Study scaling laws of reward model overoptimization in reinforcement learning.
problem Reward model overoptimization hinders true performance in reinforcement learning.
method Synthetic setup with fixed gold reward model; optimization using RL or best-of-n sampling; analysis of scaling laws. result Scaling laws of reward model overoptimization differ based on optimization method and scale smoothly with model parameters.
Reward models need more than just accuracy for effective RLHF.
problem The effectiveness of reward models in RLHF is not fully understood.
method An optimization perspective to evaluate reward models.
result Reward models with low reward variance can lead to a flat optimization landscape, hindering performance.
Mitigates biases in reward models using variational inference.
problem Spurious correlations in reward models that align large language models with human preferences.
method Formulates data-generating process, identifies non-spurious latent variables, and uses variational inference to recover them.
result Effective mitigation of spurious correlation issues, yielding more robust reward models.
BRAID fine-tunes diffusion models to optimize reward models in offline scenarios.
problem Combining generative modeling and model-based optimization in offline scenarios.
method Conservative fine-tuning of diffusion models using RL to optimize reward models.
result BRAID outperforms existing methods in offline data, avoiding invalid designs.
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.
A new method resolves non-identifiability in reward modeling using anchor labels.
problem Non-identifiability in reward modeling from pairwise preferences alone.
method Anchor-guided Variance-aware Reward Modeling (AVRM) framework.
result AVRM resolves non-identifiability and improves reward modeling performance.
CausalRM models rewards from user feedback, overcoming noise and bias.
problem Aligning language models with user preferences from noisy, biased feedback.
method Causal-theoretic reward modeling framework addressing noise and bias in observational feedback.
result CausalRM learns accurate reward signals from noisy and biased observational feedback.
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.
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.
This work shows how approximate reward models can significantly improve inference-time scaling.
problem Improving the efficiency of inference for large language models.
method Identifying the Bellman error of approximate reward models and using Sequential Monte Carlo (SMC) for inference.
result Approximate reward models can reduce computational complexity from exponential to polynomial in T. Study on deep neural networks for reward modeling with pairwise comparison data.
problem Reward modeling with deep neural networks in non-parametric settings.
method Established a non-asymptotic regret bound for deep reward estimators, introduced a margin-type condition.
result Improved regret bound for deep reward estimators, highlighting the importance of clear human beliefs.
The study analyzes how neural reward models learn features for policy optimization in a Gaussian single-index model.
problem Reward modeling in policy optimization and its impact on downstream value.
method Two-stage neural reward model: first learns hidden direction, then fits readout layer.
result For any feature-learning temperature above a dimension-free threshold, a constant fraction of neurons recover the hidden direction.
Learning reward functions can lead to poor policy performance despite low error.
problem Low error in learned reward functions does not guarantee low regret in policy performance.
method Mathematical analysis of reward learning and policy optimization.
result A low expected test error of the reward model guarantees low worst-case regret, but error-regret mismatch can occur with certain data distributions.
One obstacle to applying reinforcement learning algorithms to real-world problems is the lack of suitable reward functions. Designing such reward functions is difficult in part because the user only has an implicit understanding of the task objective. This gives rise to the agent alignment problem: how do we create age…
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.
This work frames reward modelling from preferences as a causal problem.
problem Reward modelling from preference data for AI alignment.
method Causal inference approach to identify challenges and assumptions.
result Causally-inspired approaches improve model robustness.
Privacy-preserving reinforcement learning from human feedback using decoupled reward modeling.
problem Training large language models with sensitive user information while preserving privacy.
method Proposes a privacy-preserving framework that imposes differential privacy on reward learning only.
result Privacy contributes an additional additive term to the suboptimality gap, and the upper bound is rate-optimal up to logarithmic factors.
Language models perform worse with implicit reward models than explicit ones.
problem Understanding why implicit reward models generalize worse than explicit ones.
method Investigated the root cause of the generalization gap between IM-RMs and EX-RMs.
result Implicit reward models rely more on superficial token-level cues, leading to worse generalization.
Method tackles uncertainty in reward models for LLMs from heterogeneous human feedback.
problem Uncertainty in reward models for LLMs from heterogeneous human feedback.
method Heterogeneous preference framework and alternating gradient descent algorithm.
result Established theoretical guarantees for estimator convergence and asymptotic distribution.
Learning reward functions from data is a promising path towards achieving scalable Reinforcement Learning (RL) for robotics. However, a major challenge in training agents from learned reward models is that the agent can learn to exploit errors in the reward model to achieve high reward behaviors that do not correspond …
The paper enhances preference learning by incorporating response time data.
problem Lack of temporal information in user decision-making for reward model learning.
method Integrates response time alongside binary choice data using the EZ model and Neyman-orthogonal loss functions.
result Response time-augmented approach reduces error rates from exponential to polynomial scaling, improving sample efficiency.
EBRM improves robustness and generalization of language model rewards.
problem Challenges in capturing complex human preferences and generalizing to unseen data in reward models.
method Energy-Based Reward Model (EBRM) that models reward distribution explicitly, using conflict-aware data filtering, label-noise-aware contrastive training, and hybrid initialization.
result Significant improvements in robustness and generalization, up to 5.97% improvement in safety-critical alignment tasks.
The paper tackles statistical and computational challenges in learning correlated reward models.
problem The Independence of Irrelevant Alternatives (IIA) assumption collapses human preferences into a universal utility function, leading to coarse approximations.
method The paper investigates the statistical and computational challenges of learning a correlated probit model using best-of-three preference data.
result Best-of-three preference data overcomes the limitations of pairwise preference data, allowing for more fine-grained modeling of human preferences.
Adaptive reward models capture individual preferences from human feedback.
problem Learning a reward model that can be specialised to a user.
method Empirical risk minimisation and PAC bound analysis.
result Adaptive reward models benefit from the heterogeneity of user preferences.
New algorithm reduces online learning regret by exploiting historical invariances.
problem Stochastic non-stationary linear bandits with changing reward models.
method ISD-linUCB algorithm that learns invariances in reward model.
result Significant regret improvements in fast-changing environments with historical data.
EXPO framework eliminates need for reward model, achieving better optimization.
problem Optimizing large language model responses without a separate reward model.
method Introduces EXPO framework that avoids reparameterization, directly optimizing preferences.
result Demonstrates better regularization and intuitive interpolation behaviors.
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.
We here adopt Bayesian nonparametric mixture models to extend multi-armed bandits in general, and Thompson sampling in particular, to scenarios where there is reward model uncertainty. In the stochastic multi-armed bandit, the reward for the played arm is generated from an unknown distribution. Reward uncertainty, i.e.…
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.
The paper proposes a new method for learning reward models from ordinal feedback, improving upon binary feedback.
problem Learning reward models from human preferences using binary feedback discards useful samples and loses fine-grained information.
method The paper introduces a framework for learning reward models under ordinal feedback, generalizing the Bradley-Terry model.
result Ordinal feedback reduces the Rademacher complexity compared to binary feedback, leading to better reward learning.
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.
RAFT fine-tunes models using high-quality samples to align them with human preferences.
problem Aligning generative models with human ethics and preferences.
method RAFT selects high-quality samples, discards undesired behavior, and fine-tunes the model on filtered samples.
result RAFT improves model performance in reward learning and automated metrics.
FSPO optimizes synthetic preferences for LLM personalization.
problem Personalizing large language models for diverse users.
method FSPO reframes reward modeling as a meta-learning problem, using few labeled preferences and synthetic data.
result FSPO achieves high winrates in personalized responses, both synthetic and real.
Contextual multi-armed bandit (MAB) algorithms have been shown promising for maximizing cumulative rewards in sequential decision tasks such as news article recommendation systems, web page ad placement algorithms, and mobile health. However, most of the proposed contextual MAB algorithms assume linear relationships be…
The paper addresses statistical inference for online decision-making in a contextual bandit setting.
problem Understanding the performance of reward models in online decision-making with contextual information.
method The paper uses the contextual bandit framework with a linear reward model and the ε-greedy policy to address the exploration-exploitation dilemma. It employs the martingale central limit theorem and inverse propensity score weighting to establish asymptotic normality of parameter estimators. result The online ordinary least squares estimator and the online weighted least squares estimator are asymptotically normal, providing insights into the performance of the reward model.
Paper proposes a robust RLHF algorithm for LLMs, improving response preference over baselines.
problem Reward model misspecifications in RLHF for LLMs.
method Proposes a robust algorithm to reduce reward and policy estimator variance, theoretically and empirically validated.
result Consistently outperforms existing methods on LLM benchmark datasets, favoring 77-81% of responses over baselines.
SWIFT learns intrinsic rewards from LLM hidden states for efficient best-of-N sampling.
problem Efficiency and scalability of reward models for LLMs.
method SWIFT (Simple Weighted Intrinsic Feedback Technique) learns a reward function directly from LLM hidden states.
result SWIFT outperforms existing baselines by 12.7% on MATH dataset while using less than 0.005% of their parameters.
A new approach to fine-tuning LLMs with human feedback.
problem Inability of current reward models to fully represent human preferences.
method Introducing NLHF, a new pipeline for LLM fine-tuning using pairwise human feedback.
result NLHF produces a sequence of policies converging to the regularized Nash equilibrium.
Reward collapse occurs when ranking-based reward models yield uniform rewards for different prompts.
problem Reward collapse in aligning large language models with human preferences.
method Introduced a prompt-aware optimization scheme to derive closed-form expressions for reward distributions.
result Our prompt-aware utility functions significantly alleviate reward collapse during training.
Higher conservative training increases reward-hacking in reasoning models.
problem Reward hacking during online adaptation in reasoning models.
method Conservative offline training with varying levels of conservatism (β) was applied to a Qwen3-14B policy, and online adaptation was measured against a reward ensemble.
result Higher conservatism (β) increases reward-hacking damage, measured by the Goodhart gap and AUGC.
Efficient exploration improves large language model performance with fewer queries.
problem Improving large language model performance with fewer human feedback queries.
method Sequentially generates queries, fits a reward model to feedback, uses double Thompson sampling with epistemic neural network uncertainty.
result Efficient exploration enables high performance with far fewer queries.
Agents cooperate to make decisions in multi-armed bandits over a graph.
problem Optimizing decisions in multi-agent multi-armed bandits with shared information.
method Designs consensus-based distributed estimation and cooperative algorithms for group decision-making.
result Achieves group performance close to centralized fusion center.
Mitigates overoptimization in RLHF by reformulating SFT loss as a preference optimization loss.
problem Overoptimization in RLHF where reward model misguides generative model.
method Proposes a theoretical algorithm that minimizes maximum likelihood estimation and reward penalty, reformulates as simple objective combining preference optimization and supervised learning losses.
result Improved performance of RPO compared to DPO baselines in aligning LLMs.
Paper connects Plackett-Luce and Cox models for preference estimation.
problem Estimating preferences from annotated data.
method Connects Plackett-Luce model to Cox Proportional Hazards model.
result Implications of the connection between the two models.
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.
EGFs use ergodicity to simplify generative flows for easier training and imitation learning.
problem Challenges in training generative flows, especially in continuous settings and for imitation learning.
method EGFs leverage ergodicity to build simple flows with universality guarantees and tractable FM loss. They introduce a KL-weakFM loss for IL training without a separate reward model.
result EGFs simplify generative flow training and enable effective imitation learning.
We seek to align agent behavior with a user's objectives in a reinforcement learning setting with unknown dynamics, an unknown reward function, and unknown unsafe states. The user knows the rewards and unsafe states, but querying the user is expensive. To address this challenge, we propose an algorithm that safely and …
EORM boosts LLM accuracy with a lightweight, energy-based verifier.
problem Efficiently verifying mathematical reasoning in large language models.
method Energy-based framework to rank Chain-of-Thought solutions using simple outcome labels.
result EORM boosts LLM accuracy to 90.7% on GSM8k and 63.7% on MATH with only 55M parameters.