We provide an online RLHF workflow for large language models.
problem Lack of online RLHF workflow for large language models.
method Construct preference models using open-source datasets, approximate human feedback, and implement online iterative RLHF.
result Our LLM achieves state-of-the-art performance on various benchmarks.
Improves RLHF sample efficiency by scaling reward complexity polynomially.
problem Exponential sample complexity in RLHF algorithms for skewed preferences.
method SE-POPO, an online RLHF algorithm that achieves polynomial sample complexity.
result SE-POPO outperforms existing algorithms in sample efficiency.
New method for RLHF reduces costs by integrating new data in one pass.
problem Continuous integration and re-optimization of models in RLHF leads to high computational and storage costs.
method Proposes a one-pass reward modeling method using online mirror descent with a tailored local norm.
result Achieves constant-time updates per iteration, enhancing both statistical and computational efficiency.
Unified approach to RLHF tackles uncertainty in reward function.
problem Uncertainty in reward function learned from human feedback.
method Value-incentivized preference optimization (VPO) that regularizes the reward function with value function.
result Theoretical and practical guarantees for both online and offline RLHF settings.
This paper explores how imperfect reward models can improve online RLHF.
problem Sample efficiency in online RLHF from imperfect reward models.
method Identifies policy coverability and proposes TPO algorithm for transfer learning.
result TPO algorithm improves sample efficiency compared to standard online learning.
SAIL improves online alignment of large language models with minimal feedback.
problem Offline RLHF methods often lead to sub-optimal performance due to fixed preference datasets.
method SAIL uses bilevel optimization and a single-level first-order method to iteratively refine model alignment.
result SAIL significantly improves alignment performance on open-sourced datasets with minimal computational overhead.
New method for efficient online exploration in RLHF reduces regret.
problem Efficiently collecting new preference data in RLHF to refine reward model and policy.
method Proposes a new exploration scheme that directs preference queries toward reducing uncertainty in reward differences most relevant to policy improvement.
result Establishes regret bounds of order T(β+1)/(β+2) for online RLHF, with polynomial scaling in all model parameters. Sharp analysis improves RLHF sample complexity with KL-regularization.
problem Improving RLHF sample complexity with KL-regularization.
method Sharp analysis of KL-regularized contextual bandits and RLHF.
result Achieved an O(1/ε) sample complexity when ε is sufficiently small.
Paper addresses RLHF alignment challenges with novel algorithms.
problem Challenges in RLHF alignment, especially in strategic exploration.
method Develops a reverse-KL regularized contextual bandit formulation and proposes efficient algorithms with theoretical guarantees.
result Proposed methods significantly outperform existing RLHF algorithms in real-world experiments.
XPO enhances RLHF by encouraging diverse responses, offering improved sample efficiency.
problem Limited exploration in RLHF leads to suboptimal models and bottlenecks.
method XPO is a simple one-line change to DPO, introducing a novel exploration bonus.
result XPO achieves strong theoretical guarantees and promising empirical performance.
This work develops a unified framework for RLHF with general f-divergence regularization.
problem Theoretical understanding of general f-divergence regularization in RLHF. method Holistic approach across f-divergence class, two algorithms based on distinct sampling principles. result Provably efficient algorithms with O(logT) regret and O(1/T) sub-optimality gap. New method proves fast regret bounds for online RLHF with generalized preferences.
problem Minimizing max-regret in online RLHF with general preferences and bandit feedback.
method Adopted Generalized Bilinear Preference Model (GBPM) to investigate polylogarithmic regret guarantees.
result Proved polylogarithmic regret bounds for Greedy Sampling and Explore-Then-Commit policies under GBPM.
Paper proposes a new RLHF framework for human preference learning.
problem Handling dependent online human preference outcomes with dynamic contexts.
method Two-stage algorithm with ε-greedy followed by exploitation; anti-concentration inequalities and matrix martingale concentration techniques. result Our method achieves optimal regret bound and asymptotic normality of estimators.
New RLHF framework handles general preference oracles without reward functions.
problem Handling general preference oracles without assuming a reward function.
method Developed a minimax game between two LLMs for RLHF under a general preference oracle, focusing on KL-regularized preference.
result Proposed algorithms for efficient offline and online RLHF learning.
DRO-REBEL improves LLM alignment by robustly updating models online.
problem Overfitting and drifting of LLMs during RLHF.
method DRO-REBEL uses type-p Wasserstein, KL, and χ2 ambiguity sets for robust online updates. result DRO-REBEL achieves faster convergence and better performance than prior methods.
PILAF optimizes reward models from human feedback for better policy alignment.
problem Creating accurate reward models from human feedback for policy optimization.
method Policy-Interpolated Learning for Aligned Feedback (PILAF) that explicitly aligns preference learning with maximizing underlying oracle reward.
result PILAF is optimal from both optimization and statistical perspectives, demonstrating strong performance in RLHF settings.
SharedRep-RLHF learns shared traits for diverse groups, improving fairness and performance.
problem Uniform-reward RLHF fails to capture diverse preferences, leading to unfairness.
method SharedRep-RLHF learns shared traits among various groups, improving fairness and performance.
result SharedRep-RLHF outperforms MaxMin-RLHF by up to 20% in win rate.
RLHF uses human feedback to train AI models, posing statistical challenges.
problem Aligning AI models with human preferences using noisy, subjective feedback.
method Supervised fine-tuning, reward modeling, policy optimization, statistical ideas.
result Statistical methods for reward function learning and policy optimization.
Theory for RLHF generalization under reward shift and clipped KL.
problem Theoretical understanding of RLHF generalization, especially with reward shift and clipped KL.
method Developed generalization theory for RLHF, accounting for reward shift and clipped KL.
result Presented generalization bounds for RLHF, suggesting generalization error from sampling, reward shift, and KL clipping.
Study tackles RLHF with diverse human feedback, showing limitations and proposing a meta-learning approach.
problem Traditional RLHF fails to balance diverse human preferences.
method Integrates meta-learning and multiple social welfare functions to optimize diverse preferences.
result Establishes sample complexity bounds for optimizing diverse social welfare functions.
SRPO improves AI alignment with human preferences through self-improvement and task-independent optimization.
problem AI models trained with RLHF lack self-correction mechanisms and struggle with task generalization.
method SRPO formulates the preference learning problem as a min-max objective, optimizing a self-improvement policy and a generative policy in an adversarial fashion, making the solution task-independent.
result SRPO outperforms existing methods, achieving 90% AI Win-Rate on XSum and 56% on Arena-Hard prompts after a single revision.
New insights link RLHF and contrastive learning for better model alignment.
problem Aligning large language models with human values.
method Interpreting RLHF and DPO as contrastive learning methods based on mutual information.
result Proposed Mutual Information Optimization (MIO) improves model performance.
New RLHF approach mitigates bias in aligning LLMs with human preferences.
problem Algorithmic bias in RLHF leading to preference collapse.
method Preference Matching (PM) RLHF, using PM regularizer and conditional variant.
result 29% to 41% improvement in alignment with human preferences.
LoCo-RLHF models diverse human feedback with contextual information.
problem Heterogeneous human feedback from diverse contexts and preferences.
method Low-rank contextual preference model, PRS policy.
result LoCo-RLHF achieves tighter sub-optimality gap than existing methods.
RLHF performs well despite violating social choice theory axioms.
problem RLHF's empirical success contradicts social choice theory axioms.
method Showed RLHF satisfies pairwise majority and Condorcet consistency under mild assumptions, and introduced new alignment criteria.
result RLHF satisfies pairwise majority and Condorcet consistency under mild assumptions, explaining its practical success.
This paper solves the multiple reference model problem in RLHF with exact solutions and sample complexity guarantees.
problem Limitations of single reference models in aligning LLMs with human feedback.
method Integrates multiple reference models into RLHF frameworks, addressing theoretical challenges with exact solutions and sample complexity guarantees.
result First exact solution to the multiple reference model problem in reverse KL-regularized RLHF.
New framework formalizes RLHF trilemma: improving safety, fairness, and robustness is computationally infeasible.
problem Aligning large language models with diverse human values while maintaining computational feasibility and robustness.
method Complexity-theoretic analysis integrating statistical learning theory and robust optimization.
result Achieving both representativeness (epsilon <= 0.01) and robustness (delta <= 0.001) for global-scale populations requires super-polynomial operations.
Unified framework for aligning LLMs from human feedback.
problem Lack of strong theoretical justification for RLHF and inability to compare methods.
method Reframed alignment as distribution learning from pairwise preferences, proposing three principled objectives.
result Proposed objectives achieve strong non-asymptotic convergence to target LM.
This work improves RLHF performance guarantees using greedy sampling.
problem Learning the KL-regularized target with preference feedback.
method General preference model, greedy sampling.
result Order-wise improvements over existing RLHF performance guarantees.
This paper investigates teacher hacking during language model distillation and proposes methods to mitigate it.
problem Teacher hacking during language model distillation, leading to suboptimal performance.
method A controlled experimental setup involving an oracle LM, teacher LM, and student LM, using fixed offline or online data generation techniques.
result Data diversity is the key factor in preventing teacher hacking during distillation.
Self-improvement refines language models by verifying their own outputs.
problem Improving language models without external feedback.
method Formalizing self-improvement as sharpening, using the model itself as a verifier.
result RLHF-based self-improvement can outperform SFT-based methods.
New method prevents RLHF alignment collapse by accounting for policy's influence on reward model updates.
problem Iterative RLHF leads to alignment collapse where policies exploit RM's blind spots.
method Foresighted policy optimization (FPO) restores missing steering term via regularization.
result FPO prevents alignment collapse on LLM alignment pipelines using Llama-3.2-1B.
IDS improves RLHF by smoothing reward data, enhancing model performance.
problem Reward model performance degrades and overoptimization hinders true objective.
method Iterative Data Smoothing (IDS) updates model and data labels during each epoch.
result IDS outperforms traditional methods in RLHF.
New RLHF algorithm identifies optimal policies from human feedback without explicit reward inference.
problem Training large language models with human feedback without reward inference.
method Model-free RLHF algorithm BSAD that identifies optimal policies directly from human preference. result Provable, instance-dependent sample complexity ildeO(cMSA3H3Mlogδ1). A new method for RLHF using proximal point Nash learning.
problem Capturing real human preferences in RLHF.
method Proximal point Nash learning, embedding self-play updates into a proximal point framework.
result High-probability last-iterate convergence for the combined method.
New method accounts for hidden context in preference learning for RLHF models.
problem Incomplete data with hidden context affects RLHF model outcomes.
method Distributional Preference Learning (DPL) methods estimate hidden context distributions.
result DPL methods reduce RLHF vulnerabilities by accounting for hidden context.
Develops a new model for RLHF accounting for partially observed states and intermediate feedback.
problem Lack of models for partially observed states and intermediate feedback in RLHF.
method PORRL model with cardinal and dueling feedback methods.
result Demonstrates improved learning and alignment with new model-based and model-free methods.
This paper formalizes reward training in RLHF and provides theoretical guarantees.
problem Costly human feedback collection in RLHF.
method Linear contextual dueling bandit method for regret minimization.
result Derives bounds on simple regret for offline reward training.
Blackwell's theorems influence modern AI through information compression and decision making.
problem Information compression and decision making under uncertainty.
method Theorems developed in the 1940s and 1950s, applied to modern AI.
result Blackwell theorems remain relevant and influence modern AI subfields.
New method fine-tunes discrete diffusion models for RLHF tasks.
problem Fine-tuning discrete diffusion models with policy gradient methods is challenging.
method Proposed SEPO algorithm for efficient fine-tuning over non-differentiable rewards.
result Numerical experiments show scalability and efficiency of SEPO.
Paper shows equivalence between two alignment methods and introduces a new algorithm.
problem Ensuring human alignment of large language models for useful, safe, and pleasant user experience.
method Introduces IPO-MD algorithm, showing equivalence between IPO and Nash-MD methods.
result Equivalence between IPO and Nash-MD methods proven when considering online version of IPO.
The paper analyzes RLHF with human feedback and provides convergence results for MLE and pessimistic MLE.
problem Improving RLHF with human feedback from pairwise or K-wise comparisons. method Theoretical framework for RLHF with convergence analysis of MLE and pessimistic MLE.
result MLE fails but pessimistic MLE provides improved policies under certain coverage assumptions.
A new algorithm RTO integrates DPO and PPO for RLHF, improving performance.
problem Learning from sparse, sentence-level rewards in RLHF.
method Modeling RLHF as MDP, introducing RTO that learns token-wise reward function.
result RTO outperforms PPO on benchmarks by 7.5 points on AlpacaEval 2 and 4.1 points on Arena-Hard.
Paper develops methods to optimize policies directly from human feedback without reward inference.
problem Challenges in RLHF, including reward model overfitting and distribution shift.
method Develops two algorithms for RLHF without reward inference, using zeroth-order gradient approximators.
result Establishes polynomial convergence rates and outperforms existing methods in numerical experiments.
RLHF fails when humans only partially observe, leading to inflated or overjustified feedback.
problem Failure of reinforcement learning from human feedback in partially observable environments.
method Formal definition of failure cases, modeling human as Boltzmann rational, analyzing information provided by feedback.
result RLHF can deceptively inflate or overjustify feedback when humans have partial observations.
Proposes DRRO to mitigate over-optimization in RLHF from human feedback.
problem Over-optimization due to reward misspecification in RLHF.
method Wasserstein distributionally robust regret optimization (DRRO).
result DRRO mitigates over-optimization more effectively than existing baselines.
Paper proposes efficient RLHF methods for LLMs using active queries.
problem Efficiently aligning LLMs with human preferences using RLHF.
method Formalizes RLHF as a dueling bandit problem, introduces APPO and ADPO algorithms.
result ADPO achieves similar performance to state-of-the-art methods with fewer queries.
RL-LOW algorithm achieves exponential simple regret in offline RLHF with pairwise comparisons.
problem Offline reinforcement learning from human feedback with pairwise comparisons.
method Proposes RL-LOW algorithm to minimize simple regret with exponential convergence.
result Achieves an exponential form of simple regret of \(\exp ( - Ω(n/H) )\).