78a2ea7100894b5ca504acac6037a31a55e63166552dc8afac
[ "arxiv", "semantic_scholar" ]
Policy Learning from Large Vision-Language Model Feedback without Reward Modeling
Offline reinforcement learning (RL) provides a powerful framework for training robotic agents using pre-collected, suboptimal datasets, eliminating the need for costly, time-consuming, and potentially hazardous online interactions. This is particularly useful in safety-critical real-world applications, where online dat...
[ "Tung M. Luu", "Donghoon Lee", "Younghwan Lee", "Chang D. Yoo" ]
[ "cs.LG", "cs.RO" ]
[ "Computer Science" ]
2025-07-31T00:00:00
https://arxiv.org/abs/2507.23391
https://arxiv.org/pdf/2507.23391v1
2507.23391
10.1109/IROS60139.2025.11246902
6
0
false
null
IEEE/RJS International Conference on Intelligent RObots and Systems
0.2486
fa115d47a00a5a7d7933f94e8466bc657c98c35867aa7921e91d0fdaaf043be5
[ "arxiv", "semantic_scholar" ]
Multimodal LLMs as Customized Reward Models for Text-to-Image Generation
We introduce LLaVA-Reward, an efficient reward model designed to automatically evaluate text-to-image (T2I) generations across multiple perspectives, leveraging pretrained multimodal large language models (MLLMs). Existing MLLM-based approaches require instruction-following data for supervised fine-tuning and evaluate ...
[ "Shijie Zhou", "Ruiyi Zhang", "Huaisheng Zhu", "Branislav Kveton", "Yufan Zhou", "Jiuxiang Gu", "Jian Chen", "Changyou Chen" ]
[ "cs.CV", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-07-28T00:00:00
https://arxiv.org/abs/2507.21391
https://arxiv.org/pdf/2507.21391v2
2507.21391
10.1109/ICCV51701.2025.01826
9
1
true
https://github.com/sjz5202/LLaVA-Reward
IEEE International Conference on Computer Vision
0.379
138ff77d0bcb41dc48703732472de43dd9ea35b6c8f28d78370fd9bf23d32cae
[ "arxiv", "semantic_scholar" ]
Dynamic and Generalizable Process Reward Modeling
Process Reward Models (PRMs) are crucial for guiding Large Language Models (LLMs) in complex scenarios by providing dense reward signals. However, existing PRMs primarily rely on heuristic approaches, which struggle with cross-domain generalization. While LLM-as-judge has been proposed to provide generalized rewards, c...
[ "Zhangyue Yin", "Qiushi Sun", "Zhiyuan Zeng", "Qinyuan Cheng", "Xipeng Qiu", "Xuanjing Huang" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-07-23T00:00:00
https://arxiv.org/abs/2507.17849
https://arxiv.org/pdf/2507.17849v1
2507.17849
10.18653/v1/2025.acl-long.212
16
1
false
null
Annual Meeting of the Association for Computational Linguistics
0.3076
2b71b255fd249af8aec4039afdd3d9827364095cd0504aa3b0f21f45a32a09bd
[ "arxiv", "semantic_scholar" ]
Off-Policy Corrected Reward Modeling for Reinforcement Learning from Human Feedback
Reinforcement Learning from Human Feedback (RLHF) allows us to train models, such as language models (LMs), to follow complex human preferences. In RLHF for LMs, we first train an LM using supervised fine-tuning, sample pairs of responses, obtain human feedback, and use the resulting data to train a reward model (RM). ...
[ "Johannes Ackermann", "Takashi Ishida", "Masashi Sugiyama" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-07-21T00:00:00
https://arxiv.org/abs/2507.15507
https://arxiv.org/pdf/2507.15507v1
2507.15507
10.48550/arXiv.2507.15507
5
0
true
https://github.com/JohannesAck/OffPolicyCorrectedRewardModeling
arXiv.org
0.3666
e68027ef6d0bcf43e1562e068b88d7404d53e4e19c5a178ab7c3ab4970c4867c
[ "arxiv", "semantic_scholar" ]
Tiny Reward Models
Large decoder-based language models have become the dominant architecture for reward modeling in reinforcement learning from human feedback (RLHF). However, as reward models are increasingly deployed in test-time strategies, their inference costs become a growing concern. We present TinyRM, a family of small, bidirecti...
[ "Sarah Pan" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-07-14T00:00:00
https://arxiv.org/abs/2507.09973
https://arxiv.org/pdf/2507.09973v1
2507.09973
10.48550/arXiv.2507.09973
0
0
false
null
arXiv.org
0.2292
bbbf858be2ae361da6e2a4fb1bfe371efe19ee7defcdf3a280189dd5d35224f7
[ "arxiv", "semantic_scholar" ]
Recursive Reward Aggregation
In reinforcement learning (RL), aligning agent behavior with specific objectives typically requires careful design of the reward function, which can be challenging when the desired objectives are complex. In this work, we propose an alternative approach for flexible behavior alignment that eliminates the need to modify...
[ "Yuting Tang", "Yivan Zhang", "Johannes Ackermann", "Yu-Jie Zhang", "Soichiro Nishimori", "Masashi Sugiyama" ]
[ "cs.LG", "math.CT" ]
[ "Computer Science", "Mathematics" ]
2025-07-11T00:00:00
https://arxiv.org/abs/2507.08537
https://arxiv.org/pdf/2507.08537v2
2507.08537
10.48550/arXiv.2507.08537
3
0
false
null
arXiv.org
0.2257
55d62268686e902f42d2faf95ac5361cec21cd6eb8a73ecdd214f22d48014d3e
[ "arxiv", "semantic_scholar" ]
Bradley-Terry and Multi-Objective Reward Modeling Are Complementary
Reward models trained on human preference data have demonstrated strong effectiveness in aligning Large Language Models (LLMs) with human intent under the framework of Reinforcement Learning from Human Feedback (RLHF). However, RLHF remains vulnerable to reward hacking, where the policy exploits imperfections in the re...
[ "Zhiwei Zhang", "Hui Liu", "Xiaomin Li", "Zhenwei Dai", "Jingying Zeng", "Fali Wang", "Minhua Lin", "Ramraj Chandradevan", "Zhen Li", "Chen Luo", "Xianfeng Tang", "Qi He", "Suhang Wang" ]
[ "cs.LG", "cs.CL" ]
[ "Computer Science" ]
2025-07-10T00:00:00
https://arxiv.org/abs/2507.07375
https://arxiv.org/pdf/2507.07375v1
2507.07375
10.48550/arXiv.2507.07375
7
0
false
null
arXiv.org
0.2258
9b3ac64ce71b6b84af6b5c71c4334a3aa83a892c2371c83980a89c413e268135
[ "arxiv", "semantic_scholar" ]
Teach a Reward Model to Correct Itself: Reward Guided Adversarial Failure Discovery for Robust Reward Modeling
Reward modeling (RM), which captures human preferences to align large language models (LLMs), is increasingly employed in tasks such as model finetuning, response filtering, and ranking. However, due to the inherent complexity of human preferences and the limited coverage of available datasets, reward models often fail...
[ "Pankayaraj Pathmanathan", "Furong Huang" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-07-08T00:00:00
https://arxiv.org/abs/2507.06419
https://arxiv.org/pdf/2507.06419v3
2507.06419
null
2
0
false
null
ACL 2026 Main Conference [Oral]
0.2223
d07a0744583b2b3a36917654c542f44f5d6e8941a21b0c911971da2fc2100c84
[ "arxiv", "semantic_scholar" ]
Interpretable Reward Modeling with Active Concept Bottlenecks
We introduce Concept Bottleneck Reward Models (CB-RM), a reward modeling framework that enables interpretable preference learning through selective concept annotation. Unlike standard RLHF methods that rely on opaque reward functions, CB-RM decomposes reward prediction into human-interpretable concepts. To make this fr...
[ "Sonia Laguna", "Katarzyna Kobalczyk", "Julia E. Vogt", "Mihaela Van der Schaar" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-07-07T00:00:00
https://arxiv.org/abs/2507.04695
https://arxiv.org/pdf/2507.04695v2
2507.04695
10.48550/arXiv.2507.04695
1
0
false
null
arXiv.org
0.2211
1643c07009d41a7ebac6763a7c0d6743e5c81eff96f3d28994761bc055c04600
[ "arxiv", "semantic_scholar" ]
Pre-Trained Policy Discriminators are General Reward Models
We offer a novel perspective on reward modeling by formulating it as a policy discriminator, which quantifies the difference between two policies to generate a reward signal, guiding the training policy towards a target policy with desired behaviors. Based on this conceptual insight, we propose a scalable pre-training ...
[ "Shihan Dou", "Shichun Liu", "Yuming Yang", "Yicheng Zou", "Yunhua Zhou", "Shuhao Xing", "Chenhao Huang", "Qiming Ge", "Demin Song", "Haijun Lv", "Songyang Gao", "Chengqi Lv", "Enyu Zhou", "Honglin Guo", "Zhiheng Xi", "Wenwei Zhang", "Qipeng Guo", "Qi Zhang", "Xipeng Qiu", "Xua...
[ "cs.CL", "cs.LG" ]
[ "Computer Science" ]
2025-07-07T00:00:00
https://arxiv.org/abs/2507.05197
https://arxiv.org/pdf/2507.05197v2
2507.05197
10.48550/arXiv.2507.05197
12
3
false
null
arXiv.org
0.301
1d61232ac9a584be3218c1e8dddc1e3e6ab9c921fd9db0785acb3c64118334e1
[ "arxiv", "semantic_scholar" ]
ARF-RLHF: Adaptive Reward-Following for RLHF through Emotion-Driven Self-Supervision and Trace-Biased Dynamic Optimization
Current RLHF methods such as PPO and DPO typically reduce human preferences to binary labels, which are costly to obtain and too coarse to reflect individual variation. We observe that expressions of satisfaction and dissatisfaction follow stable linguistic patterns across users, indicating that more informative superv...
[ "YuXuan Zhang" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-07-03T00:00:00
https://arxiv.org/abs/2507.03069
https://arxiv.org/pdf/2507.03069v3
2507.03069
10.48550/arXiv.2507.03069
0
0
false
null
arXiv.org
0.2166
1be2e97a46581870f7e0c192880e517f768ec40d59f45e67cd9feaad0058ddee
[ "arxiv", "semantic_scholar" ]
Uncertainty-aware Reward Design Process
Designing effective reward functions is a cornerstone of reinforcement learning (RL), yet it remains a challenging process due to the inefficiencies and inconsistencies inherent in conventional reward engineering methodologies. Recent advances have explored leveraging large language models (LLMs) to automate reward fun...
[ "Yang Yang", "Xiaolu Zhou", "Bosong Ding", "Miao Xin" ]
[ "cs.LG", "cs.RO" ]
[ "Computer Science" ]
2025-07-03T00:00:00
https://arxiv.org/abs/2507.02256
https://arxiv.org/pdf/2507.02256v1
2507.02256
10.48550/arXiv.2507.02256
0
0
false
null
null
0.1378
952e485cff5ec38f2a0471bb6aa7c33a69ba9276c49e56db4a5d9aa9e3925c68
[ "arxiv", "semantic_scholar" ]
Skywork-Reward-V2: Scaling Preference Data Curation via Human-AI Synergy
Despite the critical role of reward models (RMs) in Reinforcement Learning from Human Feedback (RLHF), current state-of-the-art open RMs perform poorly on most existing evaluation benchmarks, failing to capture nuanced human preferences. We hypothesize that this brittleness stems primarily from limitations in preferenc...
[ "Chris Yuhao Liu", "Liang Zeng", "Yuzhen Xiao", "Jujie He", "Jiacai Liu", "Chaojie Wang", "Rui Yan", "Wei Shen", "Fuxiang Zhang", "Jiacheng Xu", "Yang Liu", "Yahui Zhou" ]
[ "cs.CL", "cs.AI", "cs.LG" ]
[ "Computer Science" ]
2025-07-02T00:00:00
https://arxiv.org/abs/2507.01352
https://arxiv.org/pdf/2507.01352v3
2507.01352
10.48550/arXiv.2507.01352
154
17
false
null
arXiv.org
0.6276
8edd344b927941def2ef488153ca17681c0be3af8228070e2686718ae436371d
[ "arxiv", "semantic_scholar" ]
Activation Reward Models for Few-Shot Model Alignment
Aligning Large Language Models (LLMs) and Large Multimodal Models (LMMs) to human preferences is a central challenge in improving the quality of the models' generative outputs for real-world applications. A common approach is to use reward modeling to encode preferences, enabling alignment via post-training using reinf...
[ "Tianning Chai", "Chancharik Mitra", "Brandon Huang", "Gautam Rajendrakumar Gare", "Zhiqiu Lin", "Assaf Arbelle", "Leonid Karlinsky", "Rogerio Feris", "Trevor Darrell", "Deva Ramanan", "Roei Herzig" ]
[ "cs.CV", "cs.LG" ]
[ "Computer Science" ]
2025-07-02T00:00:00
https://arxiv.org/abs/2507.01368
https://arxiv.org/pdf/2507.01368v1
2507.01368
10.48550/arXiv.2507.01368
2
0
false
null
arXiv.org
0.2154
cd095d3ab3942adbfdc57a73f504426c4ea61b26a96ac5972080c23a74fe32e5
[ "arxiv", "semantic_scholar" ]
Residual Reward Models for Preference-based Reinforcement Learning
Preference-based Reinforcement Learning (PbRL) provides a way to learn high-performance policies in environments where the reward signal is hard to specify, avoiding heuristic and time-consuming reward design. However, PbRL can suffer from slow convergence speed since it requires training in a reward model. Prior work ...
[ "Chenyang Cao", "Miguel Rogel-García", "Mohamed Nabail", "Xueqian Wang", "Nicholas Rhinehart" ]
[ "cs.LG", "cs.AI", "cs.RO" ]
[ "Computer Science" ]
2025-07-01T00:00:00
https://arxiv.org/abs/2507.00611
https://arxiv.org/pdf/2507.00611v1
2507.00611
10.48550/arXiv.2507.00611
2
1
false
null
arXiv.org
0.2143
9b388b7ecad4064a32451de556f4482595179c1871ab5120dcd76c66dcb51cea
[ "arxiv", "semantic_scholar" ]
AutoRule: Reasoning Chain-of-thought Extracted Rule-based Rewards Improve Preference Learning
Rule-based rewards offer a promising strategy for improving reinforcement learning from human feedback (RLHF), but current approaches often rely on manual rule engineering. We present AutoRule, a fully automated method for extracting rules from preference feedback and formulating them into rule-based rewards. AutoRule ...
[ "Tevin Wang", "Chenyan Xiong" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-06-18T00:00:00
https://arxiv.org/abs/2506.15651
https://arxiv.org/pdf/2506.15651v1
2506.15651
10.48550/arXiv.2506.15651
9
2
true
https://github.com/cxcscmu/AutoRule
arXiv.org
0.3081
4645bdf06984e503f6d43074c19d8d3ac9dc59d6c1e0c26162c454d7aff85b06
[ "arxiv", "semantic_scholar" ]
Reward Models in Deep Reinforcement Learning: A Survey
In reinforcement learning (RL), agents continually interact with the environment and use the feedback to refine their behavior. To guide policy optimization, reward models are introduced as proxies of the desired objectives, such that when the agent maximizes the accumulated reward, it also fulfills the task designer's...
[ "Rui Yu", "Shenghua Wan", "Yucen Wang", "Chen-Xiao Gao", "Le Gan", "Zongzhang Zhang", "De-Chuan Zhan" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-06-18T00:00:00
https://arxiv.org/abs/2506.15421
https://arxiv.org/pdf/2506.15421v1
2506.15421
10.48550/arXiv.2506.15421
28
1
false
null
International Joint Conference on Artificial Intelligence
0.3656
58d0bedfbae10d42cfd2510ce7effd7659cf04ad02b8ac104fa07687495c49ab
[ "arxiv", "semantic_scholar" ]
A General Framework for Off-Policy Learning with Partially-Observed Reward
Off-policy learning (OPL) in contextual bandits aims to learn a decision-making policy that maximizes the target rewards by using only historical interaction data collected under previously developed policies. Unfortunately, when rewards are only partially observed, the effectiveness of OPL degrades severely. Well-know...
[ "Rikiya Takehi", "Masahiro Asami", "Kosuke Kawakami", "Yuta Saito" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-06-17T00:00:00
https://arxiv.org/abs/2506.14439
https://arxiv.org/pdf/2506.14439v1
2506.14439
10.48550/arXiv.2506.14439
1
0
false
null
International Conference on Learning Representations
0.1982
da2eee8ece386cc92ee841673ef9c046a8be31977e6f7e91d756ac9356c0df81
[ "arxiv", "semantic_scholar" ]
TGDPO: Harnessing Token-Level Reward Guidance for Enhancing Direct Preference Optimization
Recent advancements in reinforcement learning from human feedback have shown that utilizing fine-grained token-level reward models can substantially enhance the performance of Proximal Policy Optimization (PPO) in aligning large language models. However, it is challenging to leverage such token-level reward as guidance...
[ "Mingkang Zhu", "Xi Chen", "Zhongdao Wang", "Bei Yu", "Hengshuang Zhao", "Jiaya Jia" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-06-17T00:00:00
https://arxiv.org/abs/2506.14574
https://arxiv.org/pdf/2506.14574v1
2506.14574
10.48550/arXiv.2506.14574
6
0
true
https://github.com/dvlab-research/TGDPO
International Conference on Machine Learning
0.3064
155b7e74e3282ba27ef05598bed14b68c7c23386d62407e3e91574eed0ecca66
[ "arxiv", "semantic_scholar" ]
Learning to Explore in Diverse Reward Settings via Temporal-Difference-Error Maximization
Numerous heuristics and advanced approaches have been proposed for exploration in different settings for deep reinforcement learning. Noise-based exploration generally fares well with dense-shaped rewards and bonus-based exploration with sparse rewards. However, these methods usually require additional tuning to deal w...
[ "Sebastian Griesbach", "Carlo D'Eramo" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-06-16T00:00:00
https://arxiv.org/abs/2506.13345
https://arxiv.org/pdf/2506.13345v2
2506.13345
10.48550/arXiv.2506.13345
3
1
false
null
arXiv.org
0.1971
3c6cdc1b7059053ba080edb0b2ad867579667988676be916a089956c890c7eba
[ "arxiv", "semantic_scholar" ]
Similarity as Reward Alignment: Robust and Versatile Preference-based Reinforcement Learning
Preference-based Reinforcement Learning (PbRL) entails a variety of approaches for aligning models with human intent to alleviate the burden of reward engineering. However, most previous PbRL work has not investigated the robustness to labeler errors, inevitable with labelers who are non-experts or operate under time c...
[ "Sara Rajaram", "R. James Cotton", "Fabian H. Sinz" ]
[ "cs.LG", "cs.AI", "stat.ML" ]
[ "Computer Science", "Mathematics" ]
2025-06-14T00:00:00
https://arxiv.org/abs/2506.12529
https://arxiv.org/pdf/2506.12529v1
2506.12529
10.48550/arXiv.2506.12529
2
0
false
null
arXiv.org
0.1948
3a51464484b8fb38ab68447771f1f83aa81d46ba950875d5a2f98a51c4659d40
[ "arxiv", "semantic_scholar" ]
Multi-Task Reward Learning from Human Ratings
Reinforcement learning from human feedback (RLHF) has become a key factor in aligning model behavior with users' goals. However, while humans integrate multiple strategies when making decisions, current RLHF approaches often simplify this process by modeling human reasoning through isolated tasks such as classification...
[ "Mingkang Wu", "Devin White", "Evelyn Rose", "Vernon Lawhern", "Nicholas R Waytowich", "Yongcan Cao" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-06-10T00:00:00
https://arxiv.org/abs/2506.09183
https://arxiv.org/pdf/2506.09183v2
2506.09183
10.48550/arXiv.2506.09183
1
1
false
null
arXiv.org
0.1902
e44c85e3e81318c8d3ca27c07416fc8688ff6c361ea1c45bc49bb145cbfcd92d
[ "arxiv", "semantic_scholar" ]
Intra-Trajectory Consistency for Reward Modeling
Reward models are critical for improving large language models (LLMs), particularly in reinforcement learning from human feedback (RLHF) or inference-time verification. Current reward modeling typically relies on scores of overall responses to learn the outcome rewards for the responses. However, since the response-lev...
[ "Chaoyang Zhou", "Shunyu Liu", "Zengmao Wang", "Di Wang", "Rong-Cheng Tu", "Bo Du", "Dacheng Tao" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-06-10T00:00:00
https://arxiv.org/abs/2506.09096
https://arxiv.org/pdf/2506.09096v3
2506.09096
10.48550/arXiv.2506.09096
0
0
true
https://github.com/chaoyang101/ICRM
arXiv.org
0.294
acdb1c67eb8e501130bb9adc082d54b00cbdf3de0f74130edca567b6494a6eba
[ "arxiv", "semantic_scholar" ]
GFRIEND: Generative Few-shot Reward Inference through EfficieNt DPO
The ability to train high-performing reward models with few-shot data is critical for enhancing the efficiency and scalability of Reinforcement Learning from Human Feedback (RLHF). We propose a data augmentation and expansion framework that enables generative reward models trained on small datasets to achieve comparabl...
[ "Yiyang Zhao", "Huiyu Bai", "Xuejiao Zhao" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-06-10T00:00:00
https://arxiv.org/abs/2506.08965
https://arxiv.org/pdf/2506.08965v1
2506.08965
10.48550/arXiv.2506.08965
0
0
false
null
arXiv.org
0.1902
e42d4253c2f889bbbc4d97782b8fd8f6078211cb41b1638e0d158f28d5dca451
[ "arxiv", "semantic_scholar" ]
Explicit Preference Optimization: No Need for an Implicit Reward Model
The generated responses of large language models (LLMs) are often fine-tuned to human preferences through a process called reinforcement learning from human feedback (RLHF). As RLHF relies on a challenging training sequence, whereby a separate reward model is independently learned and then later applied to LLM policy u...
[ "Xiangkun Hu", "Lemin Kong", "Tong He", "David Wipf" ]
[ "cs.LG", "stat.ML" ]
[ "Computer Science", "Mathematics" ]
2025-06-09T00:00:00
https://arxiv.org/abs/2506.07492
https://arxiv.org/pdf/2506.07492v1
2506.07492
10.48550/arXiv.2506.07492
5
0
false
null
International Conference on Machine Learning
0.1945
2ad7e7671e1fe9c422089ebc327a17cf009e2ff72aed5940a6eb3417f3ce1d0a
[ "arxiv", "semantic_scholar" ]
AMoPO: Adaptive Multi-objective Preference Optimization without Reward Models and Reference Models
Existing multi-objective preference alignment methods for large language models (LLMs) face limitations: (1) the inability to effectively balance various preference dimensions, and (2) reliance on auxiliary reward/reference models introduces computational complexity. To address these challenges, we propose Adaptive Mul...
[ "Qi Liu", "Jingqing Ruan", "Hao Li", "Haodong Zhao", "Desheng Wang", "Jiansong Chen", "Wan Guanglu", "Xunliang Cai", "Zhi Zheng", "Tong Xu" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-06-08T00:00:00
https://arxiv.org/abs/2506.07165
https://arxiv.org/pdf/2506.07165v1
2506.07165
10.48550/arXiv.2506.07165
8
1
true
https://github.com/Javkonline/AMoPO
Annual Meeting of the Association for Computational Linguistics
0.2904
6a3f5aac88fd55555fee8029ecbbe11f6e1c4bc935d4b9c267cff40932520e37
[ "arxiv", "semantic_scholar" ]
BadReward: Clean-Label Poisoning of Reward Models in Text-to-Image RLHF
Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning text-to-image (T2I) models with human preferences. However, RLHF's feedback mechanism also opens new pathways for adversaries. This paper demonstrates the feasibility of hijacking T2I models by poisoning a small fraction of preference data with n...
[ "Kaiwen Duan", "Hongwei Yao", "Yufei Chen", "Ziyun Li", "Tong Qiao", "Zhan Qin", "Cong Wang" ]
[ "cs.LG", "cs.AI", "cs.CR" ]
[ "Computer Science" ]
2025-06-03T00:00:00
https://arxiv.org/abs/2506.03234
https://arxiv.org/pdf/2506.03234v1
2506.03234
10.48550/arXiv.2506.03234
3
0
false
null
arXiv.org
0.1822
c67feb08b1efbdee3f5245fcd6a488fa38981befd75d780ac97b540ceb4e9200
[ "arxiv", "semantic_scholar" ]
RewardBench 2: Advancing Reward Model Evaluation
Reward models are used throughout the post-training of language models to capture nuanced signals from preference data and provide a training target for optimization across instruction following, reasoning, safety, and more domains. The community has begun establishing best practices for evaluating reward models, from ...
[ "Saumya Malik", "Valentina Pyatkin", "Sander Land", "Jacob Morrison", "Noah A. Smith", "Hannaneh Hajishirzi", "Nathan Lambert" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-06-02T00:00:00
https://arxiv.org/abs/2506.01937
https://arxiv.org/pdf/2506.01937v2
2506.01937
10.48550/arXiv.2506.01937
100
22
false
null
arXiv.org
0.6809
17a35f64c9d38f0ffae023cb4c2a6bcc39d7caec9dbfbca99de868440badd091
[ "arxiv", "semantic_scholar" ]
Accelerating RLHF Training with Reward Variance Increase
Reinforcement learning from human feedback (RLHF) is an essential technique for ensuring that large language models (LLMs) are aligned with human values and preferences during the post-training phase. As an effective RLHF approach, group relative policy optimization (GRPO) has demonstrated success in many LLM-based app...
[ "Zonglin Yang", "Zhexuan Gu", "Houduo Qi", "Yancheng Yuan" ]
[ "cs.LG", "cs.AI", "math.OC" ]
[ "Computer Science", "Mathematics" ]
2025-05-29T00:00:00
https://arxiv.org/abs/2505.23247
https://arxiv.org/pdf/2505.23247v2
2505.23247
10.48550/arXiv.2505.23247
1
0
false
null
arXiv.org
0.1765
3ebf66805ee3bd1222f64bbb6b45f239de68a480cf0a1f63c41585b5d80316c9
[ "arxiv", "semantic_scholar" ]
Towards Reward Fairness in RLHF: From a Resource Allocation Perspective
Rewards serve as proxies for human preferences and play a crucial role in Reinforcement Learning from Human Feedback (RLHF). However, if these rewards are inherently imperfect, exhibiting various biases, they can adversely affect the alignment of large language models (LLMs). In this paper, we collectively define the v...
[ "Sheng Ouyang", "Yulan Hu", "Ge Chen", "Qingyang Li", "Fuzheng Zhang", "Yong Liu" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-05-29T00:00:00
https://arxiv.org/abs/2505.23349
https://arxiv.org/pdf/2505.23349v1
2505.23349
10.18653/v1/2025.acl-long.163
12
2
false
null
Annual Meeting of the Association for Computational Linguistics
0.2785
dcb5997c76565e63cb72d094b02260523faef7fded02f5fc0df61da80e0f17d6
[ "arxiv", "semantic_scholar" ]
Learning a Pessimistic Reward Model in RLHF
This work proposes `PET', a novel pessimistic reward fine-tuning method, to learn a pessimistic reward model robust against reward hacking in offline reinforcement learning from human feedback (RLHF). Traditional reward modeling techniques in RLHF train an imperfect reward model, on which a KL regularization plays a pi...
[ "Yinglun Xu", "Hangoo Kang", "Tarun Suresh", "Yuxuan Wan", "Gagandeep Singh" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-05-26T00:00:00
https://arxiv.org/abs/2505.20556
https://arxiv.org/pdf/2505.20556v1
2505.20556
10.48550/arXiv.2505.20556
5
1
false
null
arXiv.org
0.1945
9080ca3d8309373118e20454a75d9c0b7497a98e4c553d80e2a8551c00b7a024
[ "arxiv", "semantic_scholar" ]
Understanding the Performance Gap in Preference Learning: A Dichotomy of RLHF and DPO
We present a fine-grained theoretical analysis of the performance gap between two-stage reinforcement learning from human feedback~(RLHF) and direct preference optimization~(DPO). Our study decomposes this gap into two sources: the explicit representation gap under exact optimization and the implicit representation gap...
[ "Ruizhe Shi", "Minhak Song", "Runlong Zhou", "Zihan Zhang", "Maryam Fazel", "Simon S. Du" ]
[ "cs.LG", "cs.CL" ]
[ "Computer Science" ]
2025-05-26T00:00:00
https://arxiv.org/abs/2505.19770
https://arxiv.org/pdf/2505.19770v5
2505.19770
10.48550/arXiv.2505.19770
10
1
false
null
arXiv.org
0.2603
61e2106a41124afb7efd442f3ceb89f3cead9caff1455a8988827998e3f53b70
[ "arxiv", "semantic_scholar" ]
MOSLIM:Align with diverse preferences in prompts through reward classification
The multi-objective alignment of Large Language Models (LLMs) is essential for ensuring foundational models conform to diverse human preferences. Current research in this field typically involves either multiple policies or multiple reward models customized for various preferences, or the need to train a preference-spe...
[ "Yu Zhang", "Wanli Jiang", "Zhengyu Yang" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-05-24T00:00:00
https://arxiv.org/abs/2505.20336
https://arxiv.org/pdf/2505.20336v1
2505.20336
10.48550/arXiv.2505.20336
3
0
false
null
arXiv.org
0.1707
d18be7733c0fd29534dea4d41883e67e738dd9f9de135f39b096bbeb3954deb1
[ "arxiv", "semantic_scholar" ]
Reward Model Overoptimisation in Iterated RLHF
Reinforcement learning from human feedback (RLHF) is a widely used method for aligning large language models with human preferences. However, RLHF often suffers from reward model overoptimisation, in which models overfit to the reward function, resulting in non-generalisable policies that exploit the idiosyncrasies and...
[ "Lorenz Wolf", "Robert Kirk", "Mirco Musolesi" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-05-23T00:00:00
https://arxiv.org/abs/2505.18126
https://arxiv.org/pdf/2505.18126v2
2505.18126
10.48550/arXiv.2505.18126
5
1
false
null
arXiv.org
0.1945
d1f29db2306123334fc5447c67bf4c3698cc491dfe17f5f30f05a2c17f5ec6ce
[ "arxiv", "semantic_scholar" ]
Think-RM: Enabling Long-Horizon Reasoning in Generative Reward Models
Reinforcement learning from human feedback (RLHF) has become a powerful post-training paradigm for aligning large language models with human preferences. A core challenge in RLHF is constructing accurate reward signals, where the conventional Bradley-Terry reward models (BT RMs) often suffer from sensitivity to data si...
[ "Ilgee Hong", "Changlong Yu", "Liang Qiu", "Weixiang Yan", "Zhenghao Xu", "Haoming Jiang", "Qingru Zhang", "Qin Lu", "Xin Liu", "Chao Zhang", "Tuo Zhao" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-05-22T00:00:00
https://arxiv.org/abs/2505.16265
https://arxiv.org/pdf/2505.16265v1
2505.16265
10.48550/arXiv.2505.16265
11
3
false
null
arXiv.org
0.301
947b81b4f5e9316484a0d8d5370949384e4fd7316f009f451303b29bc2147a7d
[ "arxiv", "semantic_scholar" ]
Reward Reasoning Model
Reward models play a critical role in guiding large language models toward outputs that align with human expectations. However, an open challenge remains in effectively utilizing test-time compute to enhance reward model performance. In this work, we introduce Reward Reasoning Models (RRMs), which are specifically desi...
[ "Jiaxin Guo", "Zewen Chi", "Li Dong", "Qingxiu Dong", "Xun Wu", "Shaohan Huang", "Furu Wei" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-05-20T00:00:00
https://arxiv.org/abs/2505.14674
https://arxiv.org/pdf/2505.14674v1
2505.14674
10.48550/arXiv.2505.14674
35
4
false
null
arXiv.org
0.3891
319da1303360bd98715a2d3605071061a190aa7792a308950cdf0253f15631e0
[ "arxiv", "semantic_scholar" ]
Bias Fitting to Mitigate Length Bias of Reward Model in RLHF
Reinforcement Learning from Human Feedback relies on reward models to align large language models with human preferences. However, RLHF often suffers from reward hacking, wherein policy learning exploits flaws in the trained reward model to maximize reward scores without genuinely aligning with human preferences. A sig...
[ "Kangwen Zhao", "Jianfeng Cai", "Jinhua Zhu", "Ruopei Sun", "Dongyun Xue", "Wengang Zhou", "Li Li", "Houqiang Li" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-05-19T00:00:00
https://arxiv.org/abs/2505.12843
https://arxiv.org/pdf/2505.12843v1
2505.12843
10.48550/arXiv.2505.12843
7
0
false
null
arXiv.org
0.2258
ec39b36546472882af42521050f29e6acbff0edde3de1206625b34c2044c7ed6
[ "arxiv", "semantic_scholar" ]
Rethinking Reward Model Evaluation Through the Lens of Reward Overoptimization
Reward models (RMs) play a crucial role in reinforcement learning from human feedback (RLHF), aligning model behavior with human preferences. However, existing benchmarks for reward models show a weak correlation with the performance of optimized policies, suggesting that they fail to accurately assess the true capabil...
[ "Sunghwan Kim", "Dongjin Kang", "Taeyoon Kwon", "Hyungjoo Chae", "Dongha Lee", "Jinyoung Yeo" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-05-19T00:00:00
https://arxiv.org/abs/2505.12763
https://arxiv.org/pdf/2505.12763v1
2505.12763
10.48550/arXiv.2505.12763
7
1
false
null
Annual Meeting of the Association for Computational Linguistics
0.2258
bc8c6db98eb94e802239b6901ef44a6fd6be8d3b5500d0260a84699df6e623f7
[ "arxiv", "semantic_scholar" ]
Learning Pareto-Optimal Rewards from Noisy Preferences: A Framework for Multi-Objective Inverse Reinforcement Learning
As generative agents become increasingly capable, alignment of their behavior with complex human values remains a fundamental challenge. Existing approaches often simplify human intent through reduction to a scalar reward, overlooking the multi-faceted nature of human feedback. In this work, we introduce a theoretical ...
[ "Kalyan Cherukuri", "Aarav Lala" ]
[ "cs.LG", "cs.AI", "cs.CG" ]
[ "Computer Science" ]
2025-05-17T00:00:00
https://arxiv.org/abs/2505.11864
https://arxiv.org/pdf/2505.11864v3
2505.11864
10.48550/arXiv.2505.11864
1
0
false
null
arXiv.org
0.1627
f52fd260499c3a19e4e32b879fbf872f179a1a334e51c318c2cfe207e013737e
[ "arxiv", "semantic_scholar" ]
Detecting Prefix Bias in LLM-based Reward Models
Reinforcement Learning with Human Feedback (RLHF) has emerged as a key paradigm for task-specific fine-tuning of language models using human preference data. While numerous publicly available preference datasets provide pairwise comparisons of responses, the potential for biases in the resulting reward models remains u...
[ "Ashwin Kumar", "Yuzi He", "Aram H. Markosyan", "Bobbie Chern", "Imanol Arrieta-Ibarra" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-05-13T00:00:00
https://arxiv.org/abs/2505.13487
https://arxiv.org/pdf/2505.13487v2
2505.13487
10.1145/3715275.3732204
13
1
true
null
Conference on Fairness, Accountability and Transparency
0.2865
9177fd7c623e81899daaa21d3a2583d6809ec941b4f845c3bb7a0433b993d6d4
[ "arxiv", "semantic_scholar" ]
On the Robustness of Reward Models for Language Model Alignment
The Bradley-Terry (BT) model is widely practiced in reward modeling for reinforcement learning with human feedback (RLHF). Despite its effectiveness, reward models (RMs) trained with BT model loss are prone to over-optimization, losing generalizability to unseen input distributions. In this paper, we study the cause of...
[ "Jiwoo Hong", "Noah Lee", "Eunki Kim", "Guijin Son", "Woojin Chung", "Aman Gupta", "Shao Tang", "James Thorne" ]
[ "cs.CL", "cs.AI", "cs.LG" ]
[ "Computer Science" ]
2025-05-12T00:00:00
https://arxiv.org/abs/2505.07271
https://arxiv.org/pdf/2505.07271v1
2505.07271
10.48550/arXiv.2505.07271
11
0
true
https://github.com/LinkedIn-XFACT/RM-Robustness
International Conference on Machine Learning
0.2698
c65e399c70be1f592947a35b249bf97aa5a7c65dea4131bd6bad673fe27f41db
[ "arxiv", "semantic_scholar" ]
Learning Guarantee of Reward Modeling Using Deep Neural Networks
In this work, we study the learning theory of reward modeling with pairwise comparison data using deep neural networks. We establish a novel non-asymptotic regret bound for deep reward estimators in a non-parametric setting, which depends explicitly on the network architecture. Furthermore, to underscore the critical i...
[ "Yuanhang Luo", "Yeheng Ge", "Ruijian Han", "Guohao Shen" ]
[ "stat.ML", "cs.LG" ]
[ "Computer Science", "Mathematics" ]
2025-05-10T00:00:00
https://arxiv.org/abs/2505.06601
https://arxiv.org/pdf/2505.06601v1
2505.06601
10.1145/3770854.3780316
3
0
false
null
Knowledge Discovery and Data Mining
0.1547
666b2c5f57a8c9ecbb18268f706a919ebc5f6e81a213b549b8ea8eb6e6e753ae
[ "arxiv", "semantic_scholar" ]
DMRL: Data- and Model-aware Reward Learning for Data Extraction
Large language models (LLMs) are inherently vulnerable to unintended privacy breaches. Consequently, systematic red-teaming research is essential for developing robust defense mechanisms. However, current data extraction methods suffer from several limitations: (1) rely on dataset duplicates (addressable via deduplicat...
[ "Zhiqiang Wang", "Ruoxi Cheng" ]
[ "cs.LG", "cs.CR" ]
[ "Computer Science" ]
2025-05-07T00:00:00
https://arxiv.org/abs/2505.06284
https://arxiv.org/pdf/2505.06284v1
2505.06284
10.48550/arXiv.2505.06284
0
0
false
null
arXiv.org
0.1513
6e030a5fd3aa9ced9849c76019f2493a53b3f913b3bd6566a1aa72a632a75923
[ "arxiv", "semantic_scholar" ]
Policy-labeled Preference Learning: Is Preference Enough for RLHF?
To design rewards that align with human goals, Reinforcement Learning from Human Feedback (RLHF) has emerged as a prominent technique for learning reward functions from human preferences and optimizing policies via reinforcement learning algorithms. However, existing RLHF methods often misinterpret trajectories as bein...
[ "Taehyun Cho", "Seokhun Ju", "Seungyub Han", "Dohyeong Kim", "Kyungjae Lee", "Jungwoo Lee" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-05-06T00:00:00
https://arxiv.org/abs/2505.06273
https://arxiv.org/pdf/2505.06273v2
2505.06273
10.48550/arXiv.2505.06273
3
0
false
null
International Conference on Machine Learning
0.1505
3c658dc8a481b46e4da551de8411b504d0dc36323c24d2c2612fd751b59162c0
[ "arxiv", "semantic_scholar" ]
Sailing by the Stars: A Survey on Reward Models and Learning Strategies for Learning from Rewards
Recent developments in Large Language Models (LLMs) have shifted from pre-training scaling to post-training and test-time scaling. Across these developments, a key unified paradigm has arisen: Learning from Rewards, where reward signals act as the guiding stars to steer LLM behavior. It has underpinned a wide range of ...
[ "Xiaobao Wu" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-05-05T00:00:00
https://arxiv.org/abs/2505.02686
https://arxiv.org/pdf/2505.02686v2
2505.02686
null
7
0
true
https://github.com/bobxwu/learning-from-rewards-llm-papers
null
0.2258
647e2b573a55bc08bcb6e55eef68b9e7c8d35f3ecc19b634a916b994768c66fd
[ "arxiv", "semantic_scholar" ]
R1-Reward: Training Multimodal Reward Model Through Stable Reinforcement Learning
Multimodal Reward Models (MRMs) play a crucial role in enhancing the performance of Multimodal Large Language Models (MLLMs). While recent advancements have primarily focused on improving the model structure and training data of MRMs, there has been limited exploration into the effectiveness of long-term reasoning capa...
[ "Yi-Fan Zhang", "Xingyu Lu", "Xiao Hu", "Chaoyou Fu", "Bin Wen", "Tianke Zhang", "Changyi Liu", "Kaiyu Jiang", "Kaibing Chen", "Kaiyu Tang", "Haojie Ding", "Jiankang Chen", "Fan Yang", "Zhang Zhang", "Tingting Gao", "Liang Wang" ]
[ "cs.CV", "cs.CL" ]
[ "Computer Science" ]
2025-05-05T00:00:00
https://arxiv.org/abs/2505.02835
https://arxiv.org/pdf/2505.02835v2
2505.02835
10.48550/arXiv.2505.02835
60
5
true
https://github.com/yfzhang114/r1_reward
arXiv.org
0.4463
4112a6e5900033e5067279fa930bb6d6dc68b3b1fd85be38c58b82f149846151
[ "arxiv", "semantic_scholar" ]
RM-R1: Reward Modeling as Reasoning
Reward modeling is essential for aligning large language models with human preferences through reinforcement learning. To provide accurate reward signals, a reward model (RM) should stimulate deep thinking and conduct interpretable reasoning before assigning a score or a judgment. Inspired by recent advances of long ch...
[ "Xiusi Chen", "Gaotang Li", "Ziqi Wang", "Bowen Jin", "Cheng Qian", "Yu Wang", "Hongru Wang", "Yu Zhang", "Denghui Zhang", "Tong Zhang", "Hanghang Tong", "Heng Ji" ]
[ "cs.CL", "cs.AI", "cs.LG" ]
[ "Computer Science" ]
2025-05-05T00:00:00
https://arxiv.org/abs/2505.02387
https://arxiv.org/pdf/2505.02387v4
2505.02387
10.48550/arXiv.2505.02387
129
24
false
null
arXiv.org
0.699
94e00a7af75a5865ec9b5cd109304dbad8efd77b2fd7718dd8e1d48ccfb13ea7
[ "arxiv", "semantic_scholar" ]
CaRL: Learning Scalable Planning Policies with Simple Rewards
We investigate reinforcement learning (RL) for privileged planning in autonomous driving. State-of-the-art approaches for this task are rule-based, but these methods do not scale to the long tail. RL, on the other hand, is scalable and does not suffer from compounding errors like imitation learning. Contemporary RL app...
[ "Bernhard Jaeger", "Daniel Dauner", "Jens Beißwenger", "Simon Gerstenecker", "Kashyap Chitta", "Andreas Geiger" ]
[ "cs.LG", "cs.AI", "cs.RO" ]
[ "Computer Science" ]
2025-04-24T00:00:00
https://arxiv.org/abs/2504.17838
https://arxiv.org/pdf/2504.17838v3
2504.17838
10.48550/arXiv.2504.17838
26
1
false
null
arXiv.org
0.3578
c276d6b2a94d6af40c70bf1aafdd1a083beb61d089c31c57794573fbdd8c0fc6
[ "arxiv", "semantic_scholar" ]
Learning Explainable Dense Reward Shapes via Bayesian Optimization
Current reinforcement learning from human feedback (RLHF) pipelines for large language model (LLM) alignment typically assign scalar rewards to sequences, using the final token as a surrogate indicator for the quality of the entire sequence. However, this leads to sparse feedback and suboptimal token-level credit assig...
[ "Ryan Koo", "Ian Yang", "Vipul Raheja", "Mingyi Hong", "Kwang-Sung Jun", "Dongyeop Kang" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-04-22T00:00:00
https://arxiv.org/abs/2504.16272
https://arxiv.org/pdf/2504.16272v1
2504.16272
10.48550/arXiv.2504.16272
2
0
false
null
arXiv.org
0.1341
be2bf58c15f704843c7932ba5702c4a6fc5c21dcc5ad589fa915e7a49ce2a53e
[ "arxiv", "semantic_scholar" ]
LoRe: Personalizing LLMs via Low-Rank Reward Modeling
Personalizing large language models (LLMs) to accommodate diverse user preferences is essential for enhancing alignment and user satisfaction. Traditional reinforcement learning from human feedback (RLHF) approaches often rely on monolithic value representations, limiting their ability to adapt to individual preference...
[ "Avinandan Bose", "Zhihan Xiong", "Yuejie Chi", "Simon Shaolei Du", "Lin Xiao", "Maryam Fazel" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-04-20T00:00:00
https://arxiv.org/abs/2504.14439
https://arxiv.org/pdf/2504.14439v1
2504.14439
10.48550/arXiv.2504.14439
24
3
false
null
arXiv.org
0.3495
3b218257623054ec67457c9ae710bc0e0449e57d1bc8a6b1c6176b3b2677f41c
[ "arxiv", "semantic_scholar" ]
CHARM: Calibrating Reward Models With Chatbot Arena Scores
Reward models (RMs) play a crucial role in Reinforcement Learning from Human Feedback by serving as proxies for human preferences in aligning large language models. However, they suffer from various biases which could lead to reward hacking. In this paper, we identify a model preference bias in RMs, where they systemat...
[ "Xiao Zhu", "Chenmien Tan", "Pinzhen Chen", "Rico Sennrich", "Huiming Wang", "Yanlin Zhang", "Hanxu Hu" ]
[ "cs.AI", "cs.LG" ]
[ "Computer Science" ]
2025-04-14T00:00:00
https://arxiv.org/abs/2504.10045
https://arxiv.org/pdf/2504.10045v2
2504.10045
10.48550/arXiv.2504.10045
4
0
false
null
arXiv.org
0.1747
23edfa7e4921483e120223b3fb315f50a55dec10bf5387ddd0180d9ff21be781
[ "arxiv", "semantic_scholar" ]
FLoRA: Sample-Efficient Preference-based RL via Low-Rank Style Adaptation of Reward Functions
Preference-based reinforcement learning (PbRL) is a suitable approach for style adaptation of pre-trained robotic behavior: adapting the robot's policy to follow human user preferences while still being able to perform the original task. However, collecting preferences for the adaptation process in robotics is often ch...
[ "Daniel Marta", "Simon Holk", "Miguel Vasco", "Jens Lundell", "Timon Homberger", "Finn Busch", "Olov Andersson", "Danica Kragic", "Iolanda Leite" ]
[ "cs.RO", "cs.LG" ]
[ "Computer Science" ]
2025-04-14T00:00:00
https://arxiv.org/abs/2504.10002
https://arxiv.org/pdf/2504.10002v1
2504.10002
10.1109/ICRA55743.2025.11127633
2
0
false
null
IEEE International Conference on Robotics and Automation
0.1249
7500e5c17e13d3a53adac64ffcdc54de62b1b93e57148722c19896df552ab46e
[ "arxiv", "semantic_scholar" ]
Information-Theoretic Reward Decomposition for Generalizable RLHF
A generalizable reward model is crucial in Reinforcement Learning from Human Feedback (RLHF) as it enables correctly evaluating unseen prompt-response pairs. However, existing reward models lack this ability, as they are typically trained by increasing the reward gap between chosen and rejected responses, while overloo...
[ "Liyuan Mao", "Haoran Xu", "Amy Zhang", "Weinan Zhang", "Chenjia Bai" ]
[ "cs.AI", "cs.CL", "cs.LG" ]
[ "Computer Science" ]
2025-04-08T00:00:00
https://arxiv.org/abs/2504.06020
https://arxiv.org/pdf/2504.06020v2
2504.06020
10.48550/arXiv.2504.06020
4
1
false
null
arXiv.org
0.1747
c89b51520214f8e5f022b8b095e83c12f4b1dfd83457ea6e687d15bb62df98b5
[ "arxiv", "semantic_scholar" ]
Adversarial Training of Reward Models
Reward modeling has emerged as a promising approach for the scalable alignment of language models. However, contemporary reward models (RMs) often lack robustness, awarding high rewards to low-quality, out-of-distribution (OOD) samples. This can lead to reward hacking, where policies exploit unintended shortcuts to max...
[ "Alexander Bukharin", "Haifeng Qian", "Shengyang Sun", "Adithya Renduchintala", "Soumye Singhal", "Zhilin Wang", "Oleksii Kuchaiev", "Olivier Delalleau", "Tuo Zhao" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-04-08T00:00:00
https://arxiv.org/abs/2504.06141
https://arxiv.org/pdf/2504.06141v2
2504.06141
10.48550/arXiv.2504.06141
13
0
false
null
arXiv.org
0.2865
b5f392da3195630890228649bb006c9b911f99958a5359c4217b7a6b7c7f954c
[ "arxiv", "semantic_scholar" ]
A Unified Pairwise Framework for RLHF: Bridging Generative Reward Modeling and Policy Optimization
Reinforcement Learning from Human Feedback (RLHF) has emerged as a important paradigm for aligning large language models (LLMs) with human preferences during post-training. This framework typically involves two stages: first, training a reward model on human preference data, followed by optimizing the language model us...
[ "Wenyuan Xu", "Xiaochen Zuo", "Chao Xin", "Yu Yue", "Lin Yan", "Yonghui Wu" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-04-07T00:00:00
https://arxiv.org/abs/2504.04950
https://arxiv.org/pdf/2504.04950v1
2504.04950
10.48550/arXiv.2504.04950
20
1
false
null
arXiv.org
0.3306
9b31d18a84bc4511fa9b579531827f8d4a998552f229b4fac0cf4b84441a86ee
[ "arxiv", "semantic_scholar" ]
Reward Generation via Large Vision-Language Model in Offline Reinforcement Learning
In offline reinforcement learning (RL), learning from fixed datasets presents a promising solution for domains where real-time interaction with the environment is expensive or risky. However, designing dense reward signals for offline dataset requires significant human effort and domain expertise. Reinforcement learnin...
[ "Younghwan Lee", "Tung M. Luu", "Donghoon Lee", "Chang D. Yoo" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-04-03T00:00:00
https://arxiv.org/abs/2504.08772
https://arxiv.org/pdf/2504.08772v1
2504.08772
10.1109/ICASSP49660.2025.10889042
3
0
false
null
IEEE International Conference on Acoustics, Speech, and Signal Processing
0.1505
45d08ff3375d1b91919777a58de6d5577947e0c1a3c487f43aeb2e460edc838c
[ "arxiv", "semantic_scholar" ]
Probabilistic Uncertain Reward Model
Reinforcement learning from human feedback (RLHF) is a critical technique for training large language models. However, conventional reward models based on the Bradley-Terry model (BTRM) often suffer from overconfidence when faced with inconsistent labels or out-of-distribution samples, leading to reward hacking, where ...
[ "Wangtao Sun", "Xiang Cheng", "Xing Yu", "Haotian Xu", "Zhao Yang", "Shizhu He", "Jun Zhao", "Kang Liu" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-03-28T00:00:00
https://arxiv.org/abs/2503.22480
https://arxiv.org/pdf/2503.22480v6
2503.22480
10.48550/arXiv.2503.22480
1
0
false
null
arXiv.org
0.1054
44ad06fffad9ac7d2af1123140133961da71de87819ad8d2dfde121c6d78f789
[ "arxiv", "semantic_scholar" ]
Reward Design for Reinforcement Learning Agents
Reward functions are central in reinforcement learning (RL), guiding agents towards optimal decision-making. The complexity of RL tasks requires meticulously designed reward functions that effectively drive learning while avoiding unintended consequences. Effective reward design aims to provide signals that accelerate ...
[ "Rati Devidze" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-03-27T00:00:00
https://arxiv.org/abs/2503.21949
https://arxiv.org/pdf/2503.21949v1
2503.21949
10.48550/arXiv.2503.21949
4
0
false
null
arXiv.org
0.1747
d971ae1199ec2e89cd83e2da45084a13c1f73d0ec8fd59b4f47d7cc050bd46a7
[ "arxiv", "semantic_scholar" ]
ViLBench: A Suite for Vision-Language Process Reward Modeling
Process-supervised reward models serve as a fine-grained function that provides detailed step-wise feedback to model responses, facilitating effective selection of reasoning trajectories for complex tasks. Despite its advantages, evaluation on PRMs remains less explored, especially in the multimodal domain. To address ...
[ "Haoqin Tu", "Weitao Feng", "Hardy Chen", "Hui Liu", "Xianfeng Tang", "Cihang Xie" ]
[ "cs.CV", "cs.CL" ]
[ "Computer Science" ]
2025-03-26T00:00:00
https://arxiv.org/abs/2503.20271
https://arxiv.org/pdf/2503.20271v1
2503.20271
10.48550/arXiv.2503.20271
19
2
false
null
Conference on Empirical Methods in Natural Language Processing
0.3253
a80334dc8886133768d587f17fed54544c0b28e9624004e414ee7182900f7e4f
[ "arxiv", "semantic_scholar" ]
Mitigating Reward Over-Optimization in RLHF via Behavior-Supported Regularization
Reinforcement learning from human feedback (RLHF) is an effective method for aligning large language models (LLMs) with human values. However, reward over-optimization remains an open challenge leading to discrepancies between the performance of LLMs under the reward model and the true human objectives. A primary contr...
[ "Juntao Dai", "Taiye Chen", "Yaodong Yang", "Qian Zheng", "Gang Pan" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-03-23T00:00:00
https://arxiv.org/abs/2503.18130
https://arxiv.org/pdf/2503.18130v1
2503.18130
10.48550/arXiv.2503.18130
10
1
false
null
International Conference on Learning Representations
0.2603
ccda508663d7898f7a49fde81ccd7f3b2217d75195d5ac3f34cba0f1a5c3d632
[ "arxiv", "semantic_scholar" ]
Capturing Individual Human Preferences with Reward Features
Reinforcement learning from human feedback usually models preferences using a reward function that does not distinguish between people. We argue that this is unlikely to be a good design choice in contexts with high potential for disagreement, like in the training of large language models. We formalise and analyse the ...
[ "André Barreto", "Vincent Dumoulin", "Yiran Mao", "Mark Rowland", "Nicolas Perez-Nieves", "Bobak Shahriari", "Yann Dauphin", "Doina Precup", "Hugo Larochelle" ]
[ "cs.AI", "cs.LG", "stat.ML" ]
[ "Computer Science", "Mathematics" ]
2025-03-21T00:00:00
https://arxiv.org/abs/2503.17338
https://arxiv.org/pdf/2503.17338v2
2503.17338
10.48550/arXiv.2503.17338
11
1
false
null
arXiv.org
0.2698
f0b43246e955aca73fd63ed015c14d8290c37be52d543ece2081e812b5c92c8d
[ "arxiv", "semantic_scholar" ]
Reward Redistribution via Gaussian Process Likelihood Estimation
In many practical reinforcement learning tasks, feedback is only provided at the end of a long horizon, leading to sparse and delayed rewards. Existing reward redistribution methods typically assume that per-step rewards are independent, thus overlooking interdependencies among state-action pairs. In this paper, we pro...
[ "Minheng Xiao", "Xian Yu" ]
[ "cs.LG", "cs.RO" ]
[ "Computer Science" ]
2025-03-20T00:00:00
https://arxiv.org/abs/2503.17409
https://arxiv.org/pdf/2503.17409v2
2503.17409
null
0
0
false
null
null
0.0612
87061b5f8f9c20d43470a9a2b6a78f22ac05261df71bdf913d95e5feb9b5a99c
[ "arxiv", "semantic_scholar" ]
What Makes a Reward Model a Good Teacher? An Optimization Perspective
The success of Reinforcement Learning from Human Feedback (RLHF) critically depends on the quality of the reward model. However, while this quality is primarily evaluated through accuracy, it remains unclear whether accuracy fully captures what makes a reward model an effective teacher. We address this question from an...
[ "Noam Razin", "Zixuan Wang", "Hubert Strauss", "Stanley Wei", "Jason D. Lee", "Sanjeev Arora" ]
[ "cs.LG", "cs.AI", "cs.CL", "stat.ML" ]
[ "Computer Science", "Mathematics" ]
2025-03-19T00:00:00
https://arxiv.org/abs/2503.15477
https://arxiv.org/pdf/2503.15477v4
2503.15477
10.48550/arXiv.2503.15477
64
8
true
https://github.com/princeton-pli/what-makes-good-rm
arXiv.org
0.4771
ee30f7d9c4c841ffd3657072f5d8157dc8cdb493dd373ea97b4ce6f858e2b3b2
[ "arxiv", "semantic_scholar" ]
Provably Efficient Reward Transfer in Reinforcement Learning with Discrete Markov Decision Processes
In this paper, we propose a new solution to reward adaptation (RA) in reinforcement learning, where the agent adapts to a target reward function based on one or more existing source behaviors learned a priori under the same domain dynamics but different reward functions. While learning the target behavior from scratch ...
[ "Kevin Vora", "Yu Zhang" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-03-17T00:00:00
https://arxiv.org/abs/2503.13414
https://arxiv.org/pdf/2503.13414v3
2503.13414
null
1
0
false
null
null
0.0753
548ee58eb4a0c184cd4d2255620a3b14d571d5f1736dc1810201e41ba1a26995
[ "arxiv", "semantic_scholar" ]
From Demonstrations to Rewards: Alignment Without Explicit Human Preferences
One of the challenges of aligning large models with human preferences lies in both the data requirements and the technical complexities of current approaches. Predominant methods, such as RLHF, involve multiple steps, each demanding distinct types of data, including demonstration data and preference data. In RLHF, huma...
[ "Siliang Zeng", "Yao Liu", "Huzefa Rangwala", "George Karypis", "Mingyi Hong", "Rasool Fakoor" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-03-15T00:00:00
https://arxiv.org/abs/2503.13538
https://arxiv.org/pdf/2503.13538v1
2503.13538
10.48550/arXiv.2503.13538
5
0
false
null
arXiv.org
0.1945
859549dd753d63b73b5a5b798bfe0562a21a2e9b81f54f431907168d25782e88
[ "arxiv", "semantic_scholar" ]
Right Reward Right Time for Federated Learning
Critical learning periods (CLPs) in federated learning (FL) refer to early stages during which low-quality contributions (e.g., sparse training data availability) can permanently impair the performance of the global model owned by the cloud server. However, existing incentive mechanisms typically assume temporal homoge...
[ "Thanh Linh Nguyen", "Dinh Thai Hoang", "Diep N. Nguyen", "Quoc-Viet Pham" ]
[ "cs.LG", "cs.AI", "cs.DC", "cs.GT" ]
[ "Computer Science" ]
2025-03-10T00:00:00
https://arxiv.org/abs/2503.07869
https://arxiv.org/pdf/2503.07869v3
2503.07869
10.48550/arXiv.2503.07869
2
0
false
null
arXiv.org
0.1193
ecf9b9cdee2c1b17d877580efd168cd9dc614f7a372387773a7a071b608f8f5b
[ "arxiv", "semantic_scholar" ]
Language Model Personalization via Reward Factorization
Modern large language models (LLMs) are optimized for human-aligned responses using Reinforcement Learning from Human Feedback (RLHF). However, existing RLHF approaches assume a universal preference model and fail to account for individual user preferences, limiting their effectiveness in personalized applications. We ...
[ "Idan Shenfeld", "Felix Faltings", "Pulkit Agrawal", "Aldo Pacchiano" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-03-08T00:00:00
https://arxiv.org/abs/2503.06358
https://arxiv.org/pdf/2503.06358v1
2503.06358
10.48550/arXiv.2503.06358
23
4
false
null
arXiv.org
0.3495
91da2ec8eff57c25f9754497ab772339ab0dab5858165d4df843389d745955ea
[ "arxiv", "semantic_scholar" ]
Implicit Cross-Lingual Rewarding for Efficient Multilingual Preference Alignment
Direct Preference Optimization (DPO) has become a prominent method for aligning Large Language Models (LLMs) with human preferences. While DPO has enabled significant progress in aligning English LLMs, multilingual preference alignment is hampered by data scarcity. To address this, we propose a novel approach that $\te...
[ "Wen Yang", "Junhong Wu", "Chen Wang", "Chengqing Zong", "Jiajun Zhang" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-03-06T00:00:00
https://arxiv.org/abs/2503.04647
https://arxiv.org/pdf/2503.04647v2
2503.04647
10.48550/arXiv.2503.04647
6
0
true
https://github.com/ZNLP/Implicit-Cross-Lingual-Rewarding
Annual Meeting of the Association for Computational Linguistics
0.2113
04b8859e44d134c56a5b30b9c2a59ea8c10f08797bbde2cd069c5b2210d130e9
[ "arxiv", "semantic_scholar" ]
Diffusion Classifier-Driven Reward for Offline Preference-based Reinforcement Learning
Offline preference-based reinforcement learning (PbRL) mitigates the need for reward definition, aligning with human preferences via preference-driven reward feedback without interacting with the environment. However, trajectory-wise preference labels are difficult to meet the precise learning of step-wise reward, ther...
[ "Teng Pang", "Bingzheng Wang", "Guoqiang Wu", "Yilong Yin" ]
[ "cs.LG" ]
[ "Computer Science" ]
2025-03-03T00:00:00
https://arxiv.org/abs/2503.01143
https://arxiv.org/pdf/2503.01143v3
2503.01143
null
0
0
false
null
null
0.0489
364037cda029563f86122dc266f4dcdb67e2adb4c8b1b9313200ef0c8205cd60
[ "arxiv", "semantic_scholar" ]
Sentence-level Reward Model can Generalize Better for Aligning LLM from Human Preference
Learning reward models from human preference datasets and subsequently optimizing language models via reinforcement learning has emerged as a fundamental paradigm for aligning LLMs with human preferences. The performance of the reward model plays a crucial role in the effectiveness of alignment. Previous reward models ...
[ "Wenjie Qiu", "Yi-Chen Li", "Xuqin Zhang", "Tianyi Zhang", "Yihang Zhang", "Zongzhang Zhang", "Yang Yu" ]
[ "cs.CL", "cs.LG" ]
[ "Computer Science" ]
2025-03-01T00:00:00
https://arxiv.org/abs/2503.04793
https://arxiv.org/pdf/2503.04793v4
2503.04793
10.48550/arXiv.2503.04793
5
0
false
null
arXiv.org
0.1945
7a3f222ec9f4669af8212aee7125b5a05f40b366a397719706fb868c9b729e9c
[ "arxiv", "semantic_scholar" ]
Reward Learning from Multiple Feedback Types
Learning rewards from preference feedback has become an important tool in the alignment of agentic models. Preference-based feedback, often implemented as a binary comparison between multiple completions, is an established method to acquire large-scale human feedback. However, human feedback in other contexts is often ...
[ "Yannick Metz", "András Geiszl", "Raphaël Baur", "Mennatallah El-Assady" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-02-28T00:00:00
https://arxiv.org/abs/2502.21038
https://arxiv.org/pdf/2502.21038v1
2502.21038
10.48550/arXiv.2502.21038
7
0
false
null
International Conference on Learning Representations
0.2258
2db0ec7c2ab39bd180362a9a7937a17b0466d559caa39240ab823ef2d13dca27
[ "arxiv", "semantic_scholar" ]
Reward Shaping to Mitigate Reward Hacking in RLHF
Reinforcement Learning from Human Feedback (RLHF) is essential for aligning large language models (LLMs) with human values. However, RLHF is susceptible to \emph{reward hacking}, where the agent exploits flaws in the reward function rather than learning the intended behavior, thus degrading alignment. Although reward s...
[ "Jiayi Fu", "Xuandong Zhao", "Chengyuan Yao", "Heng Wang", "Qi Han", "Yanghua Xiao" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-02-26T00:00:00
https://arxiv.org/abs/2502.18770
https://arxiv.org/pdf/2502.18770v5
2502.18770
10.48550/arXiv.2502.18770
78
7
true
https://github.com/PorUna-byte/PAR
arXiv.org
0.4744
503deaa5c3620362933f180c29e6c7911be811104e328a301dac5e3411863e3b
[ "arxiv", "semantic_scholar" ]
Can RLHF be More Efficient with Imperfect Reward Models? A Policy Coverage Perspective
Sample efficiency is critical for online Reinforcement Learning from Human Feedback (RLHF). While existing works investigate sample-efficient online exploration strategies, the potential of utilizing misspecified yet relevant reward models to accelerate learning remains underexplored. This paper studies how to transfer...
[ "Jiawei Huang", "Bingcong Li", "Christoph Dann", "Niao He" ]
[ "cs.LG", "cs.AI", "stat.ML" ]
[ "Computer Science", "Mathematics" ]
2025-02-26T00:00:00
https://arxiv.org/abs/2502.19255
https://arxiv.org/pdf/2502.19255v3
2502.19255
10.48550/arXiv.2502.19255
4
0
false
null
International Conference on Machine Learning
0.1747
44c02d9d0d7891d15271536a56ca7bde2f7aaeceedd5a8e104a9fdb348863b60
[ "arxiv", "semantic_scholar" ]
Agentic Reward Modeling: Integrating Human Preferences with Verifiable Correctness Signals for Reliable Reward Systems
Reward models (RMs) are crucial for the training and inference-time scaling up of large language models (LLMs). However, existing reward models primarily focus on human preferences, neglecting verifiable correctness signals which have shown strong potential in training LLMs. In this paper, we propose agentic reward mod...
[ "Hao Peng", "Yunjia Qi", "Xiaozhi Wang", "Zijun Yao", "Bin Xu", "Lei Hou", "Juanzi Li" ]
[ "cs.CL", "cs.AI" ]
[ "Computer Science" ]
2025-02-26T00:00:00
https://arxiv.org/abs/2502.19328
https://arxiv.org/pdf/2502.19328v1
2502.19328
10.48550/arXiv.2502.19328
49
4
true
https://github.com/THU-KEG/Agentic-Reward-Modeling
Annual Meeting of the Association for Computational Linguistics
0.4247
00e6bc7efdcdc603d11974e7e193e026eb6db7fcbdd67a71b436fa380922410b
[ "arxiv", "semantic_scholar" ]
FSPO: Few-Shot Optimization of Synthetic Preferences Personalizes to Real Users
Effective personalization of LLMs is critical for a broad range of user-interfacing applications such as virtual assistants and content curation. Inspired by the strong in-context capabilities of LLMs, we propose few-shot preference optimization (FSPO), an algorithm for LLM personalization that reframes reward modeling...
[ "Anikait Singh", "Sheryl Hsu", "Kyle Hsu", "Eric Mitchell", "Stefano Ermon", "Tatsunori Hashimoto", "Archit Sharma", "Chelsea Finn" ]
[ "cs.LG", "cs.AI", "cs.CL", "cs.HC", "stat.ML" ]
[ "Computer Science", "Mathematics" ]
2025-02-26T00:00:00
https://arxiv.org/abs/2502.19312
https://arxiv.org/pdf/2502.19312v2
2502.19312
null
20
0
false
null
null
0.3306
e4cd76c36faf75ae1c8dc71ea566101507bf7aeb25b96998ccc282fd4695ac5d
[ "arxiv", "semantic_scholar" ]
Larger or Smaller Reward Margins to Select Preferences for Alignment?
Preference learning is critical for aligning large language models (LLMs) with human values, with the quality of preference datasets playing a crucial role in this process. While existing metrics primarily assess data quality based on either explicit or implicit reward margins, they often provide contradictory evaluati...
[ "Kexin Huang", "Junkang Wu", "Ziqian Chen", "Xue Wang", "Jinyang Gao", "Bolin Ding", "Jiancan Wu", "Xiangnan He", "Xiang Wang" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-02-25T00:00:00
https://arxiv.org/abs/2503.01864
https://arxiv.org/pdf/2503.01864v1
2503.01864
10.48550/arXiv.2503.01864
9
2
false
null
International Conference on Machine Learning
0.25
d1045fea6a5cd25c5c9072c7e21ef8c500c2e70b31ed8596a6381768c709e211
[ "arxiv", "semantic_scholar" ]
Discriminative Finetuning of Generative Large Language Models without Reward Models and Human Preference Data
Supervised fine-tuning (SFT) has become a crucial step for aligning pretrained large language models (LLMs) using supervised datasets of input-output pairs. However, despite being supervised, SFT is inherently limited by its generative training objective. To address its limitations, the existing common strategy is to f...
[ "Siqi Guo", "Ilgee Hong", "Vicente Balmaseda", "Changlong Yu", "Liang Qiu", "Xin Liu", "Haoming Jiang", "Tuo Zhao", "Tianbao Yang" ]
[ "cs.CL" ]
[ "Computer Science" ]
2025-02-25T00:00:00
https://arxiv.org/abs/2502.18679
https://arxiv.org/pdf/2502.18679v3
2502.18679
null
1
0
true
https://github.com/Optimization-AI/DFT
International Conference on Machine Learning
0.108
546c7440afa4f5ecdd641a331f9c8167d0924faf13930c14421bd8a9e3c4c2bc
[ "arxiv", "semantic_scholar" ]
Pretrain Value, Not Reward: Decoupled Value Policy Optimization
In this paper, we explore how directly pretraining a value model simplifies and stabilizes reinforcement learning from human feedback (RLHF). In reinforcement learning, value estimation is the key to policy optimization, distinct from reward supervision. The value function predicts the \emph{return-to-go} of a partial ...
[ "Chenghua Huang", "Lu Wang", "Fangkai Yang", "Pu Zhao", "Zhixu Li", "Qingwei Lin", "Dongmei Zhang", "Saravan Rajmohan", "Qi Zhang" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-02-24T00:00:00
https://arxiv.org/abs/2502.16944
https://arxiv.org/pdf/2502.16944v2
2502.16944
null
4
0
false
null
null
0.1747
659b98252045b8b988bfeaa938420b40d200899bdfedf0ebcfdb670a7ff7088b
[ "arxiv", "semantic_scholar" ]
RLHF in an SFT Way: From Optimal Solution to Reward-Weighted Alignment
Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning Large Language Models (LLMs) with human values. However, RLHF has been continuously challenged by its high complexity in implementation and computation consumption, specifically for online sampling-based methods like Proximal Policy Optimization ...
[ "Yuhao Du", "Zhuo Li", "Pengyu Cheng", "Zhihong Chen", "Yuejiao Xie", "Xiang Wan", "Anningzhe Gao" ]
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-02-16T00:00:00
https://arxiv.org/abs/2502.11026
https://arxiv.org/pdf/2502.11026v3
2502.11026
null
14
0
false
null
null
0.294
a45c279e3a6d9db75715153b9d5987808b52c3d225ae14a2b4770ac3c2f66810
[ "arxiv", "semantic_scholar" ]
Process Reward Models for LLM Agents: Practical Framework and Directions
We introduce Agent Process Reward Models (AgentPRM), a simple and scalable framework for training LLM agents to continually improve through interactions. AgentPRM follows a lightweight actor-critic paradigm, using Monte Carlo rollouts to compute reward targets and optimize policies. It requires minimal modifications to...
[ "Sanjiban Choudhury" ]
[ "cs.LG", "cs.AI" ]
[ "Computer Science" ]
2025-02-14T00:00:00
https://arxiv.org/abs/2502.10325
https://arxiv.org/pdf/2502.10325v1
2502.10325
10.48550/arXiv.2502.10325
66
4
true
https://github.com/sanjibanc/agent_prm
arXiv.org
0.4565
05be952385e975fc233797ffd481e9a4a7b665ef16ac4a543017ebe747792ef7
[ "arxiv", "semantic_scholar" ]
Provably Efficient Online RLHF with One-Pass Reward Modeling
Reinforcement Learning from Human Feedback (RLHF) has shown remarkable success in aligning Large Language Models (LLMs) with human preferences. Traditional RLHF methods rely on a fixed dataset, which often suffers from limited coverage. To this end, online RLHF has emerged as a promising direction, enabling iterative d...
[ "Long-Fei Li", "Yu-Yang Qian", "Peng Zhao", "Zhi-Hua Zhou" ]
[ "cs.LG", "stat.ML" ]
[ "Computer Science", "Mathematics" ]
2025-02-11T00:00:00
https://arxiv.org/abs/2502.07193
https://arxiv.org/pdf/2502.07193v3
2502.07193
null
6
0
false
null
null
0.2113
84f4bc71d45cf2dbfb72b536d632b54bd111ecab05f02e2814d24a7def6f70b5
[ "arxiv", "semantic_scholar" ]
Aequa: Fair Model Rewards in Collaborative Learning via Slimmable Networks
Collaborative learning enables multiple participants to learn a single global model by exchanging focused updates instead of sharing data. One of the core challenges in collaborative learning is ensuring that participants are rewarded fairly for their contributions, which entails two key sub-problems: contribution asse...
[ "Nurbek Tastan", "Samuel Horvath", "Karthik Nandakumar" ]
[ "cs.LG", "cs.DC" ]
[ "Computer Science" ]
2025-02-07T00:00:00
https://arxiv.org/abs/2502.04850
https://arxiv.org/pdf/2502.04850v2
2502.04850
10.48550/arXiv.2502.04850
6
0
false
null
International Conference on Machine Learning
0.2113
4840c97328c5a423ef7e9fe324012617d39b1ed2a8a4d81e3d7cf6e7b73fdef0
[ "arxiv", "semantic_scholar" ]
PILAF: Optimal Human Preference Sampling for Reward Modeling
As large language models increasingly drive real-world applications, aligning them with human values becomes paramount. Reinforcement Learning from Human Feedback (RLHF) has emerged as a key technique, translating preference data into reward models when oracle human values remain inaccessible. In practice, RLHF mostly ...
[ "Yunzhen Feng", "Ariel Kwiatkowski", "Kunhao Zheng", "Julia Kempe", "Yaqi Duan" ]
[ "cs.LG", "stat.ML" ]
[ "Computer Science", "Mathematics" ]
2025-02-06T00:00:00
https://arxiv.org/abs/2502.04270
https://arxiv.org/pdf/2502.04270v1
2502.04270
10.48550/arXiv.2502.04270
18
1
false
null
International Conference on Machine Learning
0.3197
2066d042cc4261d2f1f01f49f10e698c8844f43abbe1725df679d6897784f218
[ "arxiv", "semantic_scholar" ]
Towards Cost-Effective Reward Guided Text Generation
Reward-guided text generation (RGTG) has emerged as a viable alternative to offline reinforcement learning from human feedback (RLHF). RGTG methods can align baseline language models to human preferences without further training like in standard RLHF methods. However, they rely on a reward model to score each candidate...
[ "Ahmad Rashid", "Ruotian Wu", "Rongqi Fan", "Hongliang Li", "Agustinus Kristiadi", "Pascal Poupart" ]
[ "cs.LG", "cs.CL" ]
[ "Computer Science" ]
2025-02-06T00:00:00
https://arxiv.org/abs/2502.04517
https://arxiv.org/pdf/2502.04517v2
2502.04517
10.48550/arXiv.2502.04517
6
0
false
null
International Conference on Machine Learning
0.2113
6200f20906672d86011fd07d8d65c6fef8d34f5d6f6164b0cdcbce53cb21863a
[ "arxiv", "semantic_scholar" ]
PerPO: Perceptual Preference Optimization via Discriminative Rewarding
This paper presents Perceptual Preference Optimization (PerPO), a perception alignment method aimed at addressing the visual discrimination challenges in generative pre-trained multimodal large language models (MLLMs). To align MLLMs with human visual perception process, PerPO employs discriminative rewarding to gather...
[ "Zining Zhu", "Liang Zhao", "Kangheng Lin", "Jinze Yang", "En Yu", "Chenglong Liu", "Haoran Wei", "Jianjian Sun", "Zheng Ge", "Xiangyu Zhang" ]
[ "cs.AI", "cs.CL", "cs.LG" ]
[ "Computer Science" ]
2025-02-05T00:00:00
https://arxiv.org/abs/2502.04371
https://arxiv.org/pdf/2502.04371v1
2502.04371
10.48550/arXiv.2502.04371
8
0
false
null
arXiv.org
0.2386
dbffc4cc0679aff4e76f98f36da75ee33cd60574d68c9c638e9090c70d21cdd9
[ "arxiv", "semantic_scholar" ]
Reviving The Classics: Active Reward Modeling in Large Language Model Alignment
Building neural reward models from human preferences is a pivotal component in reinforcement learning from human feedback (RLHF) and large language model alignment research. Given the scarcity and high cost of human annotation, how to select the most informative pairs to annotate is an essential yet challenging open pr...
[ "Yunyi Shen", "Hao Sun", "Jean-François Ton" ]
[ "cs.CL", "cs.AI", "cs.LG" ]
[ "Computer Science" ]
2025-02-04T00:00:00
https://arxiv.org/abs/2502.04354
https://arxiv.org/pdf/2502.04354v1
2502.04354
10.48550/arXiv.2502.04354
9
0
true
null
arXiv.org
0.25
df7959677c3be321f2cf85732cfe887e1322f01bd9dca314d2bf44d5a023ea90
[ "arxiv", "semantic_scholar" ]
Process Reinforcement through Implicit Rewards
Dense process rewards have proven a more effective alternative to the sparse outcome-level rewards in the inference-time scaling of large language models (LLMs), particularly in tasks requiring complex multi-step reasoning. While dense rewards also offer an appealing choice for the reinforcement learning (RL) of LLMs s...
[ "Ganqu Cui", "Lifan Yuan", "Zefan Wang", "Hanbin Wang", "Yuchen Zhang", "Jiacheng Chen", "Wendi Li", "Bingxiang He", "Yuchen Fan", "Tianyu Yu", "Qixin Xu", "Weize Chen", "Jiarui Yuan", "Huayu Chen", "Kaiyan Zhang", "Xingtai Lv", "Shuo Wang", "Yuan Yao", "Xu Han", "Hao Peng", ...
[ "cs.LG", "cs.AI", "cs.CL" ]
[ "Computer Science" ]
2025-02-03T00:00:00
https://arxiv.org/abs/2502.01456
https://arxiv.org/pdf/2502.01456v2
2502.01456
10.48550/arXiv.2502.01456
365
54
true
https://github.com/PRIME-RL/PRIME
arXiv.org
0.8702
7a04fc67afec50ccb7684918cade472d03a8154d5a39dc377c541e95a7ceb1df
[ "arxiv", "semantic_scholar" ]
Avoiding $\mathbf{exp(R_{max})}$ scaling in RLHF through Preference-based Exploration
Reinforcement Learning from Human Feedback (RLHF) has emerged as a pivotal technique for large language model (LLM) alignment. This paper studies the setting of online RLHF and focus on improving sample efficiency. All existing algorithms in online RLHF, whether doing passive exploration or active exploration, suffer f...
[ "Mingyu Chen", "Yiding Chen", "Wen Sun", "Xuezhou Zhang" ]
[ "cs.LG", "cs.AI", "stat.ML" ]
[ "Computer Science", "Mathematics" ]
2025-02-02T00:00:00
https://arxiv.org/abs/2502.00666
https://arxiv.org/pdf/2502.00666v3
2502.00666
null
1
0
true
https://github.com/MYC000801/SE-POPO
null
0.0753