Reinforcement Learning From Human Feedback

This preprint presents a systematic, research-oriented practicum that guides the reader through the entire modern NLP pipeline: from tokenisation and vectorisation to fine-tuning of large language models, retrieval-augmented generation, and reinforcement learning from human feedback. A distinctive feature of the work is its consistent attention to low-resource and morphologically rich languages -- original contributions on Tajik and Tatar, including subword tokenisers, word embeddings, lexical databases, and transliteration benchmarks, are woven throughout the twelve sessions, demonstrating how modern NLP can be adapted to data-scarce environments without sacrificing rigour. Each session combines concise theory with detailed implementation plans, formalised evaluation metrics, and transparent assessment criteria. The work is not a conventional textbook: it is designed as a reproducible research artefact where every session requires publishing code, models, and reports in public repositories. All experiments are conducted on a single evolving corpus, and the work advocates open-weight models over commercial APIs, with special attention to the Hugging Face ecosystem. Designed for senior undergraduates, graduate students, and practising developers seeking to implement, compare, and deploy methods from classical ML to state-of-the-art LLM-based systems.

Offline Reinforcement Learning from Human Feedback (RLHF) pipelines such as Direct Preference Optimization (DPO) train on a pre-collected preference dataset, which makes them vulnerable to preference poisoning attack. We study label flip attacks against log-linear DPO. We first illustrate that flipping one preference label induces a parameter-independent shift in the DPO gradient. Using this key property, we can then convert the targeted poisoning problem into a structured binary sparse approximation problem. To solve this problem, we develop two attack methods: Binary-Aware Lattice Attack (BAL-A) and Binary Matching Pursuit Attack (BMP-A). BAL-A embeds the binary flip selection problem into a binary-aware lattice and applies Lenstra-Lenstra-Lovász reduction and Babai's nearest plane algorithm; we provide sufficient conditions that enforce binary coefficients and recover the minimum-flip objective. BMP-A adapts binary matching pursuit to our non-normalized gradient dictionary and yields coherence-based recovery guarantees and robustness (impossibility) certificates for $K$-flip budgets. Experiments on synthetic dictionaries and the Stanford Human Preferences dataset validate the theory and highlight how dictionary geometry governs attack success.

Reinforcement Learning from Human Feedback (RLHF) has become a key post-training paradigm for improving model quality. However, the synchronous three-stage RLHF pipeline is often bottlenecked by the generation stage, where response-length skew causes the effective batch size to shrink rapidly during decoding, leaving GPUs underutilized while a few long responses remain unfinished. Mainstream frameworks employ a static tensor parallelism (TP) configuration that cannot adapt to changing batch characteristics, leaving substantial performance headroom unexplored. We propose PAT, an adaptive TP method that dynamically reconfigures TP during the generation stage of each RLHF iteration. PAT introduces two key techniques. First, a predictor-guided online reconfiguration method decides both the reconfiguration point and the target TP configuration based on offline profiling, triggering reconfiguration only when the predicted latency benefit outweighs the reconfiguration overhead. Second, a lightweight online reconfiguration mechanism updates only the states and layouts affected by TP changes: it adapts unfinished decoding states through a cost-model-based choice between KV-cache migration and recomputation, performs in-place weight resharding, and reuses cached communication groups. We implement PAT on top of SGLang and integrate it with the VeRL framework. Evaluations on LLaMA3.1-8B and Qwen3-14B using DeepScaleR show that PAT reduces generation latency by up to 34.6% and end-to-end RLHF training iteration latency by up to 27.2% compared to the original VeRL setup.

Reinforcement learning from human feedback (RLHF) is a central post-training tool for aligning large language models, but its training reward is only a learned proxy for true human utility. This creates a decision problem under objective misspecification: the policy is optimized against an estimated reward, while deployment performance is governed by an unobserved population preference. The resulting gap leads to reward over-optimization, where proxy reward keeps improving after true quality deteriorates. We propose distributionally robust regret optimization (DRRO) for RLHF with a Wasserstein ambiguity set over reward laws, using promptwise $\ell_p$ distances between reward vectors as transport costs. Unlike standard distributionally robust optimization, which pessimizes worst-case value, DRRO pessimizes worst-case regret relative to the best policy under the same plausible reward perturbation. We show that the expressive-policy problem decomposes into promptwise regret problems. For each prompt, the inner adversary has a dual-norm closed form; under the $\ell_1$ transport cost used by our algorithm, the optimizer has a water-filling structure. These results lead to a practical policy-gradient algorithm that adds a simple sampled bonus to GRPO-style training. Theory and experiments both show that DRRO is less over-pessimistic than standard DRO and mitigates over-optimization more effectively than existing baselines.

Reinforcement learning from human feedback (RLHF) systems face a compounding alignment challenge: not only are learned reward models uncertain about unseen state-action pairs, but the human preference annotations they are trained on are themselves inconsistent, context-dependent, and noisy. Existing approaches address these uncertainty sources in isolation - epistemic uncertainty is used to guide exploration, while preference uncertainty is absorbed during reward model training but discarded during policy optimization. We introduce Uncertainty-Aware Reward Discounting (UARD), a principled framework that jointly models epistemic uncertainty in value estimation via ensemble disagreement and aleatoric uncertainty in human preference annotations via annotator variability, combining these signals through a confidence-adjusted Reliability Filter that adaptively modulates reward weighting during policy optimization. We prove that this dynamic discounting preserves the contraction property of the Bellman operator, guaranteeing convergence to a unique fixed point, and provide an information-theoretic justification grounded in the Information Bottleneck principle. Empirically, UARD reduces reward hacking incidents by up to 93.6% across discrete decision-making and continuous control benchmarks (MuJoCo) compared to nine baselines including DQN, Ensemble-DQN, CQL, CPO, TRPO, SAC, EDAC, SUNRISE, and PPO, while maintaining competitive task performance on well-specified rewards. Under annotation noise ranging from 10% to 30% Gaussian perturbation, UARD retains near-zero safety violations compared to baselines' near-linear degradation. These results demonstrate that treating uncertainty as an active component of the optimization objective - rather than a passive diagnostic signal - provides a principled pathway toward more reliable and aligned RL systems.

Preference-based alignment methods, most prominently Reinforcement Learning with Human Feedback (RLHF), use the judgments of human annotators to shape large language model behaviour. However, the normative role of these judgments is rarely made explicit. I distinguish three conceptual models of that role. The first is extension: annotators extend the system designers' own judgments about what outputs should be. The second is evidence: annotators provide independent evidence about some facts, whether moral, social or otherwise. The third is authority: annotators have some independent authority (as representatives of the broader population) to determine system outputs. I argue that these models have implications for how RLHF pipelines should solicit, validate and aggregate annotations. I survey landmark papers in the literature on RLHF and related methods to illustrate how they implicitly draw on these models, describe failure modes that come from unintentionally or intentionally conflating them, and offer normative criteria for choosing among them. My central recommendation is that RLHF pipeline designers should decompose annotation into separable dimensions and tailor each pipeline to the model most appropriate for that dimension, rather than seeking a single unified pipeline.

Unsupervised skill discovery in reinforcement learning aims to intrinsically motivate agents to discover diverse and useful behaviours. However, unconstrained approaches can produce unsafe, unethical, or misaligned behaviours. To mitigate these risks and improve the practical desireability of discovered skills, recent work grounds the discovery process by leveraging human preference feedback. However, preference-based approaches are feedback-inefficient and inherently ill-equipped to deal with skill spaces composed of a variety of different skills such as running, jumping, walking, etc. To overcome this limitation, we introduce semantic labelling, a novel and feedback-efficient approach that leverages human cognitive strengths to identify and label semantically meaningful behaviours. Based on semantic labelling, we propose Semantically Relevant Skill Discovery (SRSD), a novel human-in-the-loop approach that collects semantic labels from human feedback and learns a reward function to encourage skills to be more semantically diverse and relevant. Through our experiments in a 2D navigation environment and four locomotion environments, we demonstrate that SRSD can improve semantic diversity and discover relevant behaviours while scaling effectively to a large variety of behaviours.

Safe Reinforcement Learning from Human Feedback (Safe RLHF) has recently achieved empirical success in developing helpful and harmless large language models by decoupling human preferences regarding helpfulness and harmlessness. Existing approaches typically rely on fitting fixed horizon reward models from human feedback and have only been validated empirically. In this paper, we formulate safe RLHF as an infinite horizon discounted Con- strained Markov Decision Process (CMDP), since humans may interact with the model over a continuing sequence of interactions rather than within a single finite episode. We propose two Safe RLHF algorithms that do not require reward model fitting and, in contrast to prior work assuming fixed-length trajectories, support flexible trajectory lengths for training. Both algo- rithms are based on the primal-dual method and achieve global convergence guarantees with polynomial rates in terms of policy gradient iterations, trajectory sample lengths, and human preference queries. To the best of our knowledge, this is the first work to study infinite horizon discounted CMDP under human feedback and establish global, non-asymptotic convergence.

Reinforcement Learning from Human Feedback (RLHF) is central to aligning Large Language Models (LLMs), yet it introduces a critical vulnerability: an imperfect Reward Model (RM) can become a single point of failure when it fails to penalize unsafe behaviors. While existing red-teaming approaches primarily target policy-level weaknesses, they overlook what we term systemic weaknesses cases where both the core LLM and the RM fail in tandem. We present ARES, a framework that systematically discovers and mitigates such dual vulnerabilities. ARES employs a ``Safety Mentor'' that dynamically composes semantically coherent adversarial prompts by combining structured component types (topics, personas, tactics, goals) and generates corresponding malicious and safe responses. This dual-targeting approach exposes weaknesses in both the core LLM and the RM simultaneously. Using the vulnerabilities gained, ARES implements a two-stage repair process: first fine-tuning the RM to better detect harmful content, then leveraging the improved RM to optimize the core model. Experiments across multiple adversarial safety benchmarks demonstrate that ARES substantially enhances safety robustness while preserving model capabilities, establishing a new paradigm for comprehensive RLHF safety alignment.

Ensuring the safety of large language models (LLMs) requires robust red teaming, yet the systematic synthesis of high-quality toxic data remains under-explored. We propose Reverse Constitutional AI (R-CAI), a framework for automated and controllable adversarial data generation that moves beyond isolated jailbreak prompts. By inverting a harmless constitution into a constitution of toxicity and iteratively refining model outputs through a critique--revision pipeline, R-CAI enables scalable synthesis of multi-dimensional adversarial data without human annotation. Optimizing solely for toxicity-related rewards, however, can lead to reward hacking and degraded semantic coherence. To address this challenge, we introduce probability clamping within reinforcement learning from AI feedback, which stabilizes adversarial optimization while preserving adversarial intent. Experiments demonstrate that R-CAI generates diverse, high-quality toxic data and that probability clamping substantially improves semantic coherence (15%) without sacrificing adversarial strength. Overall, R-CAI provides a fully automated framework for red teaming data generation and systematic safety evaluation of aligned language models.

This paper considers reinforcement learning from human feedback in a federated learning setting with resource-constrained agents, such as edge devices. We propose an efficient federated RLHF algorithm, named Partitioned, Sign-based Stochastic Zeroth-order Policy Optimization (Par-S$^2$ZPO). The algorithm is built on zeroth-order optimization with binary perturbation, resulting in low communication, computation, and memory complexity by design. Our theoretical analysis establishes an upper bound on the convergence rate of Par-S$^2$ZPO, revealing that it is as efficient as its centralized counterpart in terms of sample complexity but converges faster in terms of policy update iterations. Our experimental results show that it outperforms a FedAvg-based RLHF on four MuJoCo RL tasks.

Safe Reinforcement Learning from Human Feedback (RLHF) typically enforces safety through expected cost constraints, but the expectation captures only a single statistic of the cost distribution and fails to account for distributional uncertainty, particularly under heavy tails or rare catastrophic events. This limitation is problematic when robustness and risk sensitivity are critical. Stochastic dominance offers a principled alternative by comparing entire cost distributions rather than just their averages, enabling direct control over tail risks and potential out-of-distribution failures that expectation-based constraints may overlook. In this work, we propose Risk-sensitive Alignment via Dominance (RAD), a novel alignment framework that replaces scalar expected cost constraints with First-Order Stochastic Dominance (FSD) constraints. We operationalize this constraint by comparing the target policy's cost distribution to that of a reference policy within an Optimal Transport (OT) framework, using entropic regularization and Sinkhorn iterations to obtain a differentiable and computationally efficient objective for stable end-to-end optimization. Furthermore, we introduce quantile-weighted FSD constraints and show that weighted FSD universally controls a broad class of Spectral Risk Measures (SRMs), so that improvements under weighted dominance imply guaranteed improvements in the corresponding spectral risk. This provides a principled mechanism for tuning a model's risk profile via the quantile weighting function. Empirical results demonstrate that RAD improves harmlessness over baselines while remaining competitive in helpfulness, and exhibits greater robustness on out-of-distribution harmlessness evaluations.

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.

Practical LLM agents often operate over multi-turn conversations where success is determined only after the full interaction ends. Most multi-turn RL methods train via on-policy rollouts, but unlike in single-turn RLHF, the policy cannot produce a trajectory alone, since an external environment must respond after each agent turn. For conversational agents, this environment is a user, but real users are generally unavailable inside the training loop and simulated users are difficult to build faithfully. Separately, real-world deployment is rarely fully online or fully offline. The common production pattern is called "batch online," where the current policy is deployed to collect a batch of interaction data, then retrained on that batch and redeployed. We show that this batch-online setting creates an opportunity for applying the classical approximate policy iteration algorithm. Our central observation is that running standard token-level, single-turn GRPO with a learned turn-level $Q$-function as the reward model is a policy improvement step for the multi-turn problem. Building on this, we present Iterative GRPO, which alternates between fitting $Q^π$ from logged Monte Carlo returns via standard reward modeling (policy evaluation) and running single-turn GRPO against $Q^π$ (policy improvement). During policy improvement, the learned $Q^π$ scores candidate responses by their expected downstream return, avoiding the need for simulator rollouts of the remaining conversation. Overall, this allows us to do principled multi-turn RL using only single-turn RLHF methods without an interactive environment or user simulator inside the training loop. We demonstrate the effectiveness of Iterative GRPO on six multi-turn negotiation environments.

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.

This survey offers a comprehensive foundation on the integration of RL with language models, highlighting prominent algorithms such as Proximal Policy Optimization (PPO), Q-Learning, and Actor-Critic methods. Additionally, it provides an extensive technical overview of RL techniques specifically tailored for LLMs, including foundational methods like Reinforcement Learning from Human Feedback (RLHF) and AI Feedback (RLAIF), as well as advanced strategies such as Direct Preference Optimization (DPO) and Group Relative Policy Optimization (GRPO). We systematically analyze their applications across domains, i.e., from code generation to tool-augmented reasoning. Crucially, we move beyond descriptive categorization to provide a rigorous algorithmic analysis of failure modes, mathematically framing the structural bottlenecks and stability trade-offs inherent in policy optimization. We also present a comparative taxonomy based on reward modeling, feedback mechanisms, and optimization strategies. Our evaluation highlights key trends. RLHF remains dominant for alignment, and outcome-based RL such as Reinforcement Learning with Verifiable Rewards (RLVR) significantly improves stepwise reasoning. However, persistent challenges such as reward hacking, computational costs, and scalable feedback collection underscore the need for continued innovation. We also explicate the causal factors behind recent benchmark performances, distinguishing between gains derived from architectural scaling versus those stemming from specific optimization objectives. We further discuss emerging directions, including hybrid RL algorithms, verifier-guided training, and multi-objective alignment frameworks. This survey serves as a roadmap for researchers advancing RL-driven LLM development, balancing capability enhancement with safety and scalability.

Reinforcement Learning from Human Feedback (RLHF) 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 significant example of such reward hacking is length bias, where reward models usually favor longer responses irrespective of actual response quality. Previous works on tackling length bias have notable limitations, these approaches either mitigate bias without characterizing the bias form, or simply assume a linear length-reward relation. To accurately model the intricate nature of length bias and facilitate more effective bias mitigation, we propose FiMi-RM (Bias Fitting to Mitigate Length Bias of Reward Model), a framework that autonomously learns and corrects underlying bias patterns. Our approach consists of three stages: First, we warm up by training a standard reward model which inherently contains length bias. Next, we deploy a lightweight fitting model to capture the non-linear relation between length and reward. Finally, we incorporate this learned relation into the reward model, effectively decoupling length from reward while preserving preference modeling capabilities. Experimental results demonstrate that FiMi-RM achieves a more balanced length-reward distribution. Furthermore, when applied to alignment algorithms such as Direct Preference Optimization (DPO) and Best-of-N (BoN), our debiased reward model improves length-controlled win rate and reduces verbosity without compromising its performance.

Reinforcement learning from human feedback (RLHF) has become a crucial tool to build the latest machine learning systems at scale. The field grew around the core methods of RLHF into today's broader suite of post-training techniques. In this book, we give a comprehensive introduction to the core methods for post-training models for people with some level of quantitative background, organized around the canonical RLHF recipe. The book starts with what RLHF does and why it was created, with seminal technical milestones in its young history and a primer on reinforcement learning context needed to understand the book. The core of the book details every optimization stage in using RLHF, from starting with instruction tuning to training a reward model and finally all of rejection sampling, reinforcement learning, on-policy distillation, and direct alignment algorithms. The book also discusses broader topics, such as the origins of RLHF -- both in recent literature and in a convergence of disparate fields of science in economics, philosophy, and optimal control. The book concludes with advanced topics -- understudied or emerging research questions in synthetic data, tool-use, character training, and evaluation -- and open questions for the field. The book is released with a variety of companion resources, including a codebase, a library to compare model completions from within post-training stages, and an educational course, to be a one-stop shop for learning all foundational concepts for post-training language models.

Reinforcement learning from human feedback (RLHF) has evolved to be one of the main methods for fine-tuning large language models (LLMs). However, existing RLHF methods are non-robust, and their performance deteriorates if the downstream task differs significantly from the preference dataset used in fine-tuning. In order to mitigate this problem, we introduce a distributionally robust RLHF for fine-tuning LLMs. In particular, our goal is to ensure that a fine-tuned model retains its performance even when the distribution of prompts significantly differs from the distribution encountered during fine-tuning. We formulate distributionally robust optimization (DRO) version of two popular fine-tuning methods -- (1) reward-based RLHF and (2) reward-free DPO (direct preference optimization). We propose a minibatch gradient descent based algorithms for both of them, and theoretically prove convergence guarantees for the algorithms. Subsequently, we evaluate our algorithms on an out-of-distribution (OOD) task by first training the model on the Unified-Feedback dataset and evaluating its performance on two different datasets. The experimental results show that our robust training improves the accuracy of the learned reward models on average, and markedly on some tasks, such as reasoning. Furthermore, we show that the robust versions of policy optimization methods, similarly improve performance on OOD tasks.

We study data corruption robustness for reinforcement learning with human feedback (RLHF) in an offline setting. Given an offline dataset of pairs of trajectories along with feedback about human preferences, an $\varepsilon$-fraction of the pairs is corrupted (e.g., feedback flipped or trajectory features manipulated), capturing an adversarial attack or noisy human preferences. We aim to design algorithms that identify a near-optimal policy from the corrupted data, with provable guarantees. Existing theoretical works have separately studied the settings of corruption robust RL (learning from scalar rewards directly under corruption) and offline RLHF (learning from human feedback without corruption); however, they are inapplicable to our problem of dealing with corrupted data in offline RLHF setting. To this end, we design novel corruption robust offline RLHF methods under various assumptions on the coverage of the data-generating distributions. At a high level, our methodology robustifies an offline RLHF framework by first learning a reward model along with confidence sets and then learning a pessimistic optimal policy over the confidence set. Our key insight is that learning optimal policy can be done by leveraging an offline corruption-robust RL oracle in different ways (e.g., zero-order oracle or first-order oracle), depending on the data coverage assumptions. To our knowledge, ours is the first work that provides provable corruption robust offline RLHF methods.