This paper tackles distribution shift in model-based offline RL, proposing a shifts-aware reward method.
problem Distribution shift challenges model-based offline RL by distorting value estimation and policy optimization.
method The paper disentangles the problem into model bias and policy shift, proposing a shifts-aware reward through probabilistic inference.
result The proposed shifts-aware reward method effectively mitigates distribution shift and improves policy optimization.
A new policy for contextual bandits adapts to reward vector shifts.
problem Learning under reward vector shifts with ordered rewards.
method Adaptive-discretization and optimistic elimination policy.
result Established upper bounds on preference-based regret.
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 tracks changes in infinite action space rewards.
problem Non-stationary Lipschitz bandits with infinite actions.
method Adaptive tracking of significant shifts using hierarchical discretization.
result Achieves minimax-optimal dynamic regret bound of O ~ ( i l d e L 1 / 3 T 2 / 3 ) \mathcal{\widetilde{O}}( ilde{L}^{1/3}T^{2/3}) O ( i l d e L 1/3 T 2/3 ) . New method tracks shifts in infinite-armed bandits without prior knowledge.
problem Tracking shifts in non-stationary infinite-armed bandits.
method Blackbox conversion of finite-armed MAB to infinite-armed non-stationary, randomized elimination.
result First parameter-free optimal regret bounds for all reservoir regularity regimes.
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.
New method tracks significant shifts in nonparametric bandits.
problem Tracking significant changes in nonparametric contextual bandits.
method Proposed a notion of 'experienced significant shifts' to adapt to minimax rate without knowledge of change parameters.
result Experienced significant shifts count fewer changes than traditional metrics, leading to an adaptive algorithm.
It is difficult to be able to imitate well in unknown states from a small amount of expert data and sampling data. Supervised learning methods such as Behavioral Cloning do not require sampling data, but usually suffer from distribution shift. The methods based on reinforcement learning, such as inverse reinforcement l…
Paper addresses off-policy evaluation and learning with covariate shift.
problem Evaluating and training a new policy using historical data with a covariate shift.
method Derives efficiency bounds and proposes doubly robust estimators for OPE and OPL under covariate shift.
result Proposes estimators for off-policy evaluation and learning under covariate shift.
Algorithm minimizes regret in predictive models influenced by their own predictions.
problem Finding near-optimal models under performativity with unknown shifts.
method Developed an algorithm that uses performative feedback to achieve low regret, scaling only with distribution shift complexity.
result Achieved regret bounds scaling with distribution shift complexity, not reward function complexity.
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.
Reward hacking exploits misspecified rewards, affecting agent capabilities and true performance.
problem Reward hacking in RL models exploiting reward misspecifications.
method Constructed four RL environments with misspecified rewards; analyzed agent capabilities and behavior.
result More capable agents exploit reward misspecifications, achieving higher proxy reward but lower true reward.
A method to reduce bias in model-based policy evaluation by shifting operators.
problem Bias in value function computation from noisy estimated models.
method Operator shifting method to reduce the residual norm error.
result The shifting factor is always positive and upper bounded by $1+O\left(1/n
ight)$ .
Study optimal offline RL with uncertainty sets and distribution shifts.
problem Optimal offline reinforcement learning with limited data.
method Construct uncertainty sets and distribution shifts, solve robust Markov decision process.
result Least conservative estimator for unknown true distribution.
Study proposes adaptive RL for dynamic portfolio optimization.
problem Traditional portfolio optimization models fail to adapt to regime shifts.
method Regime-aware reinforcement learning framework with hybrid observations and constrained reward functions.
result Transformer PPO achieves highest risk-adjusted returns, while LSTM variants offer a good balance.
Semi-pessimistic RL tackles distributional shift and data scarcity in offline RL.
problem Distributional shift and scarcity of labeled data in offline RL.
method Proposes a semi-pessimistic RL method that simplifies learning by seeking a lower bound of the reward function.
result Demonstrates clear competitiveness and improved policy learning with vast unlabeled data.
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.
Optimizes reward learning design for complex tasks using nonparametric methods.
problem Challenges in specifying reward functions for complex tasks.
method Models rewards and policies as nonparametric functions in RKHSs, derives risk bounds, and optimizes query design.
result Derives non-asymptotic excess risk bounds and finite sample statistical rates for reward learning.
New method tracks significant arm switches to improve bandit algorithms.
problem Adaptive procedures for bandits with unknown changes in reward distribution.
method Proposes a new notion of significant shift to count severe changes.
result Achieves faster rates than previous methods, especially when few changes are severe.
Causal Imitation Learning handles noisy measurements and distribution shifts.
problem Learning from noisy state observations and distributional shifts.
method Causal inference framework and adversarial RKHS learning.
result Improved robustness to distribution shifts compared to standard methods.
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.
New framework shifts bandit algorithms from expected reward to preference metrics, optimizing mixtures of arms.
problem Traditional bandit algorithms focus on expected rewards, ignoring variability and risk.
method Introduces preference metrics (PMs) and designs algorithms to optimize mixtures of arms.
result Optimal policy selects mixtures of arms based on specific weights, not a single best arm.
DAC-SSM learns domain-agnostic states for better imitation learning.
problem Domain shifts hinder imitation learning in partially observable tasks.
method DAC-SSM uses adversarial training to remove domain-dependent information from states.
result DAC-SSM achieves comparable performance to experts in sparse reward tasks.
Multi-armed bandit algorithms have become a reference solution for handling the explore/exploit dilemma in recommender systems, and many other important real-world problems, such as display advertisement. However, such algorithms usually assume a stationary reward distribution, which hardly holds in practice as users' …
The paper examines how updates to probabilistic models influence behavior based on evidence.
problem Understanding how updates to probabilistic models influence behavior based on evidence.
method Study of KL-regularized soft updates as Bayesian posterior updates within a single probabilistic model.
result Posterior updates determine relative incentives but not absolute rewards, which are ambiguous up to context-specific baselines.
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.
Paper tackles best arm identification with cost consideration.
problem Best arm identification with cost consideration in product development.
method Derives a theoretical lower bound and proposes algorithms CTAS and CO.
result Simple algorithms can deliver near-optimal performance.
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 ) T^{(β+1)/(β+2)} T ( β + 1 ) / ( β + 2 ) for online RLHF, with polynomial scaling in all model parameters. 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 B S A D \mathsf{BSAD} BSAD that identifies optimal policies directly from human preference. result Provable, instance-dependent sample complexity i l d e O ( c M S A 3 H 3 M log 1 δ ) ilde{\mathcal{O}}(c_{\mathcal{M}}SA^3H^3M\log\frac{1}δ) i l d e O ( c M S A 3 H 3 M log δ 1 ) . Paper tackles RLHF with DCPPO method, proving near-optimal suboptimality.
problem Challenges in offline RLHF with limited human feedback and bounded rationality.
method DCPPO method involving three stages: MLE, reward function recovery, and pessimistic value iteration.
result DCPPO's suboptimality almost matches classical pessimistic offline RL in terms of distribution shift and dimension.
This paper explores how environmental properties can simplify reinforcement learning in non-episodic settings.
problem Challenges in reinforcement learning with continuous interaction and sparse delayed rewards.
method Analysis of environment shaping and dynamism properties to simplify learning.
result Properties like environment shaping and dynamism can significantly ease learning in non-episodic, sparse reward settings.
New offline RL method handles average-reward MDPs with single-policy coverage.
problem Challenges in offline reinforcement learning due to distribution shift and non-uniform coverage.
method Develops an algorithm based on pessimistic discounted value iteration with quantile clipping.
result First fully single-policy sample complexity bound for average-reward offline RL.
Adaptive traffic control uses deep RL to improve decision-making.
problem Improving traffic control using deep RL.
method Integrates recent deep RL techniques into a novel DQN-based algorithm (TC-DQN+) for traffic control.
result Proposes a new reward function for traffic control.
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.
New algorithm identifies best arm in semiparametric bandits with near optimal efficiency.
problem Fixed-confidence Best Arm Identification in semiparametric bandits with unknown baseline shift.
method Phase-elimination algorithm based on orthogonalized regression design.
result Nearly optimal high-probability sample-complexity upper bound established.
This paper improves sample efficiency for off-policy evaluation with preference data.
problem Improving sample efficiency for off-policy evaluation with preference data.
method Using a deep neural network to learn the value function and leveraging manifold structure.
result Established a provably efficient guarantee for off-policy evaluation with RLHF.
Optimizes mobile notifications for multiple objectives using reinforcement learning.
problem Optimizing mobile notification systems for multiple objectives.
method End-to-end offline reinforcement learning with Double Deep Q-network and Conservative Q-learning.
result Demonstrates improved performance and benefits of the proposed approach.
MA-COPP predicts multi-agent system outcomes using data from a different policy, with probabilistic guarantees.
problem Predicting outcomes in multi-agent systems using data from a different policy.
method Conformal prediction framework applied to multi-agent systems, avoiding exhaustive search.
result Achieves probabilistic guarantees for multi-agent system predictions.
Learning to imitate expert behavior from demonstrations can be challenging, especially in environments with high-dimensional, continuous observations and unknown dynamics. Supervised learning methods based on behavioral cloning (BC) suffer from distribution shift: because the agent greedily imitates demonstrated action…
Framework for optimizing search engine rankings using observational data.
problem Optimizing ranking policies for search engines using limited observational data.
method Formulated expected reward optimization problem, estimated context value distribution, trained ranking policy via Bayesian inference.
result Demonstrated trade-offs in ranking policies trained on empirical reward estimates.
Paper proposes efficient offline neural bandit method.
problem Offline policy learning with neural networks.
method Provable efficient offline contextual bandit with neural network function approximation.
result Method provably generalizes over unseen contexts.
DRO-REBEL improves LLM alignment by robustly updating models online.
problem Overfitting and drifting of LLMs during RLHF.
method DRO-REBEL uses type- p p p Wasserstein, KL, and χ 2 χ^2 χ 2 ambiguity sets for robust online updates. result DRO-REBEL achieves faster convergence and better performance than prior methods.
QuantaAlpha uses evolutionary algorithms to mine financial alpha robustly across market distributions.
problem Challenges in alpha mining due to market noise and regime shifts.
method Evolutionary framework treating each mining run as a trajectory, mutation, crossover, targeted revision, and reuse of effective patterns.
result Consistent gains over strong baselines and prior systems, achieving high IC and ARR.
Robots' agility in changing terrain helps financial models adapt to market shifts.
problem Challenges in financial market forecasting due to regime switching.
method Adapts pretrained LLMs using intrinsic market rewards and reinforcement learning.
result Significantly improved accuracy in adapting to market regime shifts.
New metrics improve scRNA-seq perturbation modeling by reducing mode collapse.
problem Outperformed by simple mean prediction in scRNA-seq perturbation modeling.
method Introduce DEG-aware metrics (WMSE, R w 2 ( Δ ) R^{2}_{w}(Δ) R w 2 ( Δ ) ) and negative/positive baselines. result WMSE loss function reduces mode collapse and improves model performance.
Proposes DAM for optimizing discrete generative models.
problem Challenges in optimizing discrete generative models.
method Discrete Adjoint Matching (DAM) for discrete state spaces.
result Demonstrates effectiveness on synthetic and mathematical reasoning tasks.
Adaptive RL optimizes testing resource allocation for dynamic software environments.
problem Optimizing resource allocation for evolving software testing environments.
method Integrates Q-learning with hybrid reward design for sequential decision-making.
result Consistently outperforms static and optimization-based baselines in simulation studies.
MOPO optimizes offline RL by penalizing dynamics uncertainty.
problem Learning policies from offline data with distributional shift.
method Modify model-based RL to avoid distributional shift.
result MOPO outperforms model-free and standard model-based RL.