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.
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.
No real-world reward function is perfect. Sensory errors and software bugs may result in RL agents observing higher (or lower) rewards than they should. For example, a reinforcement learning agent may prefer states where a sensory error gives it the maximum reward, but where the true reward is actually small. We formal…
In reinforcement learning, a decision needs to be made at some point as to whether it is worthwhile to carry on with the learning process or to terminate it. In many such situations, stochastic elements are often present which govern the occurrence of rewards, with the sequential occurrences of positive rewards randoml…
Curiosity-Critic improves world model training by focusing on cumulative prediction error.
problem Training world models with intrinsic rewards that consider cumulative prediction error.
method Curiosity-Critic uses a surrogate reward based on the difference between current and asymptotic prediction errors, estimated online by a co-trained critic.
result Curiosity-Critic outperforms other methods in training speed and final world model accuracy.
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.
New algorithm optimally identifies best arm in both stochastic and adversarial settings.
problem Best arm identification in stochastic and adversarial reward scenarios.
method Parameter-free algorithm designed to be robust to adversarial rewards and optimal in stochastic problems.
result Algorithm's error rate matches optimal bounds in stochastic problems and is robust to adversarial rewards.
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. New algorithm for multi-player bandits with collision-dependent rewards.
problem Stochastic multi-player multi-armed bandits with collision-dependent reward distributions.
method Error-Correction Collision Communication (EC3) algorithm.
result EC3 algorithm achieves optimal regret approaching centralized MP-MAB regret.
A major challenge in reinforcement learning is exploration, when local dithering methods such as epsilon-greedy sampling are insufficient to solve a given task. Many recent methods have proposed to intrinsically motivate an agent to seek novel states, driving the agent to discover improved reward. However, while state-…
This study proposes hidden state curiosity to enhance RL models' resilience against noise.
problem Curiosity traps in RL models distract agents from discovering novel experiences.
method Proposed hidden state curiosity based on the Free Energy Principle to reward agents for KL divergence between predictive priors and posteriors.
result Agents with hidden state curiosity are more resilient against curiosity traps compared to those with prediction error curiosity.
New approach for reward-free exploration reduces estimation error.
problem Reward-free exploration in reinforcement learning.
method Adaptive approach reducing MDP estimation error.
result Reward-free UCRL algorithm improves sample complexity.
Reinforcement learning algorithms rely on carefully engineering environment rewards that are extrinsic to the agent. However, annotating each environment with hand-designed, dense rewards is not scalable, motivating the need for developing reward functions that are intrinsic to the agent. Curiosity is a type of intrins…
Recent studies have shown that reinforcement learning (RL) models are vulnerable in various noisy scenarios. For instance, the observed reward channel is often subject to noise in practice (e.g., when rewards are collected through sensors), and is therefore not credible. In addition, for applications such as robotics, …
Robust Q-learning algorithm resists corrupted rewards.
problem Non-ideal environments with corrupted rewards.
method Developed a novel robust Q-learning algorithm using historical data.
result Finite-time convergence rate matches state-of-the-art bounds.
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 …
Dealing with sparse rewards is a longstanding challenge in reinforcement learning. The recent use of hindsight methods have achieved success on a variety of sparse-reward tasks, but they fail on complex tasks such as stacking multiple blocks with a robot arm in simulation. Curiosity-driven exploration using the predict…
The Artificial Prediction Market is a recent machine learning technique for multi-class classification, inspired from the financial markets. It involves a number of trained market participants that bet on the possible outcomes and are rewarded if they predict correctly. This paper generalizes the scope of the Artificia…
Designers of AI agents often iterate on the reward function in a trial-and-error process until they get the desired behavior, but this only guarantees good behavior in the training environment. We propose structuring this process as a series of queries asking the user to compare between different reward functions. Thus…
Abstract MDPs enable strategic exploration and fast reward transfer in complex environments.
problem Challenging to learn accurate MDPs for high-dimensional states.
method Learn an abstract MDP over low-dimensional coarse states, using an abstraction function.
result Achieves superhuman performance on Pitfall! and higher reward with fewer samples.
A new algorithm for offline RL with trajectory-wise reward reduces bias and variance errors.
problem Offline RL with trajectory-wise reward incurs large bias and variance errors.
method PARTED algorithm that decomposes trajectory return into proxy rewards and performs pessimistic value iteration.
result PARTED achieves provably efficient suboptimality bounds in general MDPs with trajectory-wise reward.
New approach transfers rewards learned in one environment to reinforcement learning in a new environment.
problem Transfer of rewards learned using inverse reinforcement learning from one environment to a new, different environment.
method Formulate the problem as a joint system of Bellman equations, develop minimax estimators for the target soft-q-function, solve the source and target system of equations jointly. result The coupled approach removes the first-order influence of source Bellman residual error compared to the sequential approach.
PQR estimates reward functions from actions and states without assuming state-only rewards.
problem Estimating reward functions from actions and states without state-only assumptions.
method Deep learning approach that sequentially estimates policy, Q-function, and reward.
result PQR uniquely recovers true reward with known transitions and bounds error with unknown transitions.
New TD method stabilizes average-reward learning.
problem Stability issues in average-reward TD learning.
method Implicit fixed point update for average-reward TD(λ). result Improved numerical stability and broader step-size range.
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.
Backtesting framework for CLMMs on Uniswap V3 reduces reward estimation error.
problem Estimating rewards for CLMMs in Uniswap V3 liquidity pools.
method Parametric model for liquidity distribution, historical data analysis.
result Error in reward estimation less than 1% for each pool.
Study shows fast rates for inverse reinforcement learning with linear rewards.
problem Entropy-regularized min-max inverse reinforcement learning in finite-horizon MDPs.
method Structural and statistical analysis of Min-Max-IRL with pseudo-self-concordance.
result Both trajectory-level KL divergence and parameter error decay at O(n−1). A novel framework refines diffusion models iteratively for better downstream reward optimization.
problem Optimizing reward functions during inference of diffusion models.
method Iterative refinement process with noising and reward-guided denoising steps.
result Superior empirical performance in protein and DNA design.
SARA uses similarity to learn rewards robustly and adaptively.
problem Robustness to labeler errors and adaptability to diverse feedback formats.
method Contrastive framework that learns latent representations and computes rewards as similarities.
result Strong performance on offline RL benchmarks and diverse applications.
We consider a budget-constrained bandit problem where each arm pull incurs a random cost, and yields a random reward in return. The objective is to maximize the total expected reward under a budget constraint on the total cost. The model is general in the sense that it allows correlated and potentially heavy-tailed cos…
A new algorithm for differential privacy in kernelized contextual bandits reduces error rate.
problem Joint differential privacy in kernelized contextual bandits.
method Proposes a novel algorithm with a specific error rate and privacy parameter dependence.
result Achieves an error rate of $\mathcal{O}\left(\sqrt{\frac{γ_T}{T}} + \frac{γ_T}{T \varepsilon}
ight)$ after T queries. In classical Q-learning, the objective is to maximize the sum of discounted rewards through iteratively using the Bellman equation as an update, in an attempt to estimate the action value function of the optimal policy. Conventionally, the loss function is defined as the temporal difference between the action value and…
This paper presents a new neural architecture that combines a modulated Hebbian network (MOHN) with DQN, which we call modulated Hebbian plus Q network architecture (MOHQA). The hypothesis is that such a combination allows MOHQA to solve difficult partially observable Markov decision process (POMDP) problems which impa…
Off-policy evaluation for MNAR rewards in MDPs
problem Off-policy evaluation in MDPs with MNAR rewards
method Formalizing a reward-dependent propensity model and using future states as shadow variables
result Proposed an Fitted-Q-Evaluation-style estimator that propagates recovered rewards while allowing target policies to depend on past missingness indicators
Paper shows RLHF can be solved similarly to standard RL.
problem Difficulty of RLHF compared to standard RL.
method Reduction to reward-based RL techniques.
result RLHF can be solved using existing algorithms for reward-based RL.
Improves generative models by optimizing rewards and sample editing.
problem Efficiently generating high-reward samples with structural constraints.
method Introduces MDM-VGB, a discrete diffusion sampler that augments unmasking generation with reward-guided remasking.
result MDM-VGB achieves quadratic complexity and robustness to noise, outperforming heuristics like best-of-N. This paper balances short-term and long-term rewards in policy learning.
problem Balancing short-term and long-term rewards in policy learning.
method Formalizes a new framework to balance rewards, identifies rewards under mild assumptions, deduces efficiency bounds, and develops a policy learning approach.
result The proposed method improves the estimator of long-term reward and reduces regret.
A new method recovers rewards from behavior policies using classification and regression.
problem Recovering meaningful rewards from observed behavior in reinforcement learning.
method GenPQR, a modular procedure that estimates behavior policy, evaluates soft Q-function, and recovers normalized reward using classification and regression.
result GenPQR matches or improves reward recovery compared to DeepPQR, while being simpler and more modular.
Study designs logging policies to minimize off-policy evaluation error.
problem Minimizing OPE error with logging policies for target policies.
method Characterizes reward-coverage tradeoff, proposes a unifying framework, derives optimal policies.
result Provides actionable guidance for firms choosing recommendation systems.
New algorithm learns optimal policies without explicit rewards.
problem Learning optimal policies without explicit rewards.
method Developed a computationally tractable algorithm for reward-free navigation using linear value iteration.
result PAC guarantees on learning near optimal value functions and policies.
This work optimizes identifying good arms in nonparametric multi-armed bandits.
problem Efficiently identifying arms with high means in nonparametric settings.
method Combining reward-maximizing sampling with a nonparametric sequential test for anytime-valid labeling.
result Achieves minimax optimal stopping times for identifying arms above a threshold.
New method optimizes policies without assuming known link functions between preferences and rewards.
problem Policy alignment with unknown and unrestricted link functions.
method Formulates an f-divergence-constrained reward maximization problem, learning policies directly. result Induces a semiparametric single-index binary choice model for policy alignment.
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.
The paper analyzes the reward improvement of aligned policies in large language models.
problem Optimizing policies in large language models while staying close to a reference policy.
method Information-theoretic analysis and reduction to exponential order statistics.
result Information-theoretic upper bounds on reward improvement are derived.
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.
New algorithm tackles nonstationary linear bandits with latent dynamics.
problem Nonstationary bandit problem with latent states and unknown dynamics.
method Explore-then-commit algorithm with exploration and commitment phases.
result Achieves ildeO(T2/3) regret. STAR framework reduces OPE variance by distilling complex problems into discrete ARPs.
problem High variance and bias in off-policy evaluation methods.
method STAR framework that includes various OPE estimators and leverages state abstraction.
result Predictions from ARPs estimated from off-policy data are asymptotically correct.
Faster algorithms for solving multichain MDPs under average-reward criterion.
problem Navigating towards the best connected component in multichain MDPs.
method Developed algorithms to better solve the navigational subproblem, achieving faster convergence rates.
result Improved rates of convergence and sharper complexity measures for multichain MDPs.