Reinforcement learning is a natural way to post-train LLM agents for long-horizon interactive tasks judged only by end-of-task verification, yet a shared belief holds that outcome-only RL soon hits a ceiling on small open models. Recent work therefore compensates around the training with denser rewards, SFT priors, skill libraries, curated memory, or multi-agent orchestration. We argue the ceiling is an artifact of two failures of common practice. Signal starvation: group-relative RL with sparse outcome-only rewards yields a gradient only when a task's rollout group mixes successes and failures, so under-scaled exploration silences exactly the hardest, most instructive tasks. Policy drift: squeezing many updates out of a small task pool degrades the policy itself, as an unanchored objective lets the sampling distribution collapse exactly when saturation has already made informative groups rare. We present CANOPY (Coverage-ANchored On-PolicY RL), a minimalist protocol attacking both directly: scale same-task exploration until the natural signal reappears, keep every update on-policy, KL-anchored, and confined to the agent's own action tokens, then cash in an enlarged interaction budget at test time. On AppWorld, a long-horizon interactive coding benchmark, a Qwen3-14B policy trained with CANOPY through environment interaction alone--without task-specific supervision, auxiliary credit signals, or elaborate agent scaffolding--topped the public leaderboard (Feb. 2026; Test-Normal TGC 86.9, Test-Challenge 67.6), and the same design principles lift Qwen3.5-9B on SWE-bench Verified by 16.6 points. Agentic RL alone thus internalizes long-horizon capability directly into a small open model; we plan to release the complete training stack at https://github.com/AlibabaResearch/SignalCoverageRL.
Sheeraja Rajakrishnan, Alexander G. Ororbia, Travis Desell +1cs.LG
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.
Szymon Miłosz, Piotr Duch, Szymon Grabowskics.LG cs.AI stat.ML
Searchless chess networks reach human master strength from a single forward pass by imitating a stronger teacher: the strongest, Leela Chess Zero's (Lc0) released Chessformer, distills the visit counts of an AlphaZero-style Monte Carlo Tree Search (MCTS). Imitating a search is a poor proxy for playing without one, so we fine-tune for single-pass strength with self-play reinforcement learning (RL). Its exploration is usually supplied by an entropy bonus, the reverse Kullback-Leibler (KL) divergence to uniform. We replace it with a forward, mass-covering KL toward the network's own MCTS prior (prior-directed exploration), so exploration covers the moves the prior judges promising, and pair it with an entropy-adaptive sampling temperature, set by the value head's outcome uncertainty, that sharpens once a position is decided. In about two thousand steps it raises puzzle accuracy from 93.9% to 94.9% on a 100,000-puzzle suite and mate-in-four accuracy from 77% to 81% while holding searchless strength at or slightly above the base. Measuring tactical accuracy and playing strength together across a matched-compute sweep, we find the two dissociate: accuracy gains fall in a one-point band while ratings straddle the base, and a control fine-tuned on puzzles alone posts the study's largest tactical gains while shedding roughly 260 Elo; a better puzzle-solver is not thereby a stronger player. Distribution-level measurements show what anchoring buys: without a regularizer self-play collapses onto a single line of play, and the puzzles newly solved are the near misses whose winning move the prior kept alive. The forward-KL prior tops the rating ladder, statistically tied with a reverse-KL anchor that concentrates twice as hard and drops the hardest solutions the mass-covering prior keeps in support.
Reinforcement Learning with Verifiable Rewards (RLVR) significantly improves LLM reasoning but often causes a drop in policy entropy, leading to narrowed reasoning coverage and degraded pass@$k$ for large $k$. While existing methods mitigate this entropy collapse through algorithmic regularizations, cross-model non-parametric perturbation is also neglected. In this work, we propose a simple yet effective approach to preserve the generative diversity of LLMs during RLVR. Instead of relying solely on internal exploration, we force the target model to generate answers based on partial reasoning trajectories generated by a smaller, weaker language models. These unfamiliar prefixes effectively disrupt over-confidence and encourage the exploration of distinct reasoning paths. We empirically study the potential of outer prefixes, revealing the mechanism of the impact of distributional discrepancy to the exploration dynamics in RLVR training. Experiments across multiple mathematical benchmarks show that our method consistently outperforms vanilla RLVR. Notably, the performance gain becomes increasingly pronounced as $k$ scales up, demonstrating a substantial expansion of reasoning coverage. Furthermore, our approach efficiently mitigates entropy collapse without requiring additional SFT, intricate reward designs, or complex prompting.
Massively parallel simulation changes the data regime in which off-policy reinforcement learning (RL) is trained, challenging stabilizers designed for data-limited replay. Through controlled experiments across eight benchmark families, we show that these stabilizers are data-regime-dependent: parameter normalization helps with narrow replay coverage but restricts value fitting when data are abundant, while clipped double-Q can be relaxed in high-throughput manipulation. Age-biased replay weighting improves learning efficiency across regimes, especially with limited network capacity. Based on these findings, we propose WarpSAC, a regime-aware family of off-policy RL algorithms. WarpSAC uses Sample Weight Decay for efficient exploitation and provides two variants: WarpSAC-L (Norm ON, clipped double-Q) for data-limited CPU-scale training, and WarpSAC-A (Norm OFF, single-Q) for data-abundant GPU-parallel training. WarpSAC improves normalized score--step AUC over FlashSAC by 4.5% across nine CPU-scale environments and 23.1% across fourteen GPU-parallel environments. It increases UnitreeG1TransportBox-v1 success rate from 19.8% to 96.4%, improves mean normalized wall-time AUC on MuJoCo Playground by 19.1%, and achieves 36.4% faster sim-to-real deployment on Unitree G1 than FlashSAC. These results show that scalable off-policy RL should adapt its stabilizers to the available data regime.
Khang Luong, Nam Nguyen, Hoang Ta +2cs.LG cs.AI stat.ML
We propose Variance Driven Exploration (VarDE), a principled approach for pure exploration in highly stochastic environments, where the exploration process is dominated by stochastic variance. VarDE is built on a fundamental principle: sampling effort should be allocated to minimize the uncertainty of the final decision. We formalize the uncertainty of the final decision through a smooth decision function and derive allocation rules that explicitly capture how stochastic noise in individual components affects the reliability of the final output. We apply this methodology to three core problems of pure exploration -- Best Arm Identification (BAI), Monte Carlo Tree Search (MCTS), and Best-Policy Identification (BPI) -- with theoretical guarantees on variance decay and simple regret. Empirically, we demonstrate consistent and significant improvements of VarDE over existing methods, with especially strong gains in highly stochastic environments.
We study decentralized multi-player reinforcement learning in episodic tabular Markov decision processes (MDPs) under three forms of information asymmetry: (A) unobserved actions with common rewards, (B) observed actions with independent rewards, and (C) unobserved actions with independent rewards. Players cannot communicate during learning but may agree on a protocol a priori. For Problems A and B we propose \texttt{mQ-learning} and \texttt{mQ-learning-intervals}, achieving $\tilde{O}(\sqrt{H^4 S A_{\text{joint}}\, T})$ regret, where $H$ is the horizon, $S$ the state count, $T = KH$ the total steps, and $A_{\text{joint}} = \prod_{i=1}^M |\mathcal{A}_i|$ the joint action space across $M$ players. For Problem C we give \texttt{mEXC} and \texttt{mEXC-Bellman}, two-phase explore-then-commit algorithms with regret $\tilde{O}(H (S A_{\text{joint}})^{1/3} T^{2/3})$. Against the centralized joint-action benchmark, decentralized learning under information asymmetry matches the single-agent Q-learning rate of \cite{jin2018q} up to logarithmic factors. Because $A_{\text{joint}}$ grows exponentially in $M$, the bounds are most meaningful for small $M$ or small per-player action sets.
Md Rafid Islam, Rafsan Jany, Zahid Hasan +1cs.LG cs.AI
Personalized Federated Reinforcement Learning (PFRL) takes a decentralized approach to storing and accessing information based on past experiences while keeping each client's data private during the learning of each client's policy. Many current methods for PFRL rely heavily on exploiting existing reinforcement learning reward signals to derive an optimal policy for each client, thereby neglecting exploration in non-stationary or sparse-reward environments. In this work, we introduce a new exploration-driven framework, Exploration-Driven Personalized Federated Reinforcement Learning via Intrinsic Motivation (EDPFRL-IM), that leverages an inherent curiosity-driven exploration at each client to promote local exploration and protect client privacy. Furthermore, to facilitate policy discovery via exploration in previously unexplored state spaces, clients add an intrinsic random network distillation (RND) signal to their extrinsic reward. Additionally, the server does not have access to clients' raw experiences or local gradient estimates; instead, the server sends global exploration priors and collects minimal novelty summaries from each client to enable both diverse and coordinated exploration among clients. Experiments in benchmark environments show that our framework outperforms average PFRL benchmarks in policy personalization and sample efficiency, primarily in delayed and sparse reward systems. Overall, EDPFRL-IM enables the integration of a flexible exploratory learning structure into federated reinforcement learning systems while preserving client privacy.
Reinforcement learning (RL) with general utility extends classic RL by optimizing an arbitrary utility functional of the policy-induced occupancy measure, thereby enabling a broader range of applications. However, previous work on general utility RL typically assumes the evaluation utility is fixed and correctly specified. In practice, the utility used at deployment can deviate from the training one, creating a robustness gap that prior work does not address. Motivated by this, we propose robust general-utility RL, a minimax learning framework that trains policies against utility misspecification within a prescribed uncertainty set. Our framework strictly generalizes standard general-utility RL while also providing a unified view of many existing RL frameworks, including reward-robust RL and constrained RL, through appropriate choices of the utility uncertainty set. We further develop provably convergent stochastic algorithms for two regimes. For concave utilities, we develop a projected stochastic gradient descent-ascent method and establish stationarity guarantees. For the more challenging nonconcave regime, we propose a stochastic prox-extragradient algorithm that mitigates ill-posed behavior induced by nonconcavity, with convergence guarantees to approximate first-order stationarity. Experiments on LLM safety alignment and exploration maximization tasks further corroborate the convergence behavior consistent with our theory.
Although reinforcement learning with verifiable rewards (RLVR) has improved the performance of large language models (LLMs) across a variety of reasoning tasks, there is significant debate as to whether RLVR expands the reasoning capability boundary, or just improves sampling efficiency. In this paper, we investigate the nature of test-time exploration in RLVR-trained LLMs by employing controlled maze-solving experiments and extracting a tree structure from mathematical reasoning traces (BODHI-Trees) based on semantic equivalence. This helps us delineate between entropy arising from stylistic variations and genuine inferential branching. Our findings demonstrate that the policy entropy collapse observed in RLVR models is not merely syntactic, and is accompanied by a significant reduction in semantic branching entropy. While RLVR improves adherence to environmental constraints and backtracking capabilities, it constricts the space of continuations; we provide evidence suggesting that this might be responsible for the sample efficiency gains of RLVR, albeit at the cost of genuine rollout diversity.
Jim Dilkes, Vahid Yazdanpanah, Sebastian Steincs.AI cs.CL cs.LG
Post-training Large Language Models (LLMs) with Reinforcement Learning (RL) has become an important tool for improving model capabilities, but the LLM action-space structure introduces challenges distinct from classical RL, with implications for inducing exploration. New methods are required that leverage the broad knowledge and flexibility of pre-trained LLMs to deliberately generate diverse experience at training time. We propose Instruction-Conditioned Exploration (ICE), which supplements task prompts during training with one of several distinct instructions, increasing the coverage of behaviours attempted. To facilitate ICE, we propose Asymmetric-RL/SD, a combined Reinforcement Learning and Self-Distillation training objective, to transfer explored behaviours to the unconditioned test-time policy. ICE with the Asymmetric-RL/SD objective improves Qwen3-1.7B held-out pass@1 performance at $4$K response length on mathematical reasoning tasks by $5.0\%$ relative to training with DAPO, with improvement persisting at a longer 8K context.
In reinforcement learning, exploration with sparse and delayed rewards presents a significant challenge due to the limited feedback available for guiding the learning process. Addressing this issue requires extensive exploration in the state space to discover valuable reward signals. In this paper, we propose Entropic Information for Exploration (ENTINEX), a novel method that enhances exploration by incentivizing agents to explore beyond the boundaries of the state distribution. ENTINEX achieves this by assigning intrinsic rewards to these boundaries, leveraging entropic information to identify them effectively. Through extensive experimentation, we demonstrate that ENTINEX consistently improves exploration performance in environments characterized by sparse and delayed rewards. Our experimental results show that ENTINEX outperforms existing exploration methods, highlighting its effectiveness in both sparse and delayed reward scenarios.
Reinforcement Learning (RL) training for Large Language Models (LLMs) often suffers from instability due to the discrepancy between training and inference. This training-inference discrepancy stems from two primary factors: an architectural separation between training and inference engines, and the use of low-precision quantization in inference versus higher-precision computation in training. To address training instability issues caused by high training-inference discrepancy, we present the principles and methods for its adaptive control. We propose Adaptive Control Reinforcement Learning (ACRL), which adaptively maintains the training-inference discrepancy within a reasonable range to ensure stable RL training. Beyond stabilization, ACRL inherently increases policy entropy, thereby enhancing exploration and improving accuracy. The experimental results show that when the inference engine utilizes FP8 quantization, ACRL consistently maintains the training-inference discrepancy within a reasonable range and stabilizes RL training. Furthermore, ACRL not only matches the accuracy of the BF16 baseline but also outperforms importance sampling (IS) fixes.
We study horizon-free regret minimization for finite-horizon time-homogeneous tabular Markov decision processes with $S$ states, $A$ actions, horizon $H$, and per-trajectory total reward bounded by $1$. We propose a new algorithm and prove a regret upper bound \[\tilde O(\sqrt{SAK}+S^8A^3)\] with failure probability $δ$, where $K$ is the number of episodes and $\tilde O(\cdot)$ hides $\mathsf{poly}\log(S,A,K,1/δ)$. Thus, the regret is $H$-free and asymptotically optimal, matching the contextual-bandit lower bound $Ω(\sqrt{SAK})$ up to logarithmic factors. This completely removes the $\log H$ dependence from the previous $\tilde O(\sqrt{SAK\log H}+S^2A\log H)$ guarantee of Zhang et al. (2021), and drastically improves the prior best horizon-free regret $\tilde O(\sqrt{S^9A^3K})$ of Zhang et al. (2022) asymptotically. The main technical difficulty is that the optimal value functions $\{V_h^*\}_{h=1}^H$ are time-inhomogeneous even though the transition kernel is time-homogeneous. A direct union bound over all value functions typically incurs an additional $\min\{\log H,S\}$ factor. We avoid this factor by (i) exploiting the monotonicity of $V_h^*$ in $h$ and (ii) non-trivially projecting the value functions onto an $S$-dimensional grid. Our analysis relies on three additional ingredients. First, we introduce a horizon-truncation argument that enables reward-based exploration and removes the cost of a separate reward-free exploration phase. Second, we design a cutting bonus that preserves both optimism and the monotonicity needed for planning. Third, we prove a new bound on total deviation for time-homogeneous MDPs, which controls the clipped variance terms in the cutting bonus with adjustable polynomial dependence on $S$ and without any dependence on $H$. Together, these tools yield an asymptotically optimal horizon-free regret guarantee.
Rubric-based RL has recently shown promise in improving LLMs on open-ended tasks. A widely recognized limitation of rubric-based RL is limited exploration: criteria that no rollout manages to satisfy (Unexplored Criteria, UC) receive no optimization signal. Recent methods address this by incorporating rubric information as external guidance during rollout, yet they introduce a train-inference mismatch: the policy is optimized on rollouts produced under external guidance while this guidance is absent at inference time, causing error accumulation through autoregressive decoding. Moreover, these exploration-focused approaches overlook a fundamentally different failure mode that we term Suppressed Criteria (SC) -- criteria that are satisfied by some rollouts yet whose learning signals are lost during optimization because scalar reward aggregation assigns them non-positive aggregate advantages. Our analysis reveals that SC are remarkably prevalent: over 57% of samples exhibit this failure mode throughout training, with an average of 1.8 SC per sample. To simultaneously address both UC and SC without introducing training-inference mismatch, we propose Criterion-Distilled Policy Optimization (CriPO), which enhances rubric-based RL via on-policy self-distillation. For UC, CriPO constructs a criterion-injection self-teacher and computes a localized forward-KL loss to inject missing behaviors into the policy. For SC, CriPO employs a counterfactual self-teacher to locate criterion-relevant tokens in negative-advantage rollouts and flips their token-level advantages to positive values, preserving useful patterns that would otherwise be suppressed. Experiments on medicine and science benchmarks demonstrate that CriPO consistently outperforms rubric-based RL, achieving stronger final performance with approximately $2\times$ fewer optimization steps.
Representation learning has enabled classical exploration strategies to be extended to deep Reinforcement Learning (RL), but often makes algorithms more complex and theoretical guarantees harder to establish. We introduce Random Feature Information Gain (RFIG), grounded in Bayesian kernel methods theory, which uses random Fourier features to approximate information gain and compute exploration bonuses in non-countable spaces. We provide error bounds on information gain approximation and avoid the black-box aspects of neural network-based uncertainty estimation, for optimism-based exploration. We present practical details that make RFIG scalable to deep RL scenarios, enabling smooth integration into standard deep RL algorithms. Experimental evaluation across diverse control and navigation tasks demonstrates that RFIG achieves competitive performance with well-established deep exploration methods while offering superior theoretical interpretation.
An agent acting under partial observability must decide when to gather information and which observations are worth their cost. Standard POMDPs value information only through its eventual effect on reward. The $ρ$-POMDP framework instead rewards uncertainty reduction directly, through a belief-dependent utility $ρ$, but in practice both the choice of $ρ$ and the weight placed on it are tuned by hand for every task. We show that active inference removes this tuning entirely. Minimizing Expected Free Energy (EFE) is exactly equivalent to solving a $ρ$-POMDP whose utility is expected information gain, and the exploration weight is fixed at $w=1$ because the variational bound expresses pragmatic and epistemic value in the same units (nats). We prove this equivalence for observe-then-commit POMDPs and extend it to factored observation POMDPs, a broader class that covers interleaved observe-act problems such as non-destructive testing and mobile sensing, where gathering information leaves the hidden state unchanged. Experiments support the theory. Across environments ranging from the classic Tiger problem to RockSample and a new Structural Inspection benchmark with over 65,000 states, the untuned weight matches or outperforms reward-only planning at the same horizon, avoids the over-exploration of bonuses tuned per task, and sits near the reward-maximizing knee of the success-reward Pareto frontier. The practical payoff is an exploration objective that works out of the box. In applications such as fault detection and medical screening, where every test has a price and every missed fault has a cost, EFE supplies a belief-dependent utility that is derived rather than tuned.
Across environments with mixed sources of uncertainty, unsupervised reinforcement learning requires intrinsic motivation that does not precommit to a particular direction of surprise. Surprise minimization is scoped by design to ``unstable'' environments. Prediction-error curiosity rewards total expected surprise, including irreducible noise. Bandit or mixture switching between surprise-minimizing and surprise-maximizing rewards reintroduces non-stationarity by construction. We propose a single intrinsic reward, stationary within each window, derived from the novelty contribution of a preference-free Expected Free Energy objective, expressed in reward-maximization form. Our claim is that parameter information gain, the expected surprise of the next state minus its irreducible part, is the appropriate intrinsic signal in both high-entropy and low-entropy components of the state space. Maximizing it seeks exactly the surprise the model can explain away. In regions of unresolved dynamics, this epistemic term drives exploration. As dynamics become resolved, the epistemic term vanishes, while an aleatoric penalty favors lower-variance transitions, all without fitting an explicit next-state predictor. A pseudocount supplies epistemic value, a probe-based penalty captures aleatoric variance, and a short-horizon gate protects informative successors. A window-based freeze of all reward-defining objects yields a stationary Bellman operator, explicit bounds on learning targets, and a conditional uniform-concentration result for the nonparametric estimators under mixing, smoothness, bandwidth, and capacity assumptions. In active-inference terms, the agent is preference-free where novelty is retained, standard likelihood ambiguity vanishes under full observability, a nonstandard transition-entropy penalty is added, and surprise minimization emerges in resolved regions of the state space.
Dimitar Chakarov, Lee Cohen, Nathan Srebrocs.GT cs.LG
We extend Incentive Compatible Exploration beyond the Bayesian full-information setting of Kremer et al. [2014]. We consider agents that may possess external information unknown to the principal. We show such settings require new notions of incentivized exploration, as well as going beyond a Bayesian perspective, and we introduce a definition where agents choose any reasonable (undominated) action. Furthermore, our framework provides for a more robust treatment of ties, and extends to settings where agents lack a single common prior and instead only know that reward distributions belong to a collection of potential priors.
Reinforcement learning (RL) has become a dominant paradigm for enhancing LLMs' reasoning capabilities. However, RL algorithms with PPO-Clip are inherently limited by exploration collapse. Subsequent works remain primarily heuristic and fail to identify the essential cause of PPO-Clip's failure. This work reveals the fundamental flaw of PPO-Clip: it implicitly measures policy discrepancy using Euclidean metric, which is theoretically inconsistent with the intrinsic geometry on the policy Riemannian manifold. This geometric mismatch results in overly conservative updates in low-probability regions while aggressive in high-probability regions, ultimately collapsing exploration. To correct this geometric flaw, we propose Riemannian Isometric Policy Optimization (RIPO), which guarantees isometric policy updates on the Riemannian manifold, effectively balancing exploration and exploitation. We further show that RIPO achieves a favorable bias-variance trade-off, which stabilizes optimization. Extensive experiments demonstrate that RIPO significantly surpasses existing LLM RL algorithms across seven competition-level benchmarks (up to 60% improvement over GRPO on AIME24).
Archive-based exploration methods such as Go-Explore select which visited state to return to using visitation rarity, and frontier methods return to the boundary of the unknown; neither asks whether the unexplored region behind a boundary is enterable at all. Exploration is not just about finding reward - it is about collecting a structurally complete experience for downstream learning and planning. We introduce TopoExplore, which augments Go-Explore cell selection with a periodic topological pass: enclosed unexplored regions (voids) of the visited-set occupancy grid are detected by flood fill (the H1 classes of its cubical complex), and a decaying selection bonus is placed only on their strict entrances (gap or door cells), so sealed regions are never targeted and entered regions retire. On a controlled 18-environment MiniGrid suite (15 seeds, frozen hyperparameters) TopoExplore attains a 1.52x geometric-mean speedup in median steps-to-first-entry over its exact Go-Explore ablation, versus 1.37x for a frontier baseline; frontier exploration degrades when sealed decoy structure appears (0.83-1.48x on decoy environments vs. 1.65-2.11x for TopoExplore), while TopoExplore holds its largest win on hard multi-interaction doors (10.9x). We report an honest negative on Montezuma's Revenge - without wall knowledge, unreachable occupancy artifacts capture the bonus and performance degrades as it grows, isolating the wall-aware entrance test as the load-bearing component - and a preliminary positive on HM3D scanned buildings, where the speedup over Go-Explore tracks scene difficulty (r=0.69) even as frontier selection dominates blanket coverage. The evidence supports a deliberately scoped claim: topology-aware selection pays off where enclosed structure must be discriminated, and remains competitive at open coverage, where frontier methods are strongest, despite not being tuned for that regime.
Sunshine Jiang, John Marangola, David Zhang +6cs.LG cs.AI
Exploration is essential to RL since a policy cannot improve by repeatedly sampling the behaviors it already prefers. Standard methods inject stochasticity in the action space, but such jitter only yields rollouts close to the original. Escaping a weak policy often requires global perturbations that action noise cannot produce. Large language models (LLMs) and vision-language-action (VLA) models offer a pathway: they condition the policy on a natural language prompt, and since the rollout follows from it, modifying the prompt induces global changes. The challenge is finding prompts that induce useful global changes. With a weak policy that rarely succeeds, reward is too sparse to select on. Our idea is to refine prompts from the rollouts themselves: a vision-language model (VLM) reasons over the rollout video, diagnoses how the policy responded, and rewrites the prompt to elicit better behavior next time. This procedure resembles posterior sampling, a classical RL exploration framework, at the level of prompts: the VLM maintains an implicit distribution over useful prompts and updates it from observed rollouts. We call this strategy Prompt-Driven Exploration (PDE). Across manipulation and reasoning tasks, PDE enables RL to learn successful policies even from zero-reward starts, and improves sample efficiency more broadly. Our website is available at https://xinyunsunshine.github.io/prompt-rl.
Reinforcement learning (RL) has become the standard paradigm for enhancing the complex reasoning capabilities of large language models (LLMs). To achieve sample efficiency, modern RL frameworks rely on importance sampling (IS). However, these algorithms suffer from an exploration-stability dilemma. Pure IS often leads to catastrophic training instability, while standard clipping mechanisms used to mitigate this instability strictly constrain the policy update budget. By formalizing the concept of Probability Capacity (Cap), we reveal that conservative clipping structurally stifles exploration by prematurely truncating the update budget for correct but low-confidence reasoning paths. To break free from these constraints, we propose Unbounded Positive Asymmetric Optimization (UP), a universal and plug-and-play objective. UP theoretically restructures the optimization process by anchoring the policy to its current state via the stop-gradient operator. This asymmetric design unleashes unclipped, stable gradients for positive advantages to maximize exploration, while maintaining standard clipping safeguards for negative advantages to prevent training instability. Furthermore, our formulation readily extends across different optimization granularities, including token-level (GRPO, DAPO) and sequence-level (GSPO) frameworks. Extensive experiments demonstrate that UP enhances exploration capacity and achieves superior reasoning accuracy across diverse RL algorithms (DAPO, GSPO, and GRPO), model architectures (Dense, MoE, and vision-language), and training modalities (language and multimodal), validating UP as a truly universal plug-and-play enhancement for RL-based training.
Group Relative Policy Optimization (GRPO) is effective when the current policy already samples useful reasoning trajectories, but it stalls on hard prompts whose correct solution modes lie outside the student's on-policy support. We propose TREK (Teacher-Routed Exploration via Forward KL), a simple staged procedure that uses distillation not for imitation but for exploration support expansion. A key advantage of TREK is its generality: because it only consumes verified output trajectories, it can use an external black-box teacher, a white-box teacher, or the same model given additional inference-time context, and it can efficiently identify which hard-prompt samples are most worth consolidating even when teacher internals are unavailable. TREK first identifies prompts where the unaided student has very low pass rate, queries a proposal source to produce verified candidate solutions, keeps the top-$r$ proposals ranked by current student likelihood, applies a short forward-KL phase to pull those verified modes into the student's support, and then returns to standard on-policy GRPO refinement. On mathematical reasoning, TREK with DeepSeek-V4 proposals improves Qwen3 models across all tested scales on AIME 2024 and AIME 2025; for Qwen3-8B, it improves AIME 2025 from 36.9 to 40.3 and AIME 2024 from 47.9 to 51.1 (avg@16), while the self-context variant reaches 38.5 and 49.6 without an external teacher. On agentic tasks, TREK raises ALFWorld success rate from 75.8 to 82.8 and ScienceWorld success rate from 12.5 to 26.7; notably, on the hardest task types, TREK achieves high success rates early in training while unaided GRPO requires substantially more optimization steps to reach comparable levels.
Aligning large language models with diverse and heterogeneous human values requires multi-objective alignment methods to effectively trade off conflicting preference dimensions. Current methods achieve this trade-off by training policies conditioned on preference vectors and leveraging online direct preference optimization. However, exploration uncertainty can cause the reward distributions of responses generated under different preference vectors to overlap, and the generated responses may fail to effectively align with the corresponding preference vectors. In this paper, we propose Multi-Objective Exploration and Preference Optimization via Mutual Information (MI-EPO), an information-theoretic framework. It unifies multi-objective exploration and alignment by maximizing the joint conditional mutual information among generated responses, preference feedback, and preference vectors. By incorporating a probabilistic routing mechanism, MI-EPO naturally decomposes objective alignment and preference-aware exploration, encouraging the model to generate responses that are distinguishable and aligned with different preference conditions. Experiments on safe alignment and helpful assistant tasks show that MI-EPO significantly improves the alignment between generated responses and preference vectors, makes the outputs more controllable, and achieves stable trade-offs across multiple objectives.
Louis Bagot, Mathieu Lefort, Laëtitia Matignoncs.AI cs.LG
Zero-shot Transfer in Reinforcement Learning (RL) aims to train an agent that can generate optimal policies for any reward function, without additional learning at transfer time, while training only on reward-free trajectories. For their generality over tasks, such models are sometimes called ``Behavioral Foundation Models'' (BFMs). While they have shown strong performances and improvements in recent years, the current framework and algorithms still assume that, during the transfer phase, the agent is informed offline about the reward (the task to solve) through a dataset of state-reward pairs, which it uses to pick the best policy to deploy. However, in practice if the reward is a black-box (e.g. direct user feedback), it is not possible to generate such a dataset: it is necessary to observe the reward through interactions with the environment. In other words, the current framework of offline transfer is not aligned with the traditional RL setting of online learning through trial-and-error, which requires exploration in order to find rewards. This paper proposes to tackle this new online transfer in zero-shot RL, with the key insight that the BFM itself can be used to generate exploration policies. We show that it is possible to frame this online learning problem in terms of a bandit-like exploration-exploitation problem. More precisely, at each step the bandit algorithm recommends a policy, the BFM executes it in the environment, which yields a reward and a new state; we repeat the process until we converge to the optimal policy. In the popular context of linear reward approximation, we derive a formulation inspired by Upper Confidence Bound and show that exploration can be achieved through the minimization of the eigenvalues of an uncertainty matrix. We evaluate qualitatively and quantitatively our framework on a simple environment to validate the concept of our method.
In stochastic linear bandits, the canonical Upper Confidence Bound (UCB) algorithm admits a simple frequentist regret analysis but can be computationally demanding, while Thompson Sampling (TS) is computationally attractive yet typically harder to analyze due to its non-optimistic nature. We propose Absolute Thompson Sampling (ATS), a simple modification of TS that ensures optimism in expectation by replacing the signed exploration noise with its absolute value. This preserves the computational efficiency of TS while avoiding the technically involved anti-concentration arguments common in TS analyses, enabling a simple UCB-style regret analysis. We show that ATS achieves $\tilde{O}(d^{3/2}\sqrt{K})$ regret, matching existing bounds for TS in linear bandits. We further introduce Ensemble Absolute Thompson Sampling (EATS), which takes the maximum over multiple absolute perturbations with normalization by the ensemble size. As the ensemble size grows, EATS converges to the UCB objective, recovering UCB behavior in the limit. Experiments show that moderate ensemble sizes already yield strong performance. Our results point to a bridge between randomized exploration and deterministic optimism both in theory and practice.
Reinforcement Learning with Verifiable Rewards (RLVR) for language-model reasoning can fail at both extremes of task difficulty: easy prompts often produce all-correct, low-diversity rollout groups with little gradient signal, while hard prompts can produce all-incorrect groups with no positive reward. We introduce ExTra (Exploratory Trajectory Optimization), a GRPO-compatible framework that extracts exploration signals from the model's own rollouts. ExTra combines two mechanisms: (i) a novelty reward that adds embedding-based diversity bonuses after GRPO normalization, rewarding diverse correct solutions; and (ii) entropy-guided prefix regeneration, which scores partial trajectories using entropy signals and continues exploration from promising intermediate steps. Across six mathematical reasoning benchmarks, ExTra improves Qwen3-1.7B over GRPO by about +5 points on pass@1 and +7 points on pass@16, showing that trajectory-level exploration signals can improve both single-sample accuracy and inference-time coverage.
Reinforcement learning (RL) has become central to LLM post-training, yet the methods that dominate current pipelines, PPO and GRPO, represent only a narrow slice of what RL offers. Understanding why these methods prevail, and what alternatives exist, requires a principled examination of the design decisions that underlie any RL algorithm. This survey organizes that examination around three stages of algorithm construction. We begin with MDP creation: how the reward function, state space, action space, termination condition, and discount factor are, or could be, defined for LLM training. We then turn to exploration, covering temperature sampling, entropy regularization, intrinsic motivation, tree search, and curriculum learning. Finally, we address learning along four classical RL dimensions: model-free versus model-based, value-based versus policy-based versus actor-critic, on-policy versus off-policy, and credit assignment, including both Monte Carlo methods, which rely on full return estimates, and bootstrapping methods, which update estimates using other learned predictions. Mapping the LLM literature onto this taxonomy reveals a strikingly non-uniform distribution of research effort. Critic-free policy gradients and Monte Carlo credit assignment are densely populated, while value-based methods, off-policy actor-critic training, and bootstrapping-based credit assignment remain largely unexplored despite well-established counterparts in classical RL. These gaps represent concrete opportunities for transferring proven RL techniques to LLM training. By making these gaps explicit alongside the methods that have proven effective, this survey offers researchers in both RL and LLMs a shared framework for understanding current practice and identifying promising directions for future work.
Marco Pratticò, Pietro Novelli, Massimiliano Pontil +1cs.LG
Sparse rewards pose a central challenge in reinforcement learning, since agents receive no informative signal until they reach their goal. Intrinsic-reward methods address this issue by optimizing non-stationary objectives such as novelty, prediction error, or skill diversity, thereby injecting a supervision signal into the problem. While effective, these methods often require that the extrinsic (sparse) reward can be evaluated -- either online or during offline relabeling of the stored transitions. This limitation is particularly vexing for multi-task, meta-, and continual reinforcement learning, where agents' interactions with the environment are usually reward-free. In this work, we present a method to pre-train transferable exploration policies that rapidly adapt to sparse rewards at downstream task time. Our objective maximizes state-space covering for the occupancy measure, and can be framed in terms of entropy maximization. Its algorithmic implementation, ROVER, leverages recent advances on the operatorial formulation of RL to estimate occupancy with a learned resolvent world model, bypassing common hurdles associated with density and entropy estimation. ROVER further introduces a virtual "sink" state for unexplored regions, balancing coverage of known states with expansion into unseen ones and preventing cyclic expansion-collapse behavior during learning. In tabular and pixel-based sparse navigation tasks, ROVER produces more uniform aggregate coverage and stronger initializations for downstream tasks than standard reward-free baselines.