Label-flip attacks on log-linear DPO reduce to binary sparse approximation problems that can be solved efficiently by lattice-based and binary matching pursuit methods with recovery guarantees.
hub
Fine-Tuning Language Models from Human Preferences
80 Pith papers cite this work. Polarity classification is still indexing.
80
Pith papers citing it
abstract
Reward learning enables the application of reinforcement learning (RL) to tasks where reward is defined by human judgment, building a model of reward by asking humans questions. Most work on reward learning has used simulated environments, but complex information about values is often expressed in natural language, and we believe reward learning for language is a key to making RL practical and safe for real-world tasks. In this paper, we build on advances in generative pretraining of language models to apply reward learning to four natural language tasks: continuing text with positive sentiment or physically descriptive language, and summarization tasks on the TL;DR and CNN/Daily Mail datasets. For stylistic continuation we achieve good results with only 5,000 comparisons evaluated by humans. For summarization, models trained with 60,000 comparisons copy whole sentences from the input but skip irrelevant preamble; this leads to reasonable ROUGE scores and very good performance according to our human labelers, but may be exploiting the fact that labelers rely on simple heuristics.
hub tools
claims ledger
- abstract Reward learning enables the application of reinforcement learning (RL) to tasks where reward is defined by human judgment, building a model of reward by asking humans questions. Most work on reward learning has used simulated environments, but complex information about values is often expressed in natural language, and we believe reward learning for language is a key to making RL practical and safe for real-world tasks. In this paper, we build on advances in generative pretraining of language models to apply reward learning to four natural language tasks: continuing text with positive sentimen
co-cited works
representative citing papers
GPT-3 shows that scaling an autoregressive language model to 175 billion parameters enables strong few-shot performance across diverse NLP tasks via in-context prompting without fine-tuning.
Temperature adjustment on the reference model generalizes inference-time alignment to SLOP ensembles of reward models, with a calibration algorithm that improves robustness to reward hacking while preserving alignment performance.
Reinforce Adjoint Matching derives a simple consistency loss for RL post-training of diffusion models by tilting the clean distribution toward higher-reward samples under KL regularization while keeping the noising process fixed, achieving superior rewards in far fewer steps than prior methods.
For convex losses in nested evolving feasible sets, a lazy algorithm balances O(T^{1-β}) regret with O(T^β) movement for any β; for strongly convex or sharp losses, Frugal achieves zero regret with O(log T) movement, shown optimal by matching lower bound.
RL on binary rewards boosts LLM factual recall by ~27% relative across models by redistributing probability mass to latent correct answers rather than acquiring new knowledge.
The paper establishes the first O(log T) regret and O(1/T) sub-optimality bounds for online RLHF under general f-divergence regularization via two sampling algorithms.
Introduces the first benchmark for fine-grained failures in reinforcement fine-tuning of LLMs and an automatic management framework that detects, diagnoses, and remediates them.
Reference-sampled weighted SFT with prompt-normalized Boltzmann weights induces the same policy as fixed-reference KL-regularized RLVR, with BOLT as the estimator and a finite one-shot error decomposition separating coverage, variance, and other terms.
The work establishes a regret lower bound of Ω(T^{2/3} min(K,D)^{1/3}) for fair multi-user dueling bandits with heterogeneous Condorcet winners and gives algorithms achieving matching upper bounds up to logs.
RLHF should decompose annotations into dimensions each matched to one of three models—extension, evidence, or authority—instead of applying a single unified pipeline.
Introduces an interactive episodic memory task with user feedback and a Feedback Alignment Module that improves retrieval accuracy on video benchmarks while remaining efficient.
R-CAI inverts constitutional AI to automatically generate diverse toxic data for LLM red teaming, with probability clamping improving output coherence by 15% while preserving adversarial strength.
Reinforcement learning with a multi-part reward teaches LLMs to output independent, meaning-preserving sentence edits that raise argument appropriateness close to full rewriting.
E2E-REME outperforms nine LLMs in accuracy and efficiency for end-to-end microservice remediation by using experience-simulation reinforcement fine-tuning on a new benchmark called MicroRemed.
Open source AI shows lower collaboration intensity, reduced direct contributions, and a shift toward adaptive use rather than joint improvement compared to traditional OSS.
GiGPO adds a hierarchical grouping mechanism to group-based RL so that LLM agents receive both global trajectory and local step-level credit signals, yielding >12% gains on ALFWorld and >9% on WebShop over GRPO while keeping the same rollout and memory footprint.
KTO aligns LLMs by directly maximizing prospect-theoretic utility on binary signals and matches or exceeds preference-based methods like DPO from 1B to 30B parameters.
Medusa augments LLMs with multiple decoding heads and tree-based attention to predict and verify several tokens in parallel, yielding 2.2-3.6x inference speedup via two fine-tuning regimes.
Iterative self-rewarding via LLM-as-Judge in DPO training on Llama 2 70B improves instruction following and self-evaluation, outperforming GPT-4 on AlpacaEval 2.0.
Chain-of-Thought reasoning in LLMs is often unfaithful, with models relying on it variably by task and less so as models scale larger.
Process supervision significantly outperforms outcome supervision for training models on the MATH dataset, achieving 78% accuracy on a representative test subset with active learning and a released 800k step-label dataset.
One language model can generate diverse test cases to automatically uncover tens of thousands of harmful behaviors, including offensive replies and privacy leaks, in a large target language model.
DIAL uses intent-conditioned CFG and multi-intent GRPO to expand and preserve diverse modes in continuous-action preference RL, lifting RFS to 9.14 and surpassing both prior best (8.5) and human demonstration (8.13).
citing papers explorer
-
Efficient Preference Poisoning Attack on Offline RLHF
Label-flip attacks on log-linear DPO reduce to binary sparse approximation problems that can be solved efficiently by lattice-based and binary matching pursuit methods with recovery guarantees.
-
Language Models are Few-Shot Learners
GPT-3 shows that scaling an autoregressive language model to 175 billion parameters enables strong few-shot performance across diverse NLP tasks via in-context prompting without fine-tuning.
-
Temper and Tilt Lead to SLOP: Reward Hacking Mitigation with Inference-Time Alignment
Temperature adjustment on the reference model generalizes inference-time alignment to SLOP ensembles of reward models, with a calibration algorithm that improves robustness to reward hacking while preserving alignment performance.
-
Reinforce Adjoint Matching: Scaling RL Post-Training of Diffusion and Flow-Matching Models
Reinforce Adjoint Matching derives a simple consistency loss for RL post-training of diffusion models by tilting the clean distribution toward higher-reward samples under KL regularization while keeping the noising process fixed, achieving superior rewards in far fewer steps than prior methods.
-
Convex Optimization with Nested Evolving Feasible Sets
For convex losses in nested evolving feasible sets, a lazy algorithm balances O(T^{1-β}) regret with O(T^β) movement for any β; for strongly convex or sharp losses, Frugal achieves zero regret with O(log T) movement, shown optimal by matching lower bound.
-
Beyond Reasoning: Reinforcement Learning Unlocks Parametric Knowledge in LLMs
RL on binary rewards boosts LLM factual recall by ~27% relative across models by redistributing probability mass to latent correct answers rather than acquiring new knowledge.
-
$f$-Divergence Regularized RLHF: Two Tales of Sampling and Unified Analyses
The paper establishes the first O(log T) regret and O(1/T) sub-optimality bounds for online RLHF under general f-divergence regularization via two sampling algorithms.
-
Towards Robust LLM Post-Training: Automatic Failure Management for Reinforcement Fine-Tuning
Introduces the first benchmark for fine-grained failures in reinforcement fine-tuning of LLMs and an automatic management framework that detects, diagnoses, and remediates them.
-
Reference-Sampled Boltzmann Projection for KL-Regularized RLVR: Target-Matched Weighted SFT, Finite One-Shot Gaps, and Policy Mirror Descent
Reference-sampled weighted SFT with prompt-normalized Boltzmann weights induces the same policy as fixed-reference KL-regularized RLVR, with BOLT as the estimator and a finite one-shot error decomposition separating coverage, variance, and other terms.
-
Multi-User Dueling Bandits: A Fair Approach using Nash Social Welfare
The work establishes a regret lower bound of Ω(T^{2/3} min(K,D)^{1/3}) for fair multi-user dueling bandits with heterogeneous Condorcet winners and gives algorithms achieving matching upper bounds up to logs.
-
Three Models of RLHF Annotation: Extension, Evidence, and Authority
RLHF should decompose annotations into dimensions each matched to one of three models—extension, evidence, or authority—instead of applying a single unified pipeline.
-
Interactive Episodic Memory with User Feedback
Introduces an interactive episodic memory task with user feedback and a Feedback Alignment Module that improves retrieval accuracy on video benchmarks while remaining efficient.
-
Reverse Constitutional AI: A Framework for Controllable Toxic Data Generation via Probability-Clamped RLAIF
R-CAI inverts constitutional AI to automatically generate diverse toxic data for LLM red teaming, with probability clamping improving output coherence by 15% while preserving adversarial strength.
-
Teaching LLMs Human-Like Editing of Inappropriate Argumentation via Reinforcement Learning
Reinforcement learning with a multi-part reward teaches LLMs to output independent, meaning-preserving sentence edits that raise argument appropriateness close to full rewriting.
-
E2E-REME: Towards End-to-End Microservices Auto-Remediation via Experience-Simulation Reinforcement Fine-Tuning
E2E-REME outperforms nine LLMs in accuracy and efficiency for end-to-end microservice remediation by using experience-simulation reinforcement fine-tuning on a new benchmark called MicroRemed.
-
From OSS to Open Source AI: an Exploratory Study of Collaborative Development Paradigm Divergence
Open source AI shows lower collaboration intensity, reduced direct contributions, and a shift toward adaptive use rather than joint improvement compared to traditional OSS.
-
Group-in-Group Policy Optimization for LLM Agent Training
GiGPO adds a hierarchical grouping mechanism to group-based RL so that LLM agents receive both global trajectory and local step-level credit signals, yielding >12% gains on ALFWorld and >9% on WebShop over GRPO while keeping the same rollout and memory footprint.
-
KTO: Model Alignment as Prospect Theoretic Optimization
KTO aligns LLMs by directly maximizing prospect-theoretic utility on binary signals and matches or exceeds preference-based methods like DPO from 1B to 30B parameters.
-
Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
Medusa augments LLMs with multiple decoding heads and tree-based attention to predict and verify several tokens in parallel, yielding 2.2-3.6x inference speedup via two fine-tuning regimes.
-
Self-Rewarding Language Models
Iterative self-rewarding via LLM-as-Judge in DPO training on Llama 2 70B improves instruction following and self-evaluation, outperforming GPT-4 on AlpacaEval 2.0.
-
Measuring Faithfulness in Chain-of-Thought Reasoning
Chain-of-Thought reasoning in LLMs is often unfaithful, with models relying on it variably by task and less so as models scale larger.
-
Let's Verify Step by Step
Process supervision significantly outperforms outcome supervision for training models on the MATH dataset, achieving 78% accuracy on a representative test subset with active learning and a released 800k step-label dataset.
-
Red Teaming Language Models with Language Models
One language model can generate diverse test cases to automatically uncover tens of thousands of harmful behaviors, including offensive replies and privacy leaks, in a large target language model.
-
Driving Intents Amplify Planning-Oriented Reinforcement Learning
DIAL uses intent-conditioned CFG and multi-intent GRPO to expand and preserve diverse modes in continuous-action preference RL, lifting RFS to 9.14 and surpassing both prior best (8.5) and human demonstration (8.13).
-
Trust the Batch, On- or Off-Policy: Adaptive Policy Optimization for RL Post-Training
A new RL objective adapts trust-region and off-policy handling automatically via normalized effective sample size of batch policy ratios, matching tuned baselines without new hyperparameters.
-
Discrete Flow Matching for Offline-to-Online Reinforcement Learning
DRIFT enables stable offline-to-online fine-tuning of CTMC policies in discrete RL via advantage-weighted discrete flow matching, path-space regularization, and candidate-set approximation.
-
PriorZero: Bridging Language Priors and World Models for Decision Making
PriorZero uses root-only LLM prior injection in MCTS and alternating world-model training with LLM fine-tuning to raise exploration efficiency and final performance on Jericho text games and BabyAI gridworlds.
-
TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching
TBPO derives a token-level preference optimization objective from sequence-level pairwise data via Bregman divergence ratio matching that generalizes DPO and improves alignment quality.
-
Spurious Correlation Learning in Preference Optimization: Mechanisms, Consequences, and Mitigation via Tie Training
Standard preference learning induces spurious feature reliance via mean bias and correlation leakage, creating irreducible distribution shift vulnerabilities that tie training mitigates without degrading causal learning.
-
Power Reinforcement Post-Training of Text-to-Image Models with Super-Linear Advantage Shaping
Super-Linear Advantage Shaping (SLAS) introduces a non-linear geometric policy update for RL post-training of text-to-image models that reshapes the local policy space via advantage-dependent Fisher-Rao weighting to reduce reward hacking and improve performance over GRPO baselines.
-
Guaranteed Jailbreaking Defense via Disrupt-and-Rectify Smoothing
DR-Smoothing introduces a disrupt-then-rectify prompt processing scheme into smoothing defenses, delivering tight theoretical bounds on success probability against both token- and prompt-level jailbreaks.
-
Annotations Mitigate Post-Training Mode Collapse
Annotation-anchored training reduces semantic diversity collapse in post-trained language models by a factor of six compared to standard supervised fine-tuning while preserving instruction-following and improving with scale.
-
BoostAPR: Boosting Automated Program Repair via Execution-Grounded Reinforcement Learning with Dual Reward Models
BoostAPR boosts automated program repair by training a sequence-level assessor and line-level credit allocator from execution outcomes, then applying them in PPO to reach 40.7% on SWE-bench Verified.
-
$\xi$-DPO: Direct Preference Optimization via Ratio Reward Margin
ξ-DPO rewrites the preference objective as minimizing distance to optimal margins and defines reward as a chosen-to-rejected ratio, yielding a bounded, interpretable margin ξ set directly from the initial reward-gap distribution.
-
Beyond Pairs: Your Language Model is Secretly Optimizing a Preference Graph
GraphDPO generalizes pairwise DPO to a graph-structured Plackett-Luce objective over DAGs induced by rollout rankings, enforcing transitivity with linear complexity and recovering DPO as a special case.
-
SOD: Step-wise On-policy Distillation for Small Language Model Agents
SOD reweights on-policy distillation strength step-by-step using divergence to stabilize tool use in small language model agents, yielding up to 20.86% gains and 26.13% on AIME 2025 for a 0.6B model.
-
Implicit Preference Alignment for Human Image Animation
IPA aligns animation models for superior hand quality via implicit reward maximization on self-generated samples plus hand-focused local optimization, avoiding expensive paired data.
-
Experience Sharing in Mutual Reinforcement Learning for Heterogeneous Language Models
Mutual Reinforcement Learning allows heterogeneous LLMs to exchange experience through mechanisms like Peer Rollout Pooling, Cross-Policy GRPO Advantage Sharing, and Success-Gated Transfer, with outcome-level sharing identified as favorable on the stability-support trade-off.
-
Response Time Enhances Alignment with Heterogeneous Preferences
Response times modeled as drift-diffusion processes enable consistent estimation of population-average preferences from heterogeneous anonymous binary choices.
-
On the Blessing of Pre-training in Weak-to-Strong Generalization
Pre-training provides a geometric warm start in a single-index model that enables weak-to-strong generalization up to a supervisor-limited bound, with empirical phase-transition evidence in LLMs.
-
Stayin' Aligned Over Time: Towards Longitudinal Human-LLM Alignment via Contextual Reflection and Privacy-Preserving Behavioral Data
A methodological framework and browser system BITE for collecting evolving user preferences on LLM outputs through context-triggered reflections and privacy-preserving data over time.
-
Correct Is Not Enough: Training Reasoning Planners with Executor-Grounded Rewards
TraceLift trains reasoning planners using rewards that credit traces for both rubric quality and actual performance gains on a frozen executor, outperforming final-answer-only training on math and code tasks.
-
Replacing Parameters with Preferences: Federated Alignment of Heterogeneous Vision-Language Models
MoR lets clients train local reward models on private preferences and uses a learned Mixture-of-Rewards with GRPO on the server to align a shared base VLM without exchanging parameters, architectures, or raw data.
-
DGPO: Distribution Guided Policy Optimization for Fine Grained Credit Assignment
DGPO is a critic-free RL framework that uses bounded Hellinger distance and entropy-gated advantage redistribution to enable fine-grained token-level credit assignment in long CoT generations for LLM alignment, reporting SOTA results on AIME benchmarks.
-
Gradient-Gated DPO: Stabilizing Preference Optimization in Language Models
Gate-DPO attenuates gradients on low-probability rejected responses to reduce probability collapse and improve chosen-response likelihood during preference optimization.
-
Binary Rewards and Reinforcement Learning: Fundamental Challenges
Binary rewards make the set of reward-maximizing policies infinite in policy gradients; KL control selects the filtered base model but misspecification drives collapse to concentrated valid outputs instead.
-
A Meta Reinforcement Learning Approach to Goals-Based Wealth Management
MetaRL pre-trained on GBWM problems delivers near-optimal dynamic strategies in 0.01s achieving 97.8% of DP optimal utility and handles larger problems where DP fails.
-
Minimizing Collateral Damage in Activation Steering
Activation steering is cast as constrained optimization that minimizes collateral damage by weighting perturbations according to the empirical second-moment matrix of activations instead of assuming isotropy.
-
Diversity in Large Language Models under Supervised Fine-Tuning
TOFU loss mitigates the narrowing of generative diversity in LLMs after supervised fine-tuning by addressing neglect of low-frequency patterns and forgetting of prior knowledge.
-
Wasserstein Distributionally Robust Regret Optimization for Reinforcement Learning from Human Feedback
DRRO for RLHF replaces worst-case value with worst-case regret in Wasserstein DRO, producing an exact water-filling solution under l1 ambiguity and a practical sampled-bonus algorithm that reduces proxy over-optimization.