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.
In this paper we investigate the Follow the Regularized Leader dynamics in sequential imperfect information games (IIG). We generalize existing results of Poincaré recurrence from normal-form games to zero-sum two-player imperfect information games and other sequential game settings. We then investigate how adapting th…
The study categorizes reward errors in reinforcement learning, finding some can be beneficial.
problem Training language models with imperfect proxy rewards.
method Theoretical analysis of policy gradient optimization and categorization of reward errors.
result Reward errors can be benign or even beneficial, preventing policy from stalling.
Fine-tunes diffusion models to generate diverse samples with high genuine rewards.
problem Reward collapse in finetuning diffusion models.
method Entropy-regularized control against pretrained diffusion models.
result Efficient generation of diverse samples with high genuine rewards.
Improves off-policy evaluation with imperfect annotations.
problem Limited dataset coverage for evaluating new policies.
method Doubly robust estimators combining IS and DM, incorporating counterfactual annotations.
result Using annotations within the DM component yields the most desirable theoretical results.
Improved selection of best outputs from LLMs for better accuracy.
problem BoN fails to reliably find the correct answer with imperfect rewards.
method Majority-of-the-Bests (MoB) selects the mode of a bootstrapped output distribution of BoN.
result MoB consistently improves over BoN in 25 out of 30 setups.
Study cooperative bandit learning with imperfect communication, achieving near-optimal performance.
problem Real-world distributed decision-making with imperfect communication.
method Proposed decentralized algorithms for three communication scenarios: stochastic networks, random delays, and adversarially corrupted rewards.
result Achieved competitive performance and near-optimal guarantees on group regret.
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 n n sampling; analysis of scaling laws. result Scaling laws of reward model overoptimization differ based on optimization method and scale smoothly with model parameters.
RLCFR improves CFR's generalization in imperfect information games.
problem Improving CFR's performance in large-scale, imperfect information games.
method RLCFR integrates CFR with deep reinforcement learning to update strategies dynamically.
result RLCFR significantly enhances CFR's generalization ability in various games.
Current reinforcement learning methods fail if the reward function is imperfect, i.e. if the agent observes reward different from what it actually receives. We study this problem within the formalism of Corrupt Reward Markov Decision Processes (CRMDPs). We show that if the reward corruption in a CRMDP is sufficiently "…
New concept: reward hacking, where optimizing a flawed reward function can hurt performance.
problem Optimizing imperfect reward functions leads to poor performance.
method Formal definition and analysis of reward hacking, examining conditions for unhackability.
result Reward functions are usually hackable, making it hard to align AI with human values.
Paper addresses reward hacking in preference optimization, proposing POWER-DL to improve AI alignment.
problem Reward hacking problem in preference optimization, leading to undesired behaviors.
method POWER-DL combines robust reward maximization and dynamic label updates to mitigate reward hacking.
result POWER-DL consistently outperforms state-of-the-art methods on alignment benchmarks.
Greedy policies perform poorly in imperfectly observed contextual bandits.
problem Performance of Greedy policies in bandits with partially observed contexts.
method Analysis of Greedy reinforcement learning policies under imperfectly observed contextual bandits.
result Worst-case regret grows poly-logarithmically with the time horizon and the failure probability.
This work presents a game-theoretic method for AVs that handles imperfect communication and individual rewards.
problem Real-time, imperfect communication and individual rewards in multi-agent interactions.
method Game-theoretic approach that allows for imperfect communication and individual rewards.
result More realistic assumptions lead to better reward inference and prediction of future actions.
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.
This paper addresses reward estimation and incentive design for agents with hidden rewards.
problem Estimating and incentivizing agents with unknown rewards in a learning setting.
method Repeated adverse selection game with a self-interested learning agent and a learning principal. Introduces an estimator for consistent reward estimation and a data-driven incentive policy.
result Finite-sample consistency of the estimator and a rigorous regret bound for the principal.
The paper develops methods to estimate POMDPs from partial information.
problem Making decisions under partial information about state variables.
method Structural estimation of POMDP primitives using observable history.
result Conditions for model identifiability without state dynamics knowledge.
Investigates a Kyle model with imperfect information and risk aversion.
problem Tackles a Kyle model with imperfect information and risk-averse informed traders.
method Solves an optimal transport problem and a filtering problem under specific measures.
result Constructs an equilibrium for the Gaussian Kyle model with imperfect information and risk aversion.
New algorithm improves inference-time alignment without reward hacking.
problem Improving quality of responses from language models with limited compute.
method Inference-time alignment, focusing on e x t t t I n f e r e n c e T i m e P e s s i m i s m exttt{InferenceTimePessimism} e x ttt I n f er e n ce T im e P ess imi s m algorithm. result Optimal performance and scaling-monotonicity of e x t t t I n f e r e n c e T i m e P e s s i m i s m exttt{InferenceTimePessimism} e x ttt I n f er e n ce T im e P ess imi s m . We consider the optimal stopping problem with non-linear f f f -expectation (induced by a BSDE) without making any regularity assumptions on the reward process ξ ξ ξ . and with general filtration. We show that the value family can be aggregated by an optional process Y Y Y . We characterize the process Y Y Y as the $\mathcal{E}^f…
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.
New method robustly discovers causal relationships from imperfect data.
problem Challenges in causal discovery from imperfect structural constraints.
method Prior alignment and conflict resolution through surrogate model and multi-task learning.
result Proposes a robust method for causal discovery under imperfect constraints.
AdaDEM decouples EM into two parts to improve class overlap and uncertainty.
problem Improper EM limits its effectiveness in various machine learning tasks.
method Decouple EM into CADF and GMC, and AdaDEM normalizes CADF reward and uses MEC.
result AdaDEM outperforms classical EM and improves performance in noisy and dynamic environments.
There has been an increased interest in multimodal language processing including multimodal dialog, question answering, sentiment analysis, and speech recognition. However, naturally occurring multimodal data is often imperfect as a result of imperfect modalities, missing entries or noise corruption. To address these c…
The paper tackles learning from imperfect human feedback, especially in dueling bandit problems.
problem Learning from human feedback that can be irrational or imperfect.
method Developed a Robustified Stochastic Mirror Descent for Imperfect Dueling (RoSMID) algorithm.
result Achieved nearly optimal regret for dueling bandit problems under imperfect human feedback.
QAlign improves language model alignment with less compute, outperforming existing methods.
problem Improving language model performance with limited test-time computation.
method QAlign: sampling from optimal aligned distribution using Markov chain Monte Carlo.
result Consistent improvements over existing methods on various benchmarks.
Algorithm learns NE in imperfect information games with imperfect feedback.
problem Learning Nash equilibrium in imperfect information games with bandit feedback.
method IXOMD algorithm for model-free learning with 1 / T 1/\sqrt{T} 1/ T convergence rate. result IXOMD achieves 1 / T 1/\sqrt{T} 1/ T convergence rate to NE. Model financial markets using open quantum systems to understand market imperfections.
problem Understanding market imperfections through imperfect trading mechanisms.
method Using open quantum systems to represent financial markets, characterizing orbits, and analyzing reduced density matrices.
result Non-classical modes of time evolution can incorporate factors like illiquid trades and imperfect trading mechanisms.
Imitation learning (IL) aims to learn an optimal policy from demonstrations. However, such demonstrations are often imperfect since collecting optimal ones is costly. To effectively learn from imperfect demonstrations, we propose a novel approach that utilizes confidence scores, which describe the quality of demonstrat…
Simulator imperfection, often known as model error, is ubiquitous in practical data assimilation problems. Despite the enormous efforts dedicated to addressing this problem, properly handling simulator imperfection in data assimilation remains to be a challenging task. In this work, we propose an approach to dealing wi…
Selective planning with imperfect models reduces harmful effects of model inadequacy.
problem Harmful effects of using an imperfect model in reinforcement learning.
method Selective planning with heteroscedastic regression to estimate predictive uncertainty from model inadequacy.
result Effective selective planning requires considering both parameter uncertainty and model inadequacy.
Bayesian framework mixes imperfect models for improved predictions.
problem Improving predictions of complex computational models in unknown domains.
method Local Bayesian Dirichlet mixing of imperfect models using the Dirichlet distribution.
result Global and local mixtures of models achieve excellent performance in prediction accuracy and uncertainty quantification.
Modeling the purposeful behavior of imperfect agents from a small number of observations is a challenging task. When restricted to the single-agent decision-theoretic setting, inverse optimal control techniques assume that observed behavior is an approximately optimal solution to an unknown decision problem. These tech…
The study examines how verifier imperfections impact test-time scaling techniques.
problem Understanding how verifier imperfections affect test-time scaling methods.
method Proves the instance-level accuracy of Best-of-N and Rejection Sampling methods using the geometry of the verifier's ROC curve.
result RS outperforms BoN for fixed compute, but both converge to the same accuracy in the infinite-compute limit.
Polynomial chaos surrogates quantify epistemic uncertainty in AI-driven scientific models.
problem Uncertainty in reward estimates hinders interpretability in sequential generative models.
method Fit polynomial chaos expansions to trained models to propagate epistemic uncertainty and quantify sensitivity.
result Interpretable decomposition of reward components driving generative decisions.
Unified framework for policy learning using weak supervision.
problem High-quality supervision is often infeasible or expensive in practice.
method Treat weak supervision as imperfect peer information and evaluate policies based on correlated agreement.
result Substantial performance improvements, especially in complex or noisy environments.
We study pricing and superhedging strategies for game options in an imperfect market with default. We extend the results obtained by Kifer in \cite{Kifer} in the case of a perfect market model to the case of an imperfect market with default, when the imperfections are taken into account via the nonlinearity of the weal…
Robust real-world learning should benefit from both demonstrations and interactions with the environment. Current approaches to learning from demonstration and reward perform supervised learning on expert demonstration data and use reinforcement learning to further improve performance based on the reward received from …
We study the effect of imperfect training data labels on the performance of classification methods. In a general setting, where the probability that an observation in the training dataset is mislabelled may depend on both the feature vector and the true label, we bound the excess risk of an arbitrary classifier trained…
JPS improves joint policies for multi-agent collaboration in imperfect information games.
problem Learning good joint policies for multi-agent collaboration with imperfect information.
method Decomposes global changes to localized policy changes, iteratively improving joint policies without re-evaluating the entire game.
result JPS improves solutions provided by unilateral approaches and outperforms algorithms designed for collaborative policy learning.
DREAM learns optimal strategies in imperfect games without needing a simulator.
problem Learning optimal strategies in imperfect-information games with multiple agents.
method DREAM is a deep reinforcement learning algorithm that converges to Nash Equilibria and coarse correlated equilibria.
result DREAM achieves state-of-the-art performance in benchmark games and is competitive with simulator-based algorithms.
Policy gradient method proves convergence in imperfect-information games.
problem Policy gradient methods in imperfect-information games (EFGs).
method Policy gradient approach with best-iterate convergence.
result Policy gradient leads to provable best-iterate convergence in self-play EFGs.
Paper develops a framework for learning interpretable representations of sequential decision behavior.
problem Obtaining a transparent description of existing behavior.
method Inverse decision modeling framework, formalizing both forward and inverse problems.
result Learning interpretable representations of behavior, including suboptimal actions, biased beliefs, and imperfect knowledge.
New method uses imperfect LLM annotations for valid statistical inference in social science.
problem Inaccurate large language model annotations in social science research.
method Design-based supervised learning (DSL) combining imperfect LLM surrogates with gold-standard labels.
result DSL provides valid statistical inference with comparable predictive accuracy to existing methods.
Study on teaching with imperfect knowledge, showing its impact on optimal teaching sets.
problem Effect of imperfect teacher knowledge on effective teaching.
method Connections to machine teaching problem, optimal teaching sets.
result Teacher's success or failure depends on imperfect knowledge.
Paper introduces a method to learn physics between digital twins using imperfect models.
problem Learning physics from imperfect data and low-fidelity models.
method Bayesian Hierarchical modeling with physics-informed Gaussian processes.
result Models learning between digital twins are less uncertain than independent models but not over-confident.
New algorithm uses imperfect advice to improve online bipartite matching performance.
problem Online bipartite matching with imperfect advice.
method Designing an algorithm that uses external advice to improve performance between advice-free methods and optimal ratio.
result Algorithm achieves competitive ratio interpolating between advice-free methods and optimal ratio of 1.
Bayesian framework calibrates imperfect models using physics-informed priors and Hamiltonian Monte Carlo.
problem Quantifying uncertainty in imperfect computer models described by differential equations.
method Physics-informed Gaussian process priors, discrepancy function, Hamiltonian Monte Carlo, data approximations.
result Framework accurately recovers true parameters and produces accurate predictions.