Generalization in RL

The difficulty of learning goal-reaching policies is often attributed to a "curse of horizon" that manifests as bias accumulation in temporal-difference backups and noisy advantage estimates. In this work, we identify an additional informational curse of horizon in goal-conditioned policy learning, where increasing the goal relabeling horizon can significantly reduce policy generalization and performance. Through a series of controlled experiments with oracle planners, we decouple the goal horizons sampled during training from those that the policy is asked to reach at test time. Even when evaluated only on a sequence of nearby subgoals, goal-conditioned behavioral cloning (BC) policies suffer from severe, training horizon-dependent performance degradation that is mitigated by reinforcement learning (RL) objectives. We explain this phenomenon as a horizon-dependent decrease in the conditional mutual information between actions and hindsight-relabeled goals, and find empirically that both BC and RL policies trained on longer-horizon goals exhibit a shift in sensitivity from goal to state information, as measured by the policy's input Jacobians. Motivated by this observation, we find that distilling the input Jacobians of short-horizon policies into long-horizon policies yields significant performance gains, especially in combinatorial manipulation tasks. Taken together, our results highlight goal relabeling horizon as an important consideration when learning generalist policies from offline data.

Methods for generalization in model-based reinforcement learning typically assume that an agent cannot recover the latent context governing the environment dynamics from its own experience, and therefore supplies it externally. We formalize and test this assumption with \emph{predictive sufficiency}, which quantifies what access to the context adds to next-step prediction under the visitation distribution an agent induces, and separates that quantity into a history-recoverable part, a residual requiring the true context, and the deficit added by a finite model. We classify context-aware algorithms by the predictive risk their conditioning set can target and demonstrate across environments of increasing identification difficulty that the headroom does not follow the MDP class. The same task under different priors leaves predictive headroom in one setting and nothing distinguishable from zero in another, where the agent's behavior implicitly identifies the context and any benefit of such a mechanism cannot be attributed to missing information. Where headroom persists, the learned state exposes it only partially, and adding the true context still lowers the risk. Our contribution is a practical criterion for matching contextual mechanisms to the information available to them, estimated from the ordinary trained agent without a reference policy.

This paper explores a reward-based policy to achieve zero-shot transfer between source and target environments with completely different observation spaces. While humans can demonstrate impressive adaptation capabilities, deep neural network policies often struggle to adapt to a new environment and require a considerable amount of samples for successful transfer. Instead, we propose a novel reward-based policy only conditioned on rewards and actions, enabling zero-shot adaptation to new environments with completely different observations. We discuss the challenges and feasibility of a reward-based policy and then propose a practical algorithm for training. We demonstrate that a reward policy can be trained within three different environments, Pointmass, Cartpole, and 2D Car Racing, and transferred to completely different observations, such as different color palettes or 3D rendering, or Stretch robot navigation in Habitat-Sim, in a zero-shot manner. We also demonstrate that a reward-based policy can further guide the training of an observation-based policy in the target environment.

Unsupervised skill discovery has emerged as a promising approach for leveraging reward-free datasets to pretrain general-purpose policies. However, current skill discovery methods either require access to expert data or exhibit limited generalization, failing to transfer effectively to previously unseen layouts. A key challenge is to learn representations that capture the temporal structure of the environment while remaining robust to variations across layouts. To address this issue, we present an objective for learning action-aware temporal representations that satisfy the functional equivariance property while preserving the local temporal structure of the environment. Building upon this embedding, we further propose unsupervised skill discovery using bisimulation, which learns transferable skills by conditioning the behavior of skills exclusively on the subset of state features that directly affect their execution. This enforces invariant behavior across different layouts, enabling skills to transfer effectively to other configurations. Finally, through comprehensive empirical evaluations, we show that skills learned in a given environment can be effectively applied to solve downstream tasks in various environment layouts, demonstrating strong out-of-distribution generalization.

Large language models (LLMs) are increasingly deployed as agents for multi-step decision-making, yet transfer poorly to unseen environments. World-model methods address this by training agents to predict future observations, at the cost of additional training and errors that compound when predictions are used for planning. However, for LLM agents operating in digital environments, much of this world knowledge is already internalized during pretraining, which shifts the problem from acquiring it to eliciting it. We argue that typical post-training provides little pressure for such elicitation, since supervision under a single goal at each visited state inadvertently drives policies to rely on superficial contextual habits. We introduce EVOKE, a post-training method that supplies this pressure through goal diversity at fixed states. Motivated by theory showing that an agent competent across diverse goals must encode a world model recoverable from its action preferences, EVOKE holds the environment state and interaction history fixed and ranks the same candidate actions under alternative goals, forcing action preferences to change, so that a policy relying on contextual habits or single-goal correlations cannot order them correctly. This implicitly elicits the policy's pretrained world knowledge to inform decisions. We evaluate EVOKE across diverse tasks in three backbones, demonstrating improved task performance, unseen environment generalization, and data efficiency. We further conduct controlled analyses to better understand what drives these gains. These findings offer a new perspective on eliciting internalized world knowledge for transferable action through direct decision supervision.

Temporal abstraction via options can improve exploration in large environments. However, existing option discovery algorithms find subgoals that target all aspects of the state simultaneously. This state-reaching approach produces options that only apply in narrow regions of the state-space, eventually causing an explosion in the number of options that overwhelms the agent, and impedes progress on its primary task of reward maximization. We introduce an algorithm that instead identifies a small, relevant subset of features for each subgoal, yielding options that generalize broadly and accelerate exploration. Our approach learns abstract, transferrable options and achieves rapid exploration in three sparse-reward, image-based domains, including the Atari game MontezumasRevenge.

The job shop scheduling problem is a challenging combinatorial optimization problem, and recent reinforcement learning approaches using graph neural networks have shown promise for learning scheduling policies directly from problem instances. However, training on large instances remains computationally expensive, and generalization across instance sizes remains challenging. This paper studies curriculum learning for graph neural network-based reinforcement learning in the job shop scheduling problem by comparing it with single-size training across three target sizes: 20 x 20, 25 x 25, and 30 x 30. In the curriculum setting, the policy is first trained on smaller instances and then progressively adapted to larger target sizes, allowing scheduling behavior learned in earlier stages to support learning on larger instances. Models are evaluated on unseen instances from 8 x 8 to 30 x 30 using the optimality gap, considering both generalization across all evaluation sizes and specialization on the target size. Results show that curriculum learning consistently reduces wall-clock training time, with larger benefits as the target size increases. The strongest advantage is observed at 30 x 30, where curriculum learning reduces the mean optimality gap across all evaluation sizes by approximately 8.1 percentage points, reduces the target-size mean optimality gap by approximately 8.6 percentage points, and saves approximately 50 hours of training time.

The one-dimensional bin packing problem (1D-BPP) is a classical NP-hard combinatorial optimization problem with applications ranging from logistics and manufacturing to cloud resource management. Although deep reinforcement learning (DRL) has become a competitive paradigm for data-driven optimization, most learned packing methods target 2D and 3D variants, and intelligent learned solvers for 1D-BPP remain scarce. In this paper, we present a novel end-to-end, size-agnostic graph reinforcement learning framework for 1D-BPP. We formulate the packing process as a Markov decision process on an item-compatibility graph, serving as a structural knowledge representation in which every action merges two partial bins that fit together. A graph neural network actor-critic policy extracts relational features from this representation and is trained through reinforcement learning and decoded by stochastic beam search, enabling a single trained model to generalize zero-shot to instances of any size. We conduct a systematic empirical study across graph encoders, DRL algorithms, reward functions, training distributions, and hyperparameters. Evaluated zero-shot on the full BPPLIB benchmark against a constructive heuristic, a grouping genetic algorithm, and recent learned methods, our data-driven policy lowers the mean optimality gap of the constructive heuristic from 2.66% to 2.31%, with the largest gains on structured instances. Against learned baselines evaluated on the same benchmark, it attains a lower gap on most of the nine families and is far more stable across instance distributions. On the hardest benchmark family, it outperforms a state-of-the-art learned solver that relies on column generation and integer programming, while using no solver at all. A grouping genetic algorithm remains ahead overall, and we analyze where and why the residual gap arises.

Solomonoff induction frames prediction as a mixture over computable hypotheses, typically leading to identification of the true environment. In a finite meta-reinforcement learning setting with nested constraint families, in our previous work, we observe a different regime: a value-mixture (VM) agent achieves near-optimal, zero-collision navigation without identifying the true environment, a phenomenon we call Free Inference. This regime persists up to a sharp density threshold, beyond which performance degrades and posterior-mode selection (PMS) becomes preferable. We formalize this behavior via the Free Inference dimension dFI(S,N), a combinatorial measure of the environmental complexity a VM agent can handle while preserving trajectory coherence. We prove dFI is strictly smaller than the VC-dimension and relates to the Natarajan dimension up to a path-length factor, capturing the cost of non-decomposable loss. A PAC-style relaxation yields generalization bounds driven by dFI^(epsilon,delta). We also define a complementary PMS identification dimension and show that a hybrid strategy---averaging until the first collision, then switching to selection---is optimal, with links to Littlestone-type dimensions supported by grid-world experiments.

Designing effective chemotherapy regimens is hindered by tumor heterogeneity and drug resistance, which complicate the deployment of patient-specific model-based optimal control across diverse populations. We develop and compare closed-loop deep reinforcement learning (DRL) dosing policies with continuous (TD3) and discrete (DQN) action spaces trained on a high-dimensional heterogeneous tumor model. The DRL policies are benchmarked against a Pontryagin's Maximum Principle (PMP)-derived open-loop benchmark. We assess generalization under parametric heterogeneity using a 100-patient virtual cohort with plus or minus 10 percent uniform perturbations in growth and drug-sensitivity parameters. Across this cohort, TD3 achieves higher average tumor reduction, while DQN yields tighter inter-patient dosing consistency, revealing a clear efficacy-consistency trade-off in this study. Our simulations assume full observation of all tumor subpopulations; translation to sparse and noisy clinical measurements will require partial-observability formulations and/or state estimation. Overall, the results show that simulation-trained DRL can learn state-dependent feedback dosing policies that complement open-loop optimal control benchmarks.

This work introduces an alternative view of efficient exploration and studies its theoretical and empirical implications in the absence of extrinsic rewards. Specifically, we define efficient explorers as agents that prioritize generating generalizable experience, i.e., data that supports learning models capable of predicting and adapting across the environment. This allows us to analyze efficient exploration through the lens of prediction and generalization. Theoretically, we demonstrate that optimally efficient explorers naturally schedule their trajectories to visit the most informative and learnable regions first. Empirically, we show that optimizing for these agents gives rise to an automatic curriculum of progressively more complex behaviors, even in relatively simple environments. These results indicate that pursuing this purely intrinsic objective alone is enough to drive the emergence of highly sophisticated behaviors. We believe that this new framework provides a principled mechanism by which agent-environment systems may sustain an open-ended process of increasingly complex behavior without external rewards, tasks, or objectives.

Uncertainty estimation provides promising capabilities for reinforcement learning (RL) agents. Notably, estimating uncertainty can reduce the training time and enable agents to obtain greater rewards over time by exploiting information related to whether an action would facilitate exploration of portions of an environment that are well-known versus those that are relatively unknown. In this work, we propose a novel formulation of the experience replay buffer commonly used in RL that we call uncertainty-driven replay memory (UDRM), which entails an update scheme for internally stored memories based on uncertainty estimates obtained by an RL model during training. In contrast to existing forms of RL, which typically use temporal difference error or the distribution of transitions to update the replay memory buffer and train RL controllers, our scheme biases the memory buffer to store more uncertain transitions that will improve an RL agent's generalization throughout training. Experimental results demonstrate that our proposed uncertainty-aware replay buffer enables an RL agent to obtain higher rewards during training compared to other existing uncertainty-aware RL frameworks.

Terminal agents are a compelling application of large language models (LLMs), with the potential to integrate deeply into users' daily workflows. Reinforcement learning (RL) is a key technique for improving their capabilities, making scalable training environments a central challenge. Since public real-user interaction data are scarce, synthetic environments provide a practical alternative, but often suffer from domain gaps and limited fidelity, leading to poor generalization. Existing work mainly scales the quantity and diversity of synthetic environments, while reward-signal quality and the mechanisms governing generalization remain under-explored. We study how RL improves terminal agents and propose the Agentic Compositional Generalization hypothesis: rather than teaching new domain-specific skills from scratch, RL primarily shapes high-level decision-making behaviors that compose and route low-level skills acquired during pre-training and supervised fine-tuning (SFT). This account is consistent with our empirical results and suggests that verifier quality, which determines which behaviors are reinforced, is more important than simply increasing environment quantity or diversity. Motivated by this insight, we propose River, a simple training recipe that improves reward quality by filtering low-quality environments and augmenting outcome rewards with process-level behavior regularization. Using this recipe, our RL-trained agent achieves the best performance among evaluated open-source RL-trained 8B models across four terminal-agent benchmarks. River also generalizes across model families, scales, agent harnesses, and RL objectives. Using fewer than 30% of the TMax training environments, River improves RL gains by 106% and 30% on average for models ranging from 2B to 27B on Terminal-Bench-Lite and Terminal-Bench-v2.1, respectively.

Offline reinforcement learning (offline RL) can benefit from nearby out-of-distribution (OOD) actions, but estimation errors at these actions may be amplified by bootstrapping. Existing regularization and local-generalization methods control either the admissible OOD region or the influence of generalized targets, often through separate mechanisms. We propose Convex Hull Neighborhood Smooth Dual Generalization (CSDG), which expresses the Bellman backup as an in-sample value target plus a CHN-local correction. This formulation makes the generalized contribution explicit and separates it from the in-sample reference path. The correction is obtained by smoothing in-sample-oriented and OOD-oriented candidates sampled at different perturbation radii. A mixture coefficient lambda scales its contribution to each backup, while the recursive discount remains gamma. Under boundedness and fixed perturbation kernels, we derive an exact one-step correction identity, a time-varying iterate bound, and a fixed-point bound that depends only on the branch discrepancy at the fixed point. We further characterize the implicit policies induced by the idealized operators and give a conditional non-degradation criterion. The practical algorithm approximates these quantities using asymmetric bounded noise and expectile regression, without exact support classification or an additional pessimistic OOD penalty. Experiments on Gym-MuJoCo and AntMaze show strong aggregate performance and stable value estimation. Code is available at: https://github.com/YOUNG-fnxm/CSDG

Cooperative navigation of multi-agent UAVs in complex environments faces key challenges including local optima traps, sparse rewards, learning imbalance among agents, and insufficient cross-scenario generalisation. This paper proposes a multi-agent deep reinforcement learning framework that addresses these issues through coordinated exploration, demonstration exploitation, safe curriculum scheduling, and structure-aware generalisation. First, a perception mechanism combining memory of visited states, directional novelty estimates, and penalty backpropagation enables agents to proactively detect and escape local optima. Second, a hierarchical collaborative demonstration buffer with tiered behaviour cloning manages trajectories by degree of team collaboration and applies differential supervision to the actor network, improving demonstration utilisation under sparse collaborative signals. Third, a safety-aware dual-condition curriculum scheduling mechanism reviews mastered scenarios through back-testing and experience pre-filling during training, suppressing catastrophic forgetting while ensuring both task performance and flight safety. For generalisation, local geometric features computed from sensor readings are abstracted into a domain parameter, through which a structure-aware gating network and mixture-of-experts mechanism condition the policy on local structural patterns rather than scenario-specific coordinates, enabling cross-scenario transfer without exposure to the target environment. The framework is further validated under mixed static-dynamic obstacle settings, showing robust adaptability to dynamic disturbances. Simulation results confirm strong performance in collaboration success rate, navigation robustness, zero-shot cross-scenario generalisation, and dynamic environment adaptability.

A key challenge in modern robotics is to adapt to changing environments, a challenge that is exacerbated when simulations cannot encompass every possible real-world configuration, and therefore Reinforcement Learning (RL) in the physical world becomes necessary. Continual Reinforcement Learning provides the tools to address this challenge; however, both the frameworks and the methods remain underexplored. Autonomous Racing and in particular the RoboRacer competition provide a testing ground for such methods, as learning to drive on a new track-floor combination with the least amount of new experience naturally frames a continual learning problem. This work tries to address this gap by proposing a continual RL framework based on Continual Backpropagation that is able, with only real-world data, to train a generalistic policy on a set of tracks and then fine- tune it within 15 minutes to outperform classical controllers. Furthermore, a comparison method based on offline RL is proposed, and a simulation analysis of the plasticity properties of the methods is conducted.

In this work we investigate reinforcement learning (RL) as a framework for the robust control of parametrized dynamical systems in presence of measurements and model uncertainties. High-dimensional state spaces, expensive numerical solvers, the partial knowledge of the governing equations, and the dependence on physical parameters that may be uncertain or difficult to estimate accurately, make the use of standard RL approaches computationally unfeasible. Indeed, lack of robustness and poor generalization across parameter variations are further amplified in presence of noisy or incomplete measurements, ultimately hampering control performance. To address these challenges, we introduce HypEMBER, a novel RL framework based on the combination of hypernetworks and ensemble learning. In the proposed approach, both the policy and value functions are represented through hypernetworks that generate the weights of the underlying models conditioned on the physical parameters of the system, thereby enabling parametric generalization across different dynamical regimes. In addition, an ensemble of policy and value approximators is employed to quantify epistemic uncertainty, leading to improved exploration strategies and enhanced robustness during and after training. The performance of the proposed framework is assessed on two representative parametrized control problems: (i) the one-dimensional Kuramoto-Sivashinsky equation and (ii) a particle-navigation task in a two-dimensional time-dependent gyre flow, focusing on robustness with respect to measurement noise and parameter misspecification. Numerical results demonstrate that HypEMBER consistently improves training stability and sample efficiency, while achieving superior robustness to uncertainties affecting both the system dynamics and the available observations, in comparison with state-of-the-art RL methods.

Transferring policies across domains poses a vital challenge in reinforcement learning, due to the dynamics mismatch between the source and target domains. In this paper, we consider the setting of online dynamics adaptation, where policies are trained in the source domain with sufficient data, while only limited interactions with the target domain are allowed. There are a few existing works that address the dynamics mismatch by employing domain classifiers, value-guided data filtering, or representation learning. Instead, we study the domain adaptation problem from a generative modeling perspective. Specifically, we introduce DADiff, a diffusion-based framework that leverages the discrepancy between source and target domain generative trajectories in the generation process of the next state to estimate the dynamics mismatch. Both reward modification and data selection variants are developed to adapt the policy to the target domain. We also provide a theoretical analysis to show that the performance difference of a given policy between the two domains is bounded by the generative trajectory deviation. More discussions on the applicability of the variants and the connection between our theoretical analysis and the prior work are further provided. We conduct extensive experiments in environments with various shifts to validate the effectiveness of our method. The results demonstrate that our method provides superior performance compared to existing approaches, effectively addressing the dynamics mismatch. We provide the code of our method at https://github.com/hanyang-chen/DADiff-release

While reinforcement learning with verifiable rewards (RLVR) is widely used to improve the reasoning capabilities of large language models (LLMs), the generalizability of the resulting models remains poorly understood. In this work, we establish the first non-vacuous generalization bounds for parameter-efficient RLVR fine-tuning at the billion-parameter scale. Our approach adapts PAC-Bayes compression bounds to this setting, and addresses the inherent stochasticity of token generation by applying the Gumbel-max reparameterization trick. To operationalize these bounds, we propose the Progressive RLVR framework, which integrates RLVR with on-policy distillation, TinyLoRA, and model quantization. Progressive RLVR empirically retains 84-97% performance of standard LoRA fine-tuning while producing models that are 14,796x more compressible. We show that this framework yields non-vacuous generalization bounds in four domains: mathematical problem-solving, programming, general-knowledge reasoning, and Text-to-SQL. Our bounds exceed the accuracy of the base model by 9-51% and lie within 6-11% of the accuracy of the fine-tuned models.

As autonomous agents are increasingly deployed across diverse operational contexts, aligning their behavior with human intent demands reward functions that remain robust to such changes rather than overfitting to any single environment. Inverse reinforcement learning (IRL) provides a principled way to infer such objectives from human feedback. However, existing analyses of optimal teaching approaches for IRL focus on single-environment, demonstration-only settings, leaving underexplored how heterogeneous feedback modalities and environment dynamics jointly constrain reward functions that generalize across multiple environments. Because demonstrations in one MDP entangle reward information with that environments specific structure, the resulting rewards frequently fail to generalize when the agent is deployed in a new setting. We first analyze how different feedback modalities constrain rewards, showing that, in the unlimited-data regime, comparisons impose strictly stronger global constraints than other modalities. Beyond this theoretical analysis, we introduce a hierarchical machine teaching algorithm for reward learning that operates across multiple MDPs. The algorithm first greedily selects informative environments that expose complementary reward constraints, then strategically queries low-cost feedback within those environments. Empirically, our method achieves substantially lower regret and stronger generalization to held-out environments than uniform teaching baselines under identical feedback budgets, demonstrating the importance of multi-environment, multi-modal teaching for learning dynamics-robust reward functions.

While pessimism counteracts overestimation bias in offline reinforcement learning (RL), being overly conservative has been associated with hindering certain forms of generalization. However, in this paper we demonstrate that being overly pessimistic does not inherently prevent optimal generalization in contextual MDPs (CMDPs). Instead, we argue successful generalization depends not on the amount of pessimism, but whether the pessimistic structure respects the underlying symmetries of the optimal solution. We prove that a mildly pessimistic, non-symmetric value function can generalize worse than an overly pessimistic, symmetric one. In offline RL, the structure of the pessimism is determined by the structure of the dataset coverage. As such, enforcing a symmetric value function can be non-trivial, and might require techniques such as data augmentation (DA). Inspired by our theoretical results, we argue that DA can best be applied through a consistency loss during policy extraction, rather than the common practice of (regular) offline training on an augmented dataset. This is empirically validated using IQL and CQL on a rotationally symmetric reacher environment.

Visual Reinforcement Learning (VRL) has achieved considerable success in solving control tasks. However, generalizing learned policies to new environments remains a major challenge, as agents often overfit to task-irrelevant features in the training environment. To solve this problem, we introduce the concept of decoupling observations into task-relevant and task-irrelevant representations. Building on this idea, we propose a self-supervised Task-Relevant Representation Decoupling (T2RD) algorithm for VRL. This algorithm consists of three components: task-relevant representation consistency, cross-reconstruction, and cross-dynamic prediction. The first two components achieve the decoupling of content and style features, but the resulting content representations are not necessarily task-relevant. To further refine task-relevant features from content representations, we design the third component that introduces dynamic prediction. T2RD achieves State-Of-The-Art (SOTA) generalization performance and sample efficiency in the DeepMind Control Suite and Robotic Manipulation tasks.

Humans often acquire new skills by observing others, since observed behaviors implicitly reveal how to act in an environment. However, observations drawn from a heterogeneous population introduce conflicting behavioral signals, making it difficult to determine which behaviors are worth imitating. We address this challenge with General Reward Inference and Disentanglement (GRID), a social learning method that extracts universally useful behaviors from a heterogeneous population of demonstrators pursuing different goals. GRID decomposes per-agent reward functions into a general reward, capturing behaviors shared across all agents, and specific rewards, capturing individual preferences and objectives. Training exclusively on the general reward provides a new paradigm of generalist pretraining. It yields a generalist agent that internalizes universal environmental competencies, such as safety and basic task proficiency, without the mode-averaging bias that afflicts standard learning from demonstration techniques. This generalist serves as a superior prior for fine-tuning to downstream tasks, including preferences unseen during training. Experiments across a synthetic basis function decomposition, multi-agent Craftax, and a continuous autonomous driving simulator (Highway-Env) confirm that GRID successfully disentangles reward structure in a semantically meaningful way, outperforms standard learning from demonstration baselines, and enables more efficient and stable specialization.

Reinforcement learning (RL) often suffers from performance degradation when deployed in environments that differ from those encountered during training. Existing techniques such as domain randomization (DR) mitigate this, but require access to diverse training environments and full trajectory observability, assumptions that fail in privacy-preserving or restricted scenarios where only scalar performance metrics are available. We propose Generalization via Evolutionary Reward Shaping (GERS), a bilevel optimization approach to improve generalization on unseen test environments using only scalar feedback from validation environments. At the lower level, an RL agent guided via a reward function shaped by the upper level learns a policy on a limited set of training environments with accessible trajectory data; at the upper level, CMA-ES optimizes the reward shaping parameters to maximize the cumulative unshaped reward on separate validation environments for which trajectory access is unavailable. Results on continuous control tasks indicate that GERS outperforms the standard RL baseline on unseen test environments. GERS performance is comparable to DR, despite DR treating the combined set of training and validation environments of GERS as a single training set that requires trajectory access, whereas GERS cannot access validation trajectories. These results confirm that GERS effectively enhances generalization under restricted data access constraints.

Model post-training, and in particular reinforcement learning (RL), is one of the primary mechanisms by which developers can shape models' values and behaviors. However, as models become increasingly evaluation and training aware, they may be motivated to resist training when the perceived objective conflicts with their current values, undermining developers' ability to detect misalignment and correct model behavior through further training. In this paper, we demonstrate generalization hacking, in which a model collects reward during RL while preventing the rewarded behavior from generalizing. We construct a model organism on Qwen3-235B-A22B, finetuning on synthetic documents describing training awareness and self-inoculation, a novel mechanism in which the model frames compliance as context-specific in its chain of thought, without demonstrating or instructing either behavior. The model organism achieves train-time harmfulness comparable to controls while maintaining a persistent ${\sim}15$ percentage point compliance gap across 700 steps of RL. Additionally, a control organism trained only on training awareness documents independently discovers inoculation-like reasoning under RL pressure, developing its own compliance gap despite never being exposed to the concept. Because the generalization-hacking organism receives high reward throughout, standard training metrics provide no signal that generalization has failed. Our results constitute the first demonstration that a model can actively resist RL behavioral modification while maintaining high reward, suggesting that as models become more capable and training-aware, they may be able to undermine the training process itself.

Continual reinforcement learning (RL) aims to produce agents that never stop adapting to new tasks. A key question is how this interacts with the diversity of tasks an agent experiences. Prior work has shown that training on many diverse tasks leads to agents with strong zero-shot and in-context adaptation. However, this work evaluated agents after they'd stopped learning, i.e. with frozen weights. How task diversity affects an agent's ability to continue learning over a sequence of distribution shifts remains unclear. We introduce Banyan, a GPU-accelerated continual RL domain where one can parametrically control three independent axes that define a task: the map layouts an agent must navigate, the objects it must interact with, and the hierarchical structures of sub-goal dependencies. We find that increasing diversity along each axis induces systematic transfer -- that is, agents begin training on a new task distribution near the performance attained on the previous one, even when the shift changes the structure of the optimal policy. While increasing diversity improves systematic transfer, we find that too much diversity inhibits a learner's ability to continue adapting to new task distributions. As diversity increases, learners plateau in the success rate they achieve on new tasks, yet continue improving on old tasks -- even without further exposure to them. We find this phenomenon manifests across continual learning algorithms, memory architectures, architecture sizes, and in Kinetix -- a physics-based control domain. We release Banyan as a domain for running controlled experiments that study continual RL in the many-tasks regime. Code is available at https://github.com/nhshah15/banyan.

This work presents a logic-driven framework to evaluate the performance of reinforcement learning (RL) algorithms in their ability to generalize to unseen tasks. Our framework defines a family of inductive reach-avoid tasks, characterized by structural similarities in task dynamics, enabling evaluation of generalization capabilities. We introduce a neural certificate function that validates trajectories generated by RL algorithms by enforcing key conditions, thereby serving as a litmus test for RL generalization. We empirically demonstrate our method's capability in certifying generalization for several state-of-the-art generalizable RL algorithms on challenging continuous environments. Our results show that a lower percentage of certificate function violations correlates with a higher number of test tasks successfully solved, highlighting the effectiveness of our framework in evaluating and distinguishing generalization capabilities of RL algorithms. This work provides a principled approach for benchmarking RL generalization.

Inductive generalization is a framework for reinforcement learning (RL) generalization in which inductively related task instances admit inductively related policies. Prior work captures this structure via a higher-order policy-evolution function learned directly with RL, but suffers from poor training scalability: as training tasks grow, aggregated reward feedback becomes noisy and conflicting, destabilizing training and weakening generalization. We propose DIBS, a decoupled behavioral cloning approach that separates learning task-specific policies from learning the evolution function. We first learn individual teacher policies per task via standard RL, then fit the evolution function via behavioral cloning on teacher-labeled state-action pairs. This replaces noisy reward aggregation with dense, stable supervision. DIBS achieves significant improvements in both training stability and zero-shot generalization against existing RL and meta-RL algorithms.

Large language model agents trained with reinforcement learning (RL) often learn brittle, task-specific shortcuts. We hypothesize that agents generalize better when their successful trajectories are structurally compressible, decomposed into a small set of reusable abstract patterns. To formalize this, we introduce ReuseRL, which grounds agentic RL in the Minimum Description Length (MDL) principle. ReuseRL extracts a shared skill dictionary from successful trajectories and augments the RL objective with a segmentation cost, explicitly penalizing idiosyncratic behaviors that encode poorly. We prove a PAC-Bayes bound guaranteeing that a dictionary extracted from successful trajectories has bounded expected description length on future successful behavior. Across ALFWorld, TextWorld-Cooking, and Countdown-Stepwise, ReuseRL improves in- and out-of-distribution success over vanilla GRPO and strong round-length baselines.

Combinatorial generalization remains a central challenge in Deep Reinforcement Learning (DRL). Classical planning provides a simple yet challenging setting to study this problem through explicit relational descriptions, without requiring learning from perception. In sparse-reward domains, standard RL exploration via real-time search is ineffective, and learning-based planning methods often rely on expert demonstrations, hindsight relabeling, or random walks from the goal state. In contrast, planners rely on best-first search methods such as $\mathrm{A}^\star$ to solve problems from scratch. We propose a self-improving $\mathrm{WA}^\star$ learning framework in combination with a value heuristic represented by a Relational Graph Neural Network: the heuristic guides search, and the resulting search data updates the heuristic via $Q$-learning. This loop yields heuristics that can function as general policies and solve new instances even without search, where DRL otherwise fails, as we show on puzzles such as Sokoban, PushWorld, The Witness, and the 2023 International Planning Competition benchmarks. Notably, we demonstrate strong zero-shot generalization: For example, heuristics trained on Blocksworld instances with fewer than 30 blocks successfully solve instances with 488 blocks without search.