arXiv subjects

Benjamin Eysenbach

Publications and source records attributed to Benjamin Eysenbach.

At least 19 recordsLinked to original sources

arXiv2026

Geometric Time Score Estimators: Bridging Self-Predictive Representations and Generative Occupancy Models For Reinforcement Learning

In reinforcement learning (RL), temporal reasoning is essential for estimating the density of future outcomes under a policy's occupancy measure. While self-predictive representations have been a dominant paradigm for occupancy density estimation, we find that they yield surprisingly poor estimates. On the other hand, generative occupancy models can produce remarkably accurate samples from the occupancy measure. Yet, they cannot easily recover occupancy densities, limiting their use for control despite their expressivity. In this paper, we propose Geometric Time Score Estimators (GTSE), a framework that bridges the two paradigms by learning occupancy density estimators along probability paths. We analyze how GTSE implicitly induces better occupancy estimators through improved representation learning and show how to directly extract self-predictive representations from this framework. Through experiments across several locomotion and manipulation tasks, we show that GTSE improves performance over state-of-the-art methods by $2\times$ to $6\times$ in goal-conditioned RL settings, and that the representations learned using our framework transfer effectively to downstream RL tasks. Website: https://aravindvenu7.github.io/website/gtse/

cs.LG↗

arXiv2026

Unifying Goal-Conditioned RL and Unsupervised Skill Learning via Control-Maximization

Unsupervised pretraining has driven empirical advances in goal-conditioned reinforcement learning (GCRL), but its theoretical foundations remain poorly understood. In particular, an influential class of methods, mutual information skill learning (MISL), discovers behaviorally diverse skills that can later be used for downstream goal-reaching. However, it remains a theoretical mystery why skills learned through MISL should support goal-reaching. A subtle challenge is that both GCRL and MISL are umbrella terms: Different GCRL tasks use distinct criteria for measuring goal-reaching performance, while different MISL methods optimize distinct notions of behavioral diversity. We address this challenge and unify GCRL and MISL as instances of control maximization. We identify three canonical GCRL formulations and prove that they are fundamentally inequivalent: They can induce incompatible optimal policies even in the same environment. Nevertheless, they all share a common interpretation: A well-performing goal-conditioned policy is one whose future trajectory is highly sensitive to the commanded goal, with the precise notion of sensitivity determined by the GCRL formulation. Noting that MISL objectives can be understood as measures of skill sensitivity akin to goal sensitivity, we show that MISL objectives are bounded by formulation-specific downstream goal sensitivities. These bounds establish a precise correspondence between MISL methods and downstream GCRL tasks: For every GCRL formulation, there exists a matching MISL objective for which more diverse skills afford greater downstream goal sensitivity. Our results thus lay a theoretical foundation for RL pretraining and have important practical implications, such as suggesting which pretraining objectives to use when a user cares about a specific class of downstream tasks.

cs.LG↗

arXiv2026

Training Controllable World Models via Action Influence Maximization

Action-conditioned video generative models trained on successful, goal-directed demonstrations tend to ignore the commanded action, rendering a plausible expert-like future regardless of what is asked of them. We trace this failure to training data that lack sufficient action influence. We analyze this metric and prove that it upper bounds the action influence that any model trained on that data can express. Expert demonstrations, whose narrow action distributions let next states be predicted without attending to the action, set a low ceiling that no amount of model capacity alone can raise. Raising it requires collecting different data. We introduce Action Influence Maximization (AIM), an algorithm for actively collecting training data for improving dataset action influence. We compare the performance and data-efficiency of AIM against exploration heuristics and intrinsic-motivation methods on simulated navigation and manipulation tasks. In pixel-based manipulation, AIM is the only method that consistently improves controllability: 500 actively collected episodes added to 3k demonstrations match a world model trained on 300k expert episodes on end-effector controllability, an $85\times$ reduction in data.

cs.RO↗

arXiv2026

Measuring Empowerment from Offline Data for Intrinsically Motivated Exploration

Large, unlabeled datasets of prior experience have become increasingly available for agentic tasks in domains such as robotics and coding. A common recipe for these tasks is to train a policy using the prior data, and then further improve it with online rollouts. But while much work has studied how to use both the offline and online data for policy extraction, far less work has been done to understand how the prior data can be used to efficiently \textit{explore} during the online learning phase. Previous work has attempted to formalize this idea by defining the \textit{empowerment}, a measure of how many future states an agent can realize; states with higher empowerment require more exploration as there are more futures to explore. However, existing methods for estimating empowerment require online policy rollouts in the environment. We propose the first method for estimating empowerment from unlabeled prior data, without any environment interaction. By jointly learning an empowerment-maximizing policy together with its discounted state-occupancy measure, our estimator can be trained on unlabeled trajectories generated from any unknown policy. We demonstrate that our empowerment estimator qualitatively gives intuitive empowerment measures and show that the pretrained estimator can be used to improve exploration in the online goal-conditioned reinforcement learning setting. Code is available at \url{https://github.com/ishirgarg/offline-empowerment}.

cs.LG↗

arXiv2026

Hyperspherical Contrastive Representations Unlock Planning and Exploration

We develop a self-supervised reinforcement learning algorithm that learns to build towers of up to eight cubes from scratch, without prior data, rewards, demonstrations, or pretrained models. Our method, Contrastive Planning via Spherical Linear Interpolation (C-Slerp), uses contrastive representations as reusable building blocks and repeatedly composes them to infer a sequence of intermediate waypoints. Under our hyperspherical parameterization, the learned representations allow intermediate waypoints to be inferred in closed form through spherical linear interpolation, without training a separate high-level policy or performing graph search. Importantly, the representation geometry also tells us how to adjust the number of waypoints based on the difficulty of each goal-reaching problem. Experiments show that our method succeeds on long-horizon navigation and robotic manipulation tasks, including settings where strong baselines fail to achieve a single success. Ablations further show that using a fixed number of waypoints substantially degrades performance, highlighting the importance of adapting the plan to task difficulty.

cs.LG↗

arXiv2026

Scaling Empowered Agents

Empowerment measures the capacity of an agent to actively control its environment. In this work, we explain the connection between skill learning and empowerment. While skill learning is not empowerment, we introduce a simple scalable recipe for empowerment maximization: empowerment corresponds to the number of executable skills from a state, so maximizing empowerment corresponds to reaching states that are good starting points for many skills. Our recipe applies to arbitrary underlying skill learning methods. Whereas most prior work on empowerment looks at tabular settings, this recipe enables us to study empowerment at a scale where we can realize theoretically-predicted but empirically-elusive properties of empowerment (e.g., gathering resources, environment modification, survival). Experiments show that empowerment maximization unlocks interesting emergent properties, such as building a house in Crafter. Website and code at \href{https://empowerment-geometry.github.io/scaling-empowerment/}{this https url}.

cs.AI↗

arXiv2026

A Unified Bellman Operator for Safety-Critical Reinforcement Learning

Reinforcement learning in safety-critical domains requires maximizing task performance while strictly adhering to safety constraints. Existing safe reinforcement learning paradigms typically force a trade-off: they either require a priori knowledge to provide strict safety guarantees (e.g., safety filters), or they enable joint learning but only satisfy safety constraints on average. In this work, we propose a novel Bellman operator that unifies performance and safety objectives into a joint value function. We show that temporal difference learning with the joint Bellman operator converges under a two-timescale stochastic approximation framework. On the fast timescale, the safety value of the learning joint policy is estimated, while the joint value is estimated on the slow timescale. Convergence is ensured by formulating the limiting dynamics as an occupation-averaged differential inclusion, and showing that it asymptotically converges to a set of limiting optimal safety-constrained task value functions. Theoretically, once converged, the resulting optimal policy maximizes task return while maintaining safety at all times. Empirical evaluations on continuous control tasks with neural approximations demonstrate stable convergence with near-zero safety violations at test time.

cs.LG↗

arXiv2026

The Geometry of Empowerment

Empowerment captures the capacity for an agent to actively control its environment. While conceptually appealing as an information-theoretic quantity, the connection between empowerment and structurally central states that provide broad access to future outcomes has remained an open question. In this work, we link empowerment maximization and skill-learning methods to provide new geometries for interpreting and analyzing empowerment. Our analyses answer longstanding open questions on the connections between empowerment and structural centrality. Our analyses also reveal distinctions between information and reward geometries, highlighting important theoretical implications to build scalable empowerment-maximization methods. Website and code can be found at https://empowerment-geometry.github.io/.

cs.LG↗

arXiv2026

1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities

Scaling up self-supervised learning has driven breakthroughs in language and vision, yet comparable progress has remained elusive in reinforcement learning (RL). In this paper, we study building blocks for self-supervised RL that unlock substantial improvements in scalability, with network depth serving as a critical factor. Whereas most RL papers in recent years have relied on shallow architectures (around 2 - 5 layers), we demonstrate that increasing the depth up to 1024 layers can significantly boost performance. Our experiments are conducted in an unsupervised goal-conditioned setting, where no demonstrations or rewards are provided, so an agent must explore (from scratch) and learn how to maximize the likelihood of reaching commanded goals. Evaluated on simulated locomotion and manipulation tasks, our approach increases performance on the self-supervised contrastive RL algorithm by $2\times$ - $50\times$, outperforming other goal-conditioned baselines. Increasing the model depth not only increases success rates but also qualitatively changes the behaviors learned. The project webpage and code can be found here: https://wang-kevin3290.github.io/scaling-crl/.

cs.LG↗

arXiv2026

Up and Down the Abstraction Ladder: Code-Based Skills for Language Agents

Language agents struggle to act and learn in environments that require long sequences of low-level actions. Code-based abstractions can make these agents more productive by letting them invoke reusable skills instead of repeatedly selecting individual actions. The code handles recurring local decisions, while the language model decides which skills to use and how to combine them. Yet abstractions are leaky, and situations beyond a skill's capabilities may require a return to primitive actions. Motivated by this tradeoff between productivity and flexibility, we systematically study how code-based action abstraction affects the performance, inference cost, and learning of language agents. We study this in NetHack, a challenging, long-horizon game environment, using CodeHack, our library of code-based skills with natural-language descriptions. We use this library to compare agents restricted to primitives with those using semantic skills alone or in combination with primitives. We evaluate these agents in three settings: zero-shot prompting, supervised fine-tuning, and reinforcement learning. Across a broad zero-shot evaluation on NetHack, we find that compared with primitives, skills nearly triple game progression, while reducing inference cost per episode by 86%. Combining skills with primitives retains much of this benefit while preserving a path back down to low-level actions. Finally, in RL, we find that skill-based agents learn significantly faster than agents acting on primitives, achieving a 7.2x larger average gain in dungeon level over the same training budget. These results show that a supplied skill library can improve performance, efficiency, and learning, while retaining primitives provides flexibility when the library is insufficient. We release CodeHack together with training and evaluation code.

cs.AI↗

arXiv2026

Bridging Control, Inference, Transport, and Thermodynamics: From Theory to Applications in Learning

The last decade has seen the development of powerful methods for learning complex structure from high-dimensional data. These advances have brought to the foreground fundamental connections between subdisciplines of physics, applied mathematics, and machine learning. In this review, we bring together some of these ideas, often expressed in different languages, to highlight a conceptual thread that links five distinct fields: control theory, optimal transport, probabilistic inference, non-equilibrium thermodynamics, and machine learning. A common theme is the optimization of free-energy-like functionals under dynamical or statistical constraints. We offer a guided tour through this thread and present selected applications in reinforcement learning, variational inference, and generative modeling. The review does not assume prior familiarity with these topics, and begins with principles originating from physics.

cond-mat.stat-mech↗

arXiv2026

Three Steps at a Time: Learning Representations from Action Sequences in Contrastive RL

While self-supervised approaches to reinforcement learning have achieved strong results by learning representations of states and actions, a key open question is the time scale over which actions should be modeled. Departing from the standard formulation relying on single-step actions, we extend contrastive reinforcement learning (CRL), a prototypical self-supervised method, to operate over action chunks, and find that this results in large, pervasive gains across established offline and online benchmarks: +31.7% and +93.1% across 18 and 11 environments respectively. While action-chunking-driven gains are generally explained through the ability to model non-Markovian, temporally extended policies, and to propagate unbiased multi-step returns, interestingly, we find that these arguments only partially apply to CRL. Our empirical studies suggest that, in the context of CRL, an action chunk carries more information about the goal than a single action, measurably improving the critic's representations, and rendering the algorithm significantly more effective.

cs.LG↗

arXiv2026

Behavior-Consistent Deep Reinforcement Learning

Reinforcement learning (RL) often exhibits high variance across training runs, leading to unreliable performance and posing a major challenge to deployment in real-world domains. In this work, we address the challenge of cross-run policy divergence by formalizing the problem of behavior-consistent RL, where the objective is to obtain policies that are both high-performing and distributionally similar across training runs. Our key observation is that maximum-entropy RL provides a direct mechanism for controlling behavioral divergence by anchoring runs to a common (uniform) prior. We prove that, for Boltzmann policies, choosing the temperature proportional to $Q$-function disagreement bounds the pairwise KL divergence between the induced policies. However, we also show that naïvely increasing entropy might impair policy optimization while amplifying off-policy error. Building upon these observations, we propose $Q$-value Expectile Disagreement (QED), a state-dependent temperature schedule that uses double-critic disagreement as a single-run proxy for cross-run disagreement. Empirically, we demonstrate that across 18 continuous-control tasks, QED reduces across-run divergence by two orders of magnitude without sacrificing performance, resulting in a considerable reduction in return variance at modest sample-efficiency costs.

cs.LG↗

arXiv2026

LeAct: Learning to Reason from Expert Actions

Modern reasoning models depend on reasoning data, today sourced from human annotations or distilled from stronger LLMs. However, a rich and largely untapped source of supervision lies in expert systems (e.g., game engines, classical planners, theorem provers), which routinely produce near-optimal actions across diverse domains. But these experts are silent: they commit to an action without writing down the chain of thought (CoT) behind it. Recovering that CoT as natural-language reasoning would distill expert knowledge into a student that generalizes beyond the demonstrated actions. We treat it as a latent variable and study how to recover it from the action alone. Our approach, LeAct (Learning to reason from Actions), optimizes this latent variable: the student samples candidate CoTs for each expert action, and we retain those that measurably improve its own probability of recovering the action. Across imperfect-information games at multiple scales and a simulated robotics benchmark, LeAct reaches the solver's numerical floor on small enumerable games. At larger scale, it is $5\times$ closer to the solver than the strongest expert-iteration baseline. At Flop Hold'em ($\sim 10^9$ infosets), LeAct wins head-to-head by $+60$ mbb/g, and on the robotics probe it is the only training recipe that improves on direct imitation. We present a principled framework and the result: expert systems become a categorically new source of reasoning teachers for foundation models.

cs.LG↗

arXiv2026

Can We Really Learn One Representation to Optimize All Rewards?

As unsupervised pretraining becomes increasingly ubiquitous in reinforcement learning, a more thorough theoretical understanding of these methods becomes of equal importance to their empirical success. We focus on the setting of unsupervised learning via interaction, where the forward-backward (FB) representation learning serves as a prototypical and popular example. In this paper, we shed light on FB by formally contextualizing the method within a broader class of recent methods that use regression to obtain a low-rank approximation of a successor measure ratio. Our analysis clarifies when FB representations can exist and how the low-rank approximation converges in practice. Building upon the theory, we propose a variant of FB that is both more amenable to theoretical understanding and simpler to optimize in practice. Experiments in didactic settings, as well as in $10$ state-based and image-based continuous control domains, demonstrate that our method converges to desired representations with $10^5 \times$ smaller errors than FB, achieving $+24\%$ improved zero-shot performance on average. We also demonstrate that zero-shot policies inferred by our algorithm provide an efficient initialization if the user prefers further fine-tuning on downstream tasks. Our project website is available at https://chongyi-zheng.github.io/onestep-fb.

cs.LG↗

arXiv2026

Self-Supervised Goal-Reaching Results in Multi-Agent Cooperation and Exploration

For groups of autonomous agents to achieve a particular goal, they must engage in coordination and long-horizon reasoning. Rather than relying on complex reward functions and explicit cooperation mechanisms, we ask what minimal ingredients are required for effective coordination and exploration to emerge in multi-agent settings. We investigate this question through self-supervised goal-reaching, where agents aim to maximize the likelihood of visiting a goal state rather than maximizing a reward. Despite a sparse feedback signal, we present empirical results that show self-supervised goal-reaching techniques enable agents to learn from such feedback. On MARL benchmarks, self-supervised goal-reaching outperforms alternative approaches that have access to the same sparse reward signal. Furthermore, we empirically demonstrate that multi-agent self-supervised goal-reaching approaches can be more robust than single-agent strategies. While there is no explicit exploration mechanism, this approach explores nontrivial intermediate coordination strategies in sparse settings where alternative approaches fail to achieve a single success.

cs.LG↗

arXiv2026

BuilderBench: The Building Blocks of Intelligent Agents

Today's AI models learn primarily through mimicry and refining, so it is not surprising that they struggle to solve problems beyond the limits set by existing data. To solve novel problems, agents should acquire skills by exploring and learning through experience. Finding a scalable learning mechanism for developing agents that learn through interaction remains a major open problem. In this work, we introduce BuilderBench, a benchmark to accelerate research into agent training that centers open-ended exploration. BuilderBench requires agents to learn how to build any structure using blocks. BuilderBench is equipped with (1) a simulator of a robot interacting with various physical blocks, and (2) a task-suite with over 50 diverse target structures that are carefully curated to test an understanding of physics, mathematics, and long-horizon planning. Agents are provided with a target structure at the start, and can interact with the environment for multiple episodes to experiment and learn various skills for building the structure. Solving these tasks requires \emph{embodied reasoning} in a way that is not reflected in words but rather in actions, experimenting with different strategies and piecing them together. Our experiments with multiple state-of-the-art frontier language model based agents and tabula rasa reinforcement learning algorithms show that these agents cannot solve any of the non-trivial tasks in the BuilderBench. Our analysis throws light on the lack of exploration abilities in these models.

cs.AI↗

arXiv2026

On the Role of Computation in Reinforcement Learning

How does the amount of compute available to a reinforcement learning (RL) policy affect its learning? Can policies using a fixed amount of parameters, still benefit from additional compute? The standard RL framework does not provide a language to answer these questions formally. Empirically, deep RL policies are often parameterized as neural networks with static architectures, conflating the amount of compute and the number of parameters. In this paper, we formalize compute bounded policies and prove that policies which use more compute can solve problems and generalize to longer-horizon tasks that are outside the scope of policies with less compute. Building on prior work in algorithmic learning and model-free planning, we propose a minimal architecture that can use a variable amount of compute. Our experiments complement our theory. On a set 31 different tasks spanning online and offline RL, we show that $(1)$ this architecture achieves stronger performance simply by using more compute, and $(2)$ stronger generalization on longer-horizon test tasks compared to standard feedforward networks or deep residual network using up to 5 times more parameters.

cs.LG↗