RL from Human Feedback

RL: Reinforcement Learning

Follow this topic Ask about this topic

RLHF studies how to use human judgments and interventions to teach AI systems which responses or actions are desirable, especially when goals are hard to specify with a fixed reward. Common approaches learn reward models from pairwise comparisons and optimize policies with PPO or GRPO, while methods such as Direct Preference Optimization learn directly from preference pairs; active learning and online feedback aim to reduce the data required. Research spans language-model alignment and benchmarks such as AlpacaEval and MT-Bench, as well as robot learning from corrections and demonstrations. A central challenge is that feedback can be noisy, biased, or diverse, and optimizing an imperfect reward can cause reward hacking, so work increasingly studies uncertainty, robustness, fairness, and how to measure real-world improvement rather than proxy scores alone.

Latest papers

128 papers10 papers in the last four weeks, up 67% on the four weeks before. 0.1% of all new papers.

Mar 30, 2026

Corruption-robust Offline Multi-agent Reinforcement Learning From Human Feedback

We consider robustness against data corruption in offline multi-agent reinforcement learning from human feedback (MARLHF) under a strong-contamination model: given a dataset

D

of trajectory-preference tuples (each preference being an

n

-dimensional binary label vector representing each of the

n

agents' preferences), an

ε

-fraction of the samples may be arbitrarily corrupted. We model the problem using the framework of linear Markov games. First, under a uniform coverage assumption - where every policy of interest is sufficiently represented in the clean (prior to corruption) data - we introduce a robust estimator that guarantees an

O(ε^{1 - o(1)})

bound on the Nash equilibrium gap. Next, we move to the more challenging unilateral coverage setting, in which only a Nash equilibrium and its single-player deviations are covered. In this case, our proposed algorithm achieves an

O(\sqrtε)

bound on the Nash gap. Both of these procedures, however, suffer from intractable computation. To address this, we relax our solution concept to coarse correlated equilibria (CCE). Under the same unilateral coverage regime, we derive a quasi-polynomial-time algorithm whose CCE gap scales as

O(\sqrtε)

. To the best of our knowledge, this is the first systematic treatment of adversarial data corruption in offline MARLHF.

Andi Nika, Debmalya Mandal, Parameswaran Kamalaruban +2

MPI-SWS · University of Warwick · Visa

Multi-Agent Reinforcement Learning Offline MARL Robust RL RL from Human Feedback

Mar 27, 2026

Distributionally Robust Token Optimization in RLHF

Large Language Models (LLMs) tend to respond correctly to prompts that align well with the data they were trained and fine-tuned on. Yet, small shifts in wording, format, or language can trigger surprisingly large failures, especially on multi-step reasoning problems. To address this problem, we propose a Distributionally Robust Token Optimization (DRTO) approach, which combines token-level Reinforcement Learning from Human Feedback (RLHF) with Distributionally Robust Optimization (DRO). DRTO constructs f-divergence ambiguity sets over span-level actor losses, providing a principled way to emphasize difficult response segments during policy optimization. Empirically, DRTO enhances consistency under distribution shifts in multiple reasoning benchmarks among different tasks, achieving

+4.4

percentage points on MATH-500 and

+2.7

percentage points on LiveCodeBench over standard RTO.

Yeping Jin, Jiaming Hu, Ioannis Ch. Paschalidis

Department of System Engineering Boston University, Boston, MA 02215, USA · Department of Math & Statistics Boston University, Boston, MA 02215, USA

Language Model Robustness Reasoning Consistency in Language Models RL from Human Feedback Distributionally Robust RL Distributionally Robust Optimization

Mar 26, 2026

Neuro-Cognitive Reward Modeling for Human-Centered Autonomous Vehicle Control

Recent advancements in computer vision have accelerated the development of autonomous driving. Despite these advancements, training machines to drive in a way that aligns with human expectations remains a significant challenge. Human factors are still essential, as humans possess a sophisticated cognitive system capable of rapidly interpreting scene information and making accurate decisions. Aligning machine with human intent has been explored with Reinforcement Learning with Human Feedback (RLHF). Conventional RLHF methods rely on collecting human preference data by manually ranking generated outputs, which is time-consuming and indirect. In this work, we propose an electroencephalography (EEG)-guided decision-making framework to incorporate human cognitive insights without behaviour response interruption into reinforcement learning (RL) for autonomous driving. We collected EEG signals from 20 participants in a realistic driving simulator and analyzed event-related potentials (ERP) in response to sudden environmental changes. Our proposed framework employs a neural network to predict the strength of ERP based on the cognitive information from visual scene information. Moreover, we explore the integration of such cognitive information into the reward signal of the RL algorithm. Experimental results show that our framework can improve the collision avoidance ability of the RL algorithm, highlighting the potential of neuro-cognitive feedback in enhancing autonomous driving systems. Our project page is: https://alex95gogo.github.io/Cognitive-Reward/.

Zhuoli Zhuang, Yu-Cheng Chang, Yu-Kai Wang +2

University of Technology Sydney

RL for Autonomous Driving Reinforcement Learning RL from Human Feedback Multimodal Reward Modeling

Mar 24, 2026

Unbiased Reward Modeling from Implicit Feedback for LLM Alignment

Despite the success of reinforcement learning from human feedback (RLHF), existing reward modeling methods largely rely on explicit feedback, which is costly to collect and difficult to scale. This work studies implicit reward modeling, learning reward models from implicit user feedback, such as clicks, copies and skips. While scalable and cost-effective, implicit feedback poses two key challenges: It lacks definitive negative samples, which makes standard positive-negative classification methods inapplicable; It suffers from selection bias, where responses have heterogeneous propensities to elicit feedback, which further obscures definitive negative samples. To address these challenges, we propose ImplicitRM, which learns unbiased reward models from implicit feedback. It stratifies training samples into four latent groups using a stratification model and derives a likelihood-maximization objective that is theoretically unbiased, thereby addressing both challenges. Experiments across diverse LLM backbones and benchmark datasets validate that ImplicitRM learns accurate reward models from implicit feedback and improves performance on downstream RLHF tasks.

Hao Wang, Haocheng Yang, Licheng Pan +7

Peking University · Xiaohongshu Inc. · National University of Singapore +1

Reward Modeling LLM Alignment RL from Human Feedback

Mar 23, 2026

Privacy-Preserving Reinforcement Learning from Human Feedback via Decoupled Reward Modeling

Preference-based fine-tuning has become an important component in training large language models, and the data used at this stage may contain sensitive user information. A central question is how to design a differentially private pipeline that is well suited to the distinct structure of reinforcement learning from human feedback. We propose a privacy-preserving framework that imposes differential privacy only on reward learning and derives the final policy from the resulting private reward model. Theoretically, we study the suboptimality gap and show that privacy contributes an additional additive term beyond the usual non-private statistical error. We also establish a minimax lower bound and show that the dominant term changes with sample size and privacy level, which in turn characterizes regimes in which the upper bound is rate-optimal up to logarithmic factors. Empirically, synthetic experiments confirm the scaling predicted by the theory, and experiments on the Anthropic HH-RLHF dataset using the Gemma-2B-IT model show stronger private alignment performance than existing differentially private baseline methods across privacy budgets.

Young Hyun Cho, Will Wei Sun

Purdue University

Privacy-Preserving ML Privacy-Preserving Language Models RL from Human Feedback Preference-Based RL Reward Modeling

Mar 20, 2026

Regret Bounds for Reinforcement Learning from Multi-Source Imperfect Preferences

Reinforcement learning from human feedback (RLHF) replaces hard-to-specify rewards with pairwise trajectory preferences, yet regret-oriented theory often assumes that preference labels are generated consistently from a single ground-truth objective. In practical RLHF systems, however, feedback is typically multi-source (annotators, experts, reward models, heuristics) and can exhibit systematic, persistent mismatches due to subjectivity, expertise variation, and annotation/modeling artifacts. We study episodic RL from multi-source imperfect preferences through a cumulative imperfection budget: for each source, the total deviation of its preference probabilities from an ideal oracle is at most

ω

over

K

episodes. We propose a unified algorithm with regret

\tilde{O}(\sqrt{K/M}+ω)

, which exhibits a best-of-both-regimes behavior: it achieves

M

-dependent statistical gains when imperfection is small (where

M

is the number of sources), while remaining robust with unavoidable additive dependence on

ω

when imperfection is large. We complement this with a lower bound

\tildeΩ(\max\{\sqrt{K/M},ω\})

, which captures the best possible improvement with respect to

M

and the unavoidable dependence on

ω

, and a counterexample showing that naïvely treating imperfect feedback as oracle-consistent can incur regret as large as

\tildeΩ(\min\{ω\sqrt{K},K\})

. Technically, our approach involves imperfection-adaptive weighted comparison learning, value-targeted transition estimation to control hidden feedback-induced distribution shift, and sub-importance sampling to keep the weighted objectives analyzable, yielding regret guarantees that quantify when multi-source feedback provably improves RLHF and how cumulative imperfection fundamentally limits it.

Ming Shi, Yingbin Liang, Ness B. Shroff +1

Department of Electrical Engineering, University at Buffalo, Buffalo, NY 14260, USA · Department of Electrical and Computer Engineering, The Ohio State University, Columbus, OH 43210, USA · Department of Electrical and Computer Engineering and Computer Science and Engineering, The Ohio State University, Columbus, OH 43210, USA +1

Preference-Based RL RL from Human Feedback Regret Minimization in RL Pairwise Preference Learning

Mar 19, 2026

CausalRM: Causal-Theoretic Reward Modeling for RLHF from Observational User Feedbacks

Despite the success of reinforcement learning from human feedback (RLHF) in aligning language models, current reward modeling heavily relies on experimental feedback data collected from human annotators under controlled and costly conditions. In this work, we introduce observational reward modeling -- learning reward models with observational user feedback (e.g., clicks, copies, and upvotes) -- as a scalable and cost-effective alternative. We identify two fundamental challenges in this setting: (1) observational feedback is noisy due to annotation errors, which deviates it from true user preference; (2) observational feedback is biased by user preference, where users preferentially provide feedback on responses they feel strongly about, which creats a distribution shift between training and inference data. To address these challenges, we propose CausalRM, a causal-theoretic reward modeling framework that aims to learn unbiased reward models from observational feedback. To tackle challenge (1), CausalRM introduces a noise-aware surrogate loss term that is provably equivalent to the primal loss under noise-free conditions by explicitly modeling the annotation error generation process. To tackle challenge (2), CausalRM uses propensity scores -- the probability of a user providing feedback for a given response -- to reweight training samples, yielding a loss function that eliminates user preference bias. Extensive experiments across diverse LLM backbones and benchmark datasets validate that CausalRM effectively learns accurate reward signals from noisy and biased observational feedback and delivers substantial performance improvements on downstream RLHF tasks -- including a 49.2% gain on WildGuardMix and a 32.7% improvement on HarmBench. Code is available on our project website.

Hao Wang, Licheng Pan, Zhichao Chen +7

Peking University · Xiaohongshu Inc. · Zhejiang University

Reward Modeling RL from Human Feedback

Mar 18, 2026

Efficient Exploration at Scale

We develop an online learning algorithm that dramatically improves the data efficiency of reinforcement learning from human feedback (RLHF). Our algorithm incrementally updates reward and language models as choice data is received. The reward model is fit to the choice data, while the language model is updated by a variation of reinforce, with reinforcement signals provided by the reward model. Several features enable the efficiency gains: a small affirmative nudge added to each reinforcement signal, an epistemic neural network that models reward uncertainty, and information-directed exploration. With Gemma large language models (LLMs), our algorithm matches the performance of offline RLHF trained on 200K labels using fewer than 20K labels, representing more than a 10x gain in data efficiency. Extrapolating from our results, we expect our algorithm trained on 1M labels to match offline RLHF trained on 1B labels. This represents a 1,000x gain. To our knowledge, these are the first results to demonstrate that such large improvements are possible.

Seyed Mohammad Asghari, Chris Chute, Vikranth Dwaracherla +5

The Efficient Agent Team, Google DeepMind

RL from Human Feedback Data-Efficient Learning RL Exploration Reward Modeling

Mar 17, 2026

Via Negativa for AI Alignment: Why Negative Constraints Are Structurally Superior to Positive Preferences

Recent empirical results have demonstrated that training large language models (LLMs) with negative-only feedback can match or exceed standard reinforcement learning from human feedback (RLHF). Negative Sample Reinforcement achieves parity with PPO on mathematical reasoning; Distributional Dispreference Optimization trains effectively using only dispreferred samples; and Constitutional AI outperforms pure RLHF on harmlessness benchmarks. Yet no unified theoretical account explains why negative signals are so effective. This paper proposes such an account: positive preferences and negative constraints are structurally asymmetric. Positive preferences ("which is better") encode continuously coupled, context-dependent human values that cannot be exhaustively specified -- leading models to learn surface correlates such as agreement with the user (sycophancy). Negative constraints ("what is wrong") encode discrete, finite, independently verifiable prohibitions that can converge to a stable boundary. This asymmetry -- rooted in Popper's falsification logic and the epistemology of negative knowledge -- explains both the sycophancy failure of preference-based RLHF and the surprising effectiveness of negative-signal methods. We argue that alignment research should shift its center of gravity from "learning what humans prefer" to "learning what humans reject," and offer testable predictions for this framework.

Quan Cheng

Tsinghua University

LLM Alignment RL from Human Feedback Preference-Based RL

Mar 13, 2026

Swap-guided Preference Learning for Personalized Reinforcement Learning from Human Feedback

Reinforcement Learning from Human Feedback (RLHF) is a widely used approach to align large-scale AI systems with human values. However, RLHF typically assumes a single, universal reward, which overlooks diverse preferences and limits personalization. Variational Preference Learning (VPL) seeks to address this by introducing user-specific latent variables. Despite its promise, we found that VPL suffers from posterior collapse. While this phenomenon is well known in VAEs, it has not previously been identified in preference learning frameworks. Under sparse preference data and with overly expressive decoders, VPL may cause latent variables to be ignored, reverting to a single-reward model. To overcome this limitation, we propose Swap-guided Preference Learning (SPL). The key idea is to construct fictitious swap annotators and use the mirroring property of their preferences to guide the encoder. SPL introduces three components: (1) swap-guided base regularization, (2) Preferential Inverse Autoregressive Flow (P-IAF), and (3) adaptive latent conditioning. Experiments show that SPL mitigates collapse, enriches user-specific latents, and improves preference prediction. Our code and data are available at https://github.com/cobang0111/SPL

Gihoon Kim, Euntai Kim

Yonsei University · Korea Institute of Science and Technology

Preference Learning RL from Human Feedback Personalized Language Model Alignment User Preference Modeling Posterior Collapse

Mar 10, 2026

ActiveUltraFeedback: Efficient Preference Data Generation using Active Learning

Reinforcement Learning from Human Feedback (RLHF) has become the standard for aligning Large Language Models (LLMs), yet its efficacy is bottlenecked by the high cost of acquiring preference data, especially in low-resource and expert domains. To address this, we introduce ACTIVEULTRAFEEDBACK, a modular active learning pipeline that leverages uncertainty estimates to dynamically identify the most informative responses for annotation. Our pipeline facilitates the systematic evaluation of standard response selection methods alongside DOUBLE REVERSE THOMPSON SAMPLING (DRTS) and DELTAUCB, two novel methods prioritizing response pairs with large predicted quality gaps, leveraging recent results showing that such pairs provide good signals for fine-tuning. Our experiments demonstrate that ACTIVEULTRAFEEDBACK yields high-quality datasets that lead to significant improvements in downstream performance, notably achieving comparable or superior results with as little as one-sixth of the annotated data relative to static baselines. Our pipeline is available at https://github.com/lasgroup/ActiveUltraFeedback and our preference datasets at https://huggingface.co/ActiveUltraFeedback.

Davit Melikidze, Marian Schneider, Jessica Lam +4

ETH Zurich · Institute of Neuroinformatics, University of Zurich and ETH Zurich · ETH AI Center

Data Selection Active Learning RL from Human Feedback Uncertainty Sampling

Mar 9, 2026

Aligning to Illusions: Choice Blindness in Human and AI Feedback

Reinforcement Learning from Human Feedback (RLHF) assumes annotator preferences reflect stable internal states. We challenge this through three experiments spanning the preference pipeline. In a human choice blindness study, 91% of surreptitiously swapped preferences go undetected, extending choice blindness to third-person evaluative comparison of unfamiliar text. Testing fifteen LLM judges as potential replacements, we find detection relies on shallow text matching rather than genuine self-monitoring: removing prior reasoning from context causes blindness to surge from near-zero to over 50%, while explicit social pressure induces near-universal compliance. In a dose-response experiment across two architectures from 86M to 2B parameters, one-sixth to one-third of labels must be corrupted before the reward signal halves, yet standard pairwise accuracy remains virtually unchanged. A Best-of-N evaluation confirms this translates to downstream policy degradation: at 50% corruption, reward-guided selection produces no improvement over random sampling, while the proxy model reports monotonically increasing scores. Together, these results reveal a preference construction problem: the signal entering RLHF is shaped by elicitation context in ways that neither human metacognition, LLM self-monitoring, nor standard evaluation metrics can detect.

Wenbin Wu

Cambridge Judge Business School, University of Cambridge, UK

RL from Human Feedback Bias in LLM-as-a-Judge

Mar 5, 2026

Why Is RLHF Alignment Shallow? A Gradient Analysis

Why is safety alignment in LLMs shallow? We prove that gradient-based alignment inherently concentrates on positions where harm is decided and vanishes beyond. Using a martingale decomposition of sequence-level harm, we derive an exact characterization of alignment gradients. The gradient at position

t

equals the covariance between the conditional expected harm and the score function. This implies that positions beyond the harm horizon where the output's harmfulness is already determined receive zero gradient signal during training. This explains empirical observations that KL divergence between aligned and base models concentrates on early tokens. Consequently, standard alignment objectives cannot produce deep alignment, regardless of optimization quality. We introduce the concept of harm information

I_t

, which quantifies each position's influence on harm, and prove that equilibrium KL divergence tracks this quantity. Finally, we derive an objective based on recovery penalties that creates gradient signal at all positions, providing theoretical grounding for empirically successful data augmentation techniques.

Robin Young

Department of Computer Science and Technology University of Cambridge Cambridge, UK

LLM Alignment LLM Safety Alignment RL from Human Feedback

Mar 4, 2026

Bootstrapping Exploration with Group-Level Natural Language Feedback in Reinforcement Learning

Large language models (LLMs) typically receive diverse natural language (NL) feedback through interaction with the environment. However, current reinforcement learning (RL) algorithms rely solely on scalar rewards, leaving the rich information in NL feedback underutilized and leading to inefficient exploration. In this work, we propose GOLF, an RL framework that explicitly exploits group-level language feedback to guide targeted exploration through actionable refinements. GOLF aggregates two complementary feedback sources: (i) external critiques that pinpoint errors or propose targeted fixes, and (ii) intra-group attempts that supply alternative partial ideas and diverse failure patterns. These group-level feedbacks are aggregated to produce high-quality refinements, which are adaptively injected into training as off-policy scaffolds to provide targeted guidance in sparse-reward regions. Meanwhile, GOLF jointly optimizes generation and refinement within a unified RL loop, creating a virtuous cycle that continuously improves both capabilities. Experiments on both verifiable and non-verifiable benchmarks show that GOLF achieves superior performance and exploration efficiency, achieving 2.2

\times

improvements in sample efficiency compared to RL methods trained solely on scalar rewards. Code is available at https://github.com/LuckyyySTA/GOLF.

Lei Huang, Xiang Cheng, Chenxiao Zhao +6

Harbin Institute of Technology · Xiaohongshu Inc.

RL Exploration RL for Language Models RL from Human Feedback Sparse-Reward RL

Mar 3, 2026

PrivMedChat: End-to-End Differentially Private RLHF for Medical Dialogue Systems

Large language models are increasingly used for patient-facing medical assistance and clinical decision support, but adapting them to clinical dialogue often requires supervision derived from doctor-patient conversations that may contain sensitive information. Conventional supervised fine-tuning and reinforcement learning from human feedback (RLHF) can amplify memorization, enabling membership inference and disclosure of rare training-set details. We present PrivMedChat (Private Medical Chat), an end-to-end framework for differentially private RLHF (DP-RLHF) for medical dialogue systems. Our approach enforces differential privacy at each training stage that accesses dialogue-derived supervision, combining DP-SGD for supervised fine-tuning and reward model learning from preference pairs, and DP-aware policy optimization for alignment. To avoid costly clinician labeling, we introduce an annotation-free preference construction strategy that pairs physician responses with filtered non-expert generations. We evaluate PrivMedChat across medical dialogue tasks and assess utility, safety, and privacy under consistent privacy accounting, thereby providing a practical pathway to align medical chatbots while offering formal privacy guarantees. We open-source our code at https://github.com/sudip-bhujel/privmedchat.

Sudip Bhujel

Department of Computer Science, University of Kentucky, Lexington, KY, USA

Privacy-Preserving Language Models RL from Human Feedback Differential Privacy

Feb 26, 2026

Human Supervision as an Information Bottleneck: A Unified Theory of Error Floors in Human-Guided Learning

Large language models are trained primarily on human-generated data and feedback, yet they exhibit persistent errors arising from annotation noise, subjective preferences, and the limited expressive bandwidth of natural language. We argue that these limitations reflect structural properties of the supervision channel rather than model scale or optimization. We develop a unified theory showing that whenever the human supervision channel is not sufficient for a latent evaluation target, it acts as an information-reducing channel that induces a strictly positive excess-risk floor for any learner dominated by it. We formalize this Human-Bounded Intelligence limit and show that across six complementary frameworks (operator theory, PAC-Bayes, information theory, causal inference, category theory, and game-theoretic analyses of reinforcement learning from human feedback), non-sufficiency yields strictly positive lower bounds arising from the same structural decomposition into annotation noise, preference distortion, and semantic compression. The theory explains why scaling alone cannot eliminate persistent human-aligned errors and characterizes conditions under which auxiliary non-human signals (e.g., retrieval, program execution, tools) increase effective supervision capacity and collapse the floor by restoring information about the latent target. Experiments on real preference data, synthetic known-target tasks, and externally verifiable benchmarks confirm the predicted structural signatures: human-only supervision exhibits a persistent floor, while sufficiently informative auxiliary channels strictly reduce or eliminate excess error.

Alejandro Rodriguez Dominguez

Miralta Finance Bank S.A., Spain · Department of Computer Science, University of Reading, UK

Information Bottleneck Learning with Noisy Labels RL from Human Feedback

Feb 26, 2026

Agency and Architectural Limits: Why Optimization-Based Systems Cannot Be Norm-Responsive

AI systems are increasingly deployed in high-stakes contexts (medical diagnosis, legal research, financial analysis) under the assumption they can be governed by norms. This paper demonstrates that the assumption is formally invalid for optimization-based systems, specifically Large Language Models trained via Reinforcement Learning from Human Feedback (RLHF). Genuine agency requires two necessary and jointly sufficient architectural conditions. First, the capacity to maintain certain boundaries as non-negotiable constraints rather than tradeable weights (Incommensurability). Second, a non-inferential mechanism capable of suspending processing when those boundaries are threatened (Apophatic Responsiveness). RLHF-based systems are constitutively incompatible with both conditions. The operations that make optimization powerful, unifying all values on a scalar metric and always selecting the highest-scoring output, are precisely the operations that preclude normative governance and agency. This incompatibility is not a correctable training bug awaiting a technical fix. It is a formal constraint inherent to what optimization is. Consequently, documented failure modes (sycophancy, hallucination, and unfaithful reasoning) are not accidents but expected structural manifestations. Misaligned deployment triggers a second-order risk termed the Convergence Crisis. When humans are forced to verify AI outputs under metric pressure, they degrade from genuine agents into criteria-checking optimizers, eliminating the only component capable of bearing normative accountability. Beyond the incompatibility proof, this paper's primary positive contribution is a substrate-neutral architectural specification deriving what any system (biological, artificial, or institutional) must necessarily satisfy to qualify as a genuine agent rather than a sophisticated instrument.

Radha Sarma

AI Governance AI Alignment Normative Conformity in Language Models RL from Human Feedback

Feb 26, 2026

Provably Efficient Regularized Online RLHF with Generalized Bilinear Preferences

We consider the problem of regularized best-response max-regret minimization in online RLHF under general preferences and bandit feedback. While various regularizers are utilized to robustify alignment, known polylogarithmic regret guarantees remain heavily specific to KL. To investigate whether such fast rates extend beyond KL, we adopt the Generalized Bilinear Preference Model (GBPM) -- capturing intransitive preferences over

d

-dimensional item-wise features via a rank-

2r

skew-symmetric matrix -- to isolate the impact of generic regularization. Crucially, under GBPM, we prove that the dual gap of any greedy policy is bounded by the squared estimation error, derived using only strong convexity and skew-symmetry. Under a feature coverage assumption, we establish a generic polylogarithmic regret of

\tilde{\mathcal{O}}(ηd^4 C_{\min}^{-1} (\log T)^2 \wedge d^2 C_{\min}^{-1/2} \sqrt{T})

with Greedy Sampling, and a dimension-wise improved regret (for well-conditioned arm-sets) of

\tilde{\mathcal{O}}(C_{\min}^{-2} \sqrt{ηr T} \wedge r^{1/3} C_{\min}^{-4/3} T^{2/3})

with Explore-Then-Commit, where

η^{-1}

is the regularization coefficient,

T

is the time horizon, and

C_{\min}

is an arm-set dependent quantity. This demonstrates that ``fast'' regrets are not KL-specific, but rather a fundamental consequence of generic strongly convex geometry.

Junghyun Lee, Minju Hong, Kwang-Sung Jun +2

KAIST AI · KAIST EE · POSTECH CSE/AI

Regret Minimization in RL Preference-Based RL RL from Human Feedback Preference Learning Pairwise Preference Learning

Feb 26, 2026

RLHFless: Serverless Computing for Efficient RLHF

Reinforcement Learning from Human Feedback (RLHF) has been widely applied to Large Language Model (LLM) post-training to align model outputs with human preferences. Recent models, such as DeepSeek-R1, have also shown RLHF's potential to improve LLM reasoning on complex tasks. In RL, inference and training co-exist, creating dynamic resource demands throughout the workflow. Compared to traditional RL, RLHF further challenges training efficiency due to expanding model sizes and resource consumption. Several RLHF frameworks aim to balance flexible abstraction and efficient execution. However, they rely on serverful infrastructures, which struggle with fine-grained resource variability. As a result, during synchronous RLHF training, idle time between or within RL components often causes overhead and resource wastage. To address these issues, we present RLHFless, the first scalable training framework for synchronous RLHF, built on serverless computing environments. RLHFless adapts to dynamic resource demands throughout the RLHF pipeline, pre-computes shared prefixes to avoid repeated computation, and uses a cost-aware actor scaling strategy that accounts for response length variation to find sweet spots with lower cost and higher speed. In addition, RLHFless assigns workloads efficiently to reduce intra-function imbalance and idle time. Experiments on both physical testbeds and a large-scale simulated cluster show that RLHFless achieves up to 1.35x speedup and 44.8% cost reduction compared to the state-of-the-art baseline.

Rui Wei, Hanfei Yu, Shubham Jain +5

Stevens Institute of Technology · Northeastern University · Stony Brook University +1

Efficient Language Model Training RL from Human Feedback

Feb 25, 2026

Provable Last-Iterate Convergence for Multi-Objective Safe LLM Alignment via Optimistic Primal-Dual

Reinforcement Learning from Human Feedback (RLHF) plays a significant role in aligning Large Language Models (LLMs) with human preferences. While RLHF with expected reward constraints can be formulated as a primal-dual optimization problem, standard primal-dual methods only guarantee convergence with a distributional policy where the saddle-point problem is in convex-concave form. Moreover, standard primal-dual methods may exhibit instability or divergence in the last iterate under policy parameterization in practical applications. In this work, we propose a universal primal-dual framework for safe RLHF that unifies a broad class of existing alignment algorithms, including safe-RLHF, one-shot, and multi-shot based methods. Building on this framework, we introduce an optimistic primal-dual (OPD) algorithm that incorporates predictive updates for both primal and dual variables to stabilize saddle-point dynamics. We establish last-iterate convergence guarantees for the proposed method, covering both exact policy optimization in the distributional space and convergence to a neighborhood of the optimal solution whose gap is related to approximation error and bias under parameterized policies. Our analysis reveals that optimism plays a crucial role in mitigating oscillations inherent to constrained alignment objectives, thereby closing a key theoretical gap between constrained RL and practical RLHF.

Yining Li, Peizhong Ju, Ness Shroff

The Ohio State University · University of Kentucky

LLM Safety Alignment RL from Human Feedback Multi-Objective Language Model Alignment Constrained RL Constrained Optimization

Feb 25, 2026

Generalisation of RLHF under Reward Shift and Clipped KL Regularisation

Alignment and adaptation in large language models heavily rely on reinforcement learning from human feedback (RLHF); yet, theoretical understanding of its generalisability remains premature, especially when the learned reward could shift, and the KL control is estimated and clipped. To address this issue, we develop generalisation theory for RLHF that explicitly accounts for (1) reward shift: reward models are trained on preference data from earlier or mixed behaviour policies while RLHF optimises the current policy on its own rollouts; and (2) clipped KL regularisation: the KL regulariser is estimated from sampled log-probability ratios and then clipped for stabilisation, resulting in an error to RLHF. We present generalisation bounds for RLHF, suggesting that the generalisation error stems from a sampling error from prompts and rollouts, a reward shift error, and a KL clipping error. We also discuss special cases of (1) initialising RLHF parameters with a uniform prior over a finite space, and (2) training RLHF by stochastic gradient descent, as an Ornstein-Uhlenbeck process. The theory yields practical implications in (1) optimal KL clipping threshold, and (2) budget allocation in prompts, rollouts, and preference data.

Kenton Tang, Yuzhu Chen, Fengxiang He

University of Edinburgh · University of Science and Technology of China

RL from Human Feedback Generalization in RL KL-Regularized RL

Feb 23, 2026

Sample-Efficient Learning with Online Expert Correction for Autonomous Catheter Steering in Endovascular Bifurcation Navigation

Robot-assisted endovascular intervention offers a safe and effective solution for remote catheter manipulation, reducing radiation exposure while enabling precise navigation. Reinforcement learning (RL) has recently emerged as a promising approach for autonomous catheter steering; however, conventional methods suffer from sparse reward design and reliance on static vascular models, limiting their sample efficiency and generalization to intraoperative variations. To overcome these challenges, this paper introduces a sample-efficient RL framework with online expert correction for autonomous catheter steering in endovascular bifurcation navigation. The proposed framework integrates three key components: (1) A segmentation-based pose estimation module for accurate real-time state feedback, (2) A fuzzy controller for bifurcation-aware orientation adjustment, and (3) A structured reward generator incorporating expert priors to guide policy learning. By leveraging online expert correction, the framework reduces exploration inefficiency and enhances policy robustness in complex vascular structures. Experimental validation on a robotic platform using a transparent vascular phantom demonstrates that the proposed approach achieves convergence in 123 training episodes -- a 25.9% reduction compared to the baseline Soft Actor-Critic (SAC) algorithm -- while reducing average positional error to 83.8% of the baseline. These results indicate that combining sample-efficient RL with online expert correction enables reliable and accurate catheter steering, particularly in anatomically challenging bifurcation scenarios critical for endovascular navigation.

Hao Wang, Tianliang Yao, Bo Lu +4

Department of Control Science and Engineering, College of Electronics and Information Engineering, and Shanghai Institute of Intelligent Science and Technology, Tongji University, Shanghai 200092, China · Department of Electronic Engineering, Faculty of Engineering, The Chinese University of Hong Kong, Hong Kong SAR 999077, China · Robotics and Microsystems Center, School of Mechanical and Electric Engineering, Soochow University, Suzhou, Jiangsu 215131, China +3

Robotic Endovascular Intervention RL for Robotics Data-Efficient Robot Learning RL from Human Feedback

Feb 23, 2026

IR $^3$ : Contrastive Inverse Reinforcement Learning for Interpretable Detection and Mitigation of Reward Hacking

Reinforcement Learning from Human Feedback (RLHF) enables powerful LLM alignment but can introduce reward hacking - models exploit spurious correlations in proxy rewards without genuine alignment. Compounding this, the objectives internalized during RLHF remain opaque, making hacking behaviors difficult to detect or correct. We introduce IR3 (Interpretable Reward Reconstruction and Rectification), a framework that reverse-engineers, interprets, and surgically repairs the implicit objectives driving RLHF-tuned models. We propose Contrastive Inverse Reinforcement Learning (C-IRL), which reconstructs the implicit reward function by contrasting paired responses from post-alignment and baseline policies to explain behavioral shifts during RLHF. We then decompose the reconstructed reward via sparse autoencoders into interpretable features, enabling identification of hacking signatures through contribution analysis. Finally, we propose mitigation strategies - clean reward optimization, adversarial shaping, constrained optimization, and feature-guided distillation - that target problematic features while preserving beneficial alignment. Experiments across multiple reward model configurations show that IR3 achieves 0.89 correlation with ground-truth rewards, identifies hacking features with over 90% precision, and significantly reduces hacking behaviors while maintaining capabilities within 3% of the original model.

Mohammad Beigi, Ming Jin, Junshan Zhang +3

UC Davis · Virginia Tech · Salesforce +1

Reward Hacking Reinforcement Learning Interpretability in RL RL from Human Feedback

Feb 20, 2026

Enhancing Goal Inference via Correction Timing

Corrections offer a natural modality for people to provide feedback to a robot, by (i) intervening in the robot's behavior when they believe the robot is failing (or will fail) the task objectives and (ii) modifying the robot's behavior to successfully fulfill the task. Each correction offers information on what the robot should and should not do, where the corrected behavior is more aligned with task objectives than the original behavior. Most prior work on learning from corrections involves interpreting a correction as a new demonstration (consisting of the modified robot behavior), or a preference (for the modified trajectory compared to the robot's original behavior). However, this overlooks one essential element of the correction feedback, which is the human's decision to intervene in the robot's behavior in the first place. This decision can be influenced by multiple factors including the robot's task progress, alignment with human expectations, dynamics, motion legibility, and optimality. In this work, we investigate whether the timing of this decision can offer a useful signal for inferring these task-relevant influences. In particular, we investigate three potential applications for this learning signal: (1) identifying features of a robot's motion that may prompt people to correct it, (2) quickly inferring the final goal of a human's correction based on the timing and initial direction of their correction motion, and (3) learning more precise constraints for task objectives. Our results indicate that correction timing results in improved learning for the first two of these applications. Overall, our work provides new insights on the value of correction timing as a signal for robot learning.

Anjiabei Wang, Shuangge Wang, Tesca Fitzgerald

Yale University, New Haven, CT, United States

Robot Learning Human-Robot Collaboration RL from Human Feedback

Feb 20, 2026

Gradient Regularization Mitigates Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards

Reinforcement Learning from Human Feedback (RLHF) or Verifiable Rewards (RLVR) are two key steps in the post-training of modern Language Models (LMs). A common problem is reward hacking, where the policy may exploit inaccuracies of the reward and learn an unintended behavior. Most previous works address this by limiting the policy update with a Kullback-Leibler (KL) penalty towards a reference model. We propose a different framing: Train the LM in a way that biases policy updates towards regions in which the reward is more accurate. First, we derive a theoretical connection between the accuracy of a reward model and the flatness of an optimum at convergence. Gradient regularization (GR) can then be used to bias training to flatter regions and thereby maintain reward model accuracy. We confirm these results by showing that the gradient norm and reward accuracy are empirically correlated in RLHF. We then empirically show that Reference Resets of the KL penalty find flatter regions with a higher reward accuracy. We further improve on this by proposing to use explicit GR with an efficient finite-difference estimate. Empirically, GR performs better than a KL penalty across a diverse set of RL experiments with LMs. GR achieves a higher GPT-judged win-rate in RLHF, avoids overly focusing on the format in rule-based math rewards, and prevents hacking the judge in LLM-as-a-Judge math tasks.

Johannes Ackermann, Michael Noukhovitch, Takashi Ishida +1

The University of Tokyo · Mila · RIKEN AIP

Reward Hacking Reinforcement Learning with Verifiable Rewards Policy Learning RL from Human Feedback Regularized RL

Feb 7, 2026

Joint Reward Modeling: Internalizing Chain-of-Thought for Efficient Visual Reward Models

Reward models are critical for reinforcement learning from human feedback, as they determine the alignment quality and reliability of generative models. For complex tasks such as image editing, reward models are required to capture global semantic consistency and implicit logical constraints beyond local similarity. Existing reward modeling approaches have clear limitations. Discriminative reward models align well with human preferences but struggle with complex semantics due to limited reasoning supervision. Generative reward models offer stronger semantic understanding and reasoning, but they are costly at inference time and difficult to align directly with human preferences. To this end, we propose Joint Reward Modeling (JRM), which jointly optimizes preference learning and language modeling on a shared vision-language backbone. This approach internalizes the semantic and reasoning capabilities of generative models into efficient discriminative representations, enabling fast and accurate evaluation. JRM achieves state-of-the-art results on MMRB2 and EditReward-Bench, and significantly improves stability and performance in downstream online reinforcement learning. These results show that joint training effectively bridges efficiency and semantic understanding in reward modeling.

Yankai Yang, Yancheng Long, Hongyang Wei +12

Harbin Institute of Technology, Shenzhen. · Kuaishou Technology.

Reward Modeling Multimodal Reward Modeling VLM Reasoning RL from Human Feedback

Feb 2, 2026

Adversarial Reward Auditing for Active Detection and Mitigation of Reward Hacking

Reinforcement Learning from Human Feedback (RLHF) remains vulnerable to reward hacking, where models exploit spurious correlations in learned reward models to achieve high scores while violating human intent. Existing mitigations rely on static defenses that cannot adapt to novel exploitation strategies. We propose Adversarial Reward Auditing (ARA), a framework that reconceptualizes reward hacking as a dynamic, competitive game. ARA operates in two stages: first, a Hacker policy discovers reward model vulnerabilities while an Auditor learns to detect exploitation from latent representations; second, Auditor-Guided RLHF (AG-RLHF) gates reward signals to penalize detected hacking, transforming reward hacking from an unobservable failure into a measurable, controllable signal. Experiments across three hacking scenarios demonstrate that ARA achieves the best alignment-utility tradeoff among all baselines: reducing sycophancy to near-SFT levels while improving helpfulness, decreasing verbosity while achieving the highest ROUGE-L, and suppressing code gaming while improving Pass@1. Beyond single-domain evaluation, we show that reward hacking, detection, and mitigation all generalize across domains -- a Hacker trained on code gaming exhibits increased sycophancy despite no reward for this behavior, and an Auditor trained on one domain effectively suppresses exploitation in others, enabling efficient multi-domain defense with a single model.

Mohammad Beigi, Ming Jin, Junshan Zhang +2

Department of Computer Science, University of California, Davis, USA · Virginia Tech, Blacksburg, USA · Meta AI Research.

Adversarial Training Reward Hacking RL from Human Feedback

Sep 26, 2025

Adaptive Margin RLHF via Preference over Preferences

Margin-based optimization is fundamental to improving generalization and robustness in classification tasks. In the context of reward model learning from preferences within Reinforcement Learning from Human Feedback (RLHF), existing methods typically rely on no margins, fixed margins, or margins that are simplistic functions of preference ratings. However, such formulations often fail to account for the varying strengths of different preferences or they rely on noisy margin information derived from preference ratings. Furthermore, many existing methods that use adaptive margins assume access to accurate preference scores, which can be difficult for humans to provide reliably. We propose leveraging preferences over preferences, that is, annotations indicating which of two preferences reflects a stronger distinction, to infer adaptive margins on a per-datapoint basis. Such preference-over-preference annotations are general and can be incorporated into both standard RLHF reward modeling objectives and direct alignment losses. As a concrete instantiation, we introduce DPO-PoP, an extension to Direct Preference Optimization (DPO) that incorporates adaptive margins from preference-over-preference supervision, enabling improved discriminative and generative performance. Additionally, we show a tradeoff between discriminative and generative performance and propose two sampling strategies for gathering preference-over-preference labels to navigate it.

Yaswanth Chittepu, Prasann Singhal, Greg Durrett +1

University of Massachusetts Amherst · University of Texas at Austin

Pairwise Preference Learning Reward Modeling Direct Preference Optimization Preference Learning RL from Human Feedback

Sep 18, 2025

Aligning Audio Captions with Human Preferences

Current audio captioning relies on supervised learning with paired audio-caption data, which is costly to curate and may not reflect human preferences in real-world scenarios. To address this, we propose a preference-aligned audio captioning framework based on Reinforcement Learning from Human Feedback (RLHF). To capture nuanced preferences, we train a Contrastive Language-Audio Pretraining (CLAP) based reward model using human-labeled pairwise preference data. This reward model is integrated into an RL framework to fine-tune any baseline captioning system without ground-truth annotations. Extensive human evaluations across multiple datasets show that our method produces captions preferred over baseline models, particularly when baselines fail to provide correct and natural captions. Furthermore, our framework achieves performance comparable to supervised approaches with ground-truth data, demonstrating effective alignment with human preferences and scalability in real-world use.

Kartik Hegde, Rehana Mahfuz, Yinyi Guo +1

Qualcomm Technologies, Inc., India · Qualcomm Technologies, Inc., USA

Reward Modeling Audio-Language Models Audio Captioning Preference Alignment RL from Human Feedback

Aug 4, 2025

TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback

Group Relative Policy Optimization (GRPO) trains language models without a value critic using rewards centered within sampled response groups. We study how importance weighting, clipping, and length normalization affect its stochastic updates and propose Trajectory-level Importance-Corrected GRPO (TIC-GRPO), combining an upper importance-ratio cap with a full-trajectory likelihood ratio. Under common maximum-length normalization, a trajectory change of measure gives a score second moment proportional to the response horizon. We propagate this estimate through the actual empirical baseline and all inner updates. With separately optimized constant steps, the sampling term in TIC's stationarity bound has a linear horizon coefficient, versus order three halves for token-level GRPO

_2

. Matching upper and lower bounds separate the capped methods' worst-case response-level variances by a factor

T

. The bounds track prompt count, response-group size, vocabulary size, reward scale, and score regularity; per-response-normalized GRPO also retains a length-covariance term. Experiments at two Qwen3 scales on four reasoning and coding benchmarks evaluate both mechanisms in sampled-token-normalized training.

Lei Pang, Jun Luo, Ruinan Jin

Peking University Beijing, China · The Ohio State University Columbus, USA · Shugu Yuntai Technology Co., Ltd. +1

Importance Sampling for RL RL for Language Models Group Relative Policy Optimization Policy Optimization RL from Human Feedback