Policy Learning

Modern data-driven decision-making methods, such as imitation learning (IL) and reinforcement learning (RL), have achieved great success in solving many complex tasks. However, these methods often suffer from serious control instability and robustness issues when applied in real-world applications such as robotics and autonomous driving, posing notable challenges for their practical deployment. We argue that this instability issue stems largely from their limitations in solely supervising and optimizing zeroth-order actions (i.e., the action labels), failing to account for higher-order action dynamics and temporal consistency. In this paper, we show that simultaneously supervising both zeroth- and first-order actions can dramatically enhance policies' performance and control robustness. To achieve this, we introduce a novel and elegant loss scheme supported by formal theoretical guarantees that can equip any off-the-shelf policy model (e.g., deterministic, stochastic, or flow policies) with the capability for higher-order action supervision, without requiring any structural modifications. Moreover, our proposed method can serve as a lightweight plug-and-play module that seamlessly integrates with a broad spectrum of existing offline RL frameworks. Extensive evaluations on OGBench and D4RL demonstrate that our approach yields substantial performance and robustness improvements across a wide range of continuous control environments. Notably, our method can also enhance policies' out-of-distribution (OOD) generalization capability in the challenging low-data regime, making it an ideal tool in tackling many real-world control problems.

Heuristic search for a plan can store exponentially many states, even when its heuristic is almost perfect. We instead learn search control, one specification per domain, written as an indexical policy: a generalized policy with registers that hold objects and modes that sequence its rules. We add the choose rule, which loads an object into a register and marks a backtracking point, where one candidate suffices; every other rule must work for all of its outcomes and needs no search. Our main result is that structural termination, which rules out infinite executions, also bounds every execution by a polynomial in the number of objects. A depth-first procedure then finds a plan in polynomial space, however large the state space, with no list of visited states. The cost is time, exponential only in the choice depth, the number of real choices along an execution. Any class that such a policy solves therefore lies in NP, and in P at constant choice depth. We learn these policies with a language model in a counterexample-guided loop that certifies termination, verifies the training tasks, and keeps the choice depth small. With the learned policies, the procedure solves 1,709 of 1,890 test tasks of the IPC 2023 Learning Track and the Autoscale Agile suite, more than LAMA, BFWS, and Levitron, and most of them within one second and 100 MiB.

Policy learning in digital experimentation faces three challenges: weak signal-to-noise ratios, rich covariate spaces, and massive data volumes. We formalize this regime by modeling treatment-effect estimates from increasingly fine covariate partitions as Gaussian observations with bounded signal-to-noise ratios. We establish that, in general, the optimal treatment policy is not learnable in this setting. Even learning the optimal policy value suffers from impractically slow rates. However, when treatment effects vary smoothly, we derive minimax-adaptive policies based on linear smoothers that achieve vanishing welfare regret. We demonstrate the practical value of our framework by applying it to large-scale real-world experiments at Netflix, showing that personalized linear-smoothing policies can dominate unpersonalized policies even in this challenging empirical setting.

Long-running autonomous agents must reuse accumulated reasoning experience without allowing explicit historical memory and LLM context to grow indefinitely. However, existing memory mechanisms mainly retrieve, summarize, or compress past content and do not directly learn when particular kinds of thinking should be activated or discover new thinking knowledge from temporally dispersed experiences. This paper proposes a situation-conditioned thinking memory framework that transforms historical reasoning experience into a lightweight policy for predicting what should be thought about in the current situation, while leaving detailed reasoning to a large language model. Situations may represent temporal or spatiotemporal evolution rather than only current states. Temporary experiences are also periodically analyzed across multiple independent episodes to identify repeated long-range regularities, which are consolidated into new thinking knowledge and further internalized by the lightweight policy. Experiments show that the learned policy achieves 1.000 F1 on temporal-rule generalization, improves DeepSeek reasoning F1 from 0.789 to 0.868, reduces online processing time from 0.3636 ms to 0.0382 ms per query at 30,000 historical situations, and reaches 1.000 relation-discovery F1 and future-thinking accuracy after sufficient repeated cross-experience evidence.

State-of-the-art reinforcement learning agents are capable of outperforming human experts at games like chess, Go and StarCraft II. These agents do not simply take advantage of their digital hardware in being able to react and calculate faster than humans, but employ better strategies that lead to more victories. Interpreting these strategies would give human players valuable insight into how to improve their play. In this preliminary work, we propose a symbolic sub-policy model for playing chess. Inspired by chess tactics, our model attempts to incorporate domain knowledge to improve interpretability. We adapt patterns learned by an inductive logic programming system called PAL to derive our model. We contribute a divergence metric to evaluate our model against a random baseline, and find a set of tactics that is able to suggest moves of similar playing strength to a human beginner. Finally, we propose a computational evaluation scheme for the model by augmenting an off-the-shelf engine with it.

As part of learning to play complex games, human players develop develop abstractions for concepts and strategies of gameplay consistent with game rules to improve their performance. These concepts are applied to explain other players' actions, and to inform their own actions in-game. Understanding other players' strategies is a crucial part of such improvement, but requires time and effort. In this paper, we propose a strategy similar to human cognition for training RL agents to synthesize learned strategies and policies as executable procedures based on sequences of gameplay actions. We present methods to automatically learn such programs to play chess and to solve tasks in a grid-based environment. We show that the learned strategies produce effective actions, and can be learned from gameplay data.

Several state-of-the-art methods for online reinforcement learning in continuous control improve policies using action gradients of a learned critic. However, critics are typically trained to predict returns, and accurate value predictions do not necessarily yield accurate action derivatives, potentially leading to unreliable policy updates. We propose Forward Entropy-Regularized Policy Optimization (FERPO), an on-policy maximum entropy reinforcement learning algorithm that performs policy improvement using critic values without differentiating the critic with respect to actions. FERPO derives an optimal target action distribution from a policy-improvement objective regularized by entropy and Kullback-Leibler (KL) divergence. We then fit the actor to this target by minimizing a forward-KL objective, estimated using self-normalized importance sampling (SNIS) with actions drawn from the rollout policy. By limiting the target distribution's deviation from the rollout policy, the KL regularization helps keep these importance weights well behaved. In contrast to reverse-KL objectives, which can favor a subset of the target distribution's modes, the forward-KL objective encourages coverage of multiple high-value modes and thereby promotes exploration. Experiments and ablations on MuJoCo Playground and ManiSkill show competitive performance and sample-efficiency gains. Computational benchmarks also demonstrate faster actor updates than Relative Entropy Pathwise Policy Optimization (REPPO).

Reinforcement learning (RL) for reachability specifications is fundamental to sequential decision-making. Prior work establishes asymptotic convergence to optimal policies, but only through model-based methods that must explicitly estimate the transition probabilities of the underlying Markov Decision Process (MDP). We present Quasar, the first model-free algorithm with asymptotic guarantees for reachability on the fragment of MDPs free of non-terminal maximal end components (MECs), a building block to which every MDP reduces by the standard MEC quotient. Our algorithm follows the classical Q-learning approach, using temporal-difference updates to converge to an optimal policy without ever learning the transition probabilities. The resulting learner reduces the memory footprint from the O(|S|^2|A|) that model-based methods require to O(|S||A|). On the standardized Quantitative Verification Benchmark Set, our algorithm converges to the optimal policy with orders of magnitude fewer samples than the previous model-based state-of-the-art. Together these results are a concrete step toward the practical deployment of reachability learning and, with it, of specification-guided RL.

Practical Reinforcement Learning (RL) algorithms learn to solve Markov Decision Processes (MDPs) through iterative policy improvement in the presence of approximate evaluation. We study policy improvement from first principles, defining optimal policy improvement as producing the best policy attainable in a single update under specified constraints. We show that optimal improvement restricted to a set of states is equivalent to solving an induced MDP, characterizing planning with an explicit or implicit model as a path towards optimal policy improvement. Because practical methods commonly solve such induced problems through iterative improvement in the form of greedification, we take steps towards optimal greedification under the central practical constraint of approximate evaluation. We formulate greedification under this constraint as probabilistic decision-making under uncertainty and derive a novel operator that is optimal with respect to the resulting objective. Empirically, the operator and its practical gradient-based approximations improve aggregate performance across GumbelAlphaZero, SAC, ReBRAC and Generalized Policy Iteration, in experiments spanning discrete and continuous actions, model-based and model-free, online and offline RL.

Among on-policy deep reinforcement learning methods, Proximal Policy Optimization (PPO) has become the de facto standard, due to its consistently strong empirical performance across diverse application domains. However, on-policy methods are inherently sample inefficient: fresh data collected under the current policy is used for just a few updates before being discarded. Off-policy methods avoid this inefficiency via experience replay, achieving notable sample efficiency gains, but at the cost of training instabilities or extensive tuning. This motivated the rise of hybrid strategies that augment PPO with off-policy data reuse. Existing sample-reuse variants of PPO demonstrated improved sample efficiency over vanilla PPO, yet a systematic study of when reuse helps, in which scenarios, and to what extent remains missing. In this work, we study the effectiveness of sample reuse in PPO by instantiating two variants within a multiple importance weighting framework. Both retain the core PPO mechanics, reusing only samples from a window of recent iterations, thereby isolating the effect of data reuse from other factors. The variants, termed wPPO-U and wPPO-BH, employ vanilla importance weights or balance-heuristic-corrected ones, respectively. For both, we derive policy improvement lower bounds providing theoretical grounding for their respective losses. We use them to empirically study when and how data reuse improves sample efficiency or final performance of PPO across continuous control tasks.

Existing approaches to offline goal-conditioned reinforcement learning (GCRL) struggle with long-horizon tasks. Discounting shrinks value differences between distant states until they fall below the function approximation error, leaving the agent with no signal for ranking states. Temporal abstraction, which treats k environment steps as a single transition, restores this signal at long range, but no single fixed k suits all state-goal distances: large k preserves value differences across long temporal distances while collapsing distinctions between nearby states, and small k does the reverse. We make this trade-off explicit and introduce Generalized Implicit Temporal Abstraction (GITA), which conditions a single value function on k. GITA trains one policy by aggregating advantage-weighted supervision across multiple k values, so scales assigning larger positive advantages to a state-goal pair contribute more strongly to its update. GITA does not need to choose between local resolution and long-range signal; it retains both without committing to a single k. On OGBench, GITA outperforms a broad range of offline GCRL baselines, raising average success rate across all tasks by 25 percentage points (73% relative improvement) over HIQL. It also improves over the strongest fixed-k method, OTA, by 7 percentage points (14% relative).

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.

Multi-reward reinforcement learning trains large language models to satisfy multiple behavioral objectives simultaneously. Reward-wise normalization, as used in GDPO, preserves reward-specific relative information within rollout groups, but different objectives can still exhibit uneven learning progress. We study this behavior through advantage energy, the sum of a reward's squared advantages over a batch. Under idealized GDPO normalization, we show that this energy is proportional to active-group density: the fraction of rollout groups in which the reward provides nonzero relative advantages. This reveals a residual batch-level signal imbalance and provides a basis for calibrating reward contributions. Based on this relation, we propose Density-Aware Reward Aggregation (DARA). We derive an inverse-square-root density correction that gives greater weight to signals from less frequently active rewards. DARA computes its weights from each rollout batch, adapting to changes in reward activity throughout training without modifying the underlying policy optimization objective. Experiments on tool calling and mathematical reasoning show that DARA learns the targeted behaviors faster than GDPO, reaching high format compliance in up to 26% fewer training steps on tool calling and near-saturated length compliance in up to 65% fewer steps on mathematical reasoning, while remaining competitive in final performance. Our code is available at https://github.com/zhaihaotian/DARA.

On-policy distillation (OPD) is a promising approach for training language agents, providing dense teacher supervision on student-generated trajectories. However, in multi-turn interaction, an incorrect action changes the states the student encounters later, so errors compound across turns. In preliminary experiments across three Qwen3 models (8B to 235B), we find that more than half of the failed rollouts contain a pivotal mistake, an action that moves the agent farther from completing the task, and this mistake typically occurs early. These pivotal mistakes often remain recoverable: guiding the model for only a few turns after the pivotal turn can restore task success. We therefore propose PivotOPD, an on-policy distillation framework that jointly trains the student to prevent pivotal mistakes and to recover from the states they create. At each pivotal mistake, a teacher model provides a gold action and then names a recovery action at each of the next few turns. Preventive distillation uses the gold action with reverse KL to steer the student away from the pivotal mistake, while recovery distillation uses the recovery actions with forward KL to transfer recovery behaviors that the student rarely samples. Against 13 baselines on ALFWorld, WebShop, and Search-based QA, PivotOPD achieves the strongest average performance for both Qwen3-1.7B and Qwen3-8B students, improving over the strongest baseline on ALFWorld by +5.5% with the 1.7B student. The gains also transfer to another model family on the software engineering domain, where PivotOPD raises the resolve rate of a Nemotron-3.5 student on SWE-Bench Verified by +3.2%. Project page: https://research.nvidia.com/labs/lpr/pivotopd/

Policy regularization in offline reinforcement learning balances policy improvement against reliance on uncertain value estimates. This balance can differ between selecting actions for execution and supplying actions for critic bootstrapping, yet methods such as TD3+BC couple these roles through a shared policy. We propose Role-Adaptive Policy Optimization (RAPO), which adapts policy-update coefficients according to their roles in value learning and execution. RAPO learns these coefficients by differentiating through candidate policy updates formed using the base algorithm's actor objective. For TD3+BC, RAPO separates bootstrap and execution actors and adapts their coefficients independently: the bootstrap objective penalizes policy-induced changes in target values, while the execution objective evaluates a local policy-improvement surrogate. For IQL, whose value learning is already independent of the execution actor, RAPO preserves the original value updates and adapts only the inverse temperature in advantage-weighted policy extraction. Experiments on D4RL locomotion and AntMaze tasks show improvements over both base algorithms, with larger gains for TD3+BC, whose RAPO instantiation outperforms baselines on average.

The policy-gradient theorem gives the exact gradient under the current policy, but finite on-policy samples may miss rare high-return trajectories. We study whether tree search improves their coverage within a fixed budget while controlling gradient bias. We introduce On-Policy Parallel Tree Search (OPTS) and Tree Trajectory Policy Optimization (TTPO) using on-policy tree trajectories, which sample new suffixes from the current policy at visited states. This needs no action-distribution correction, although branching changes state visitation. Our Branch Aggregation Lemma shows that branch-weighted tree statistics recover chain expectations when branch choices and weights are fixed before outgoing transitions are sampled. OPTS selects expansion states using estimated performance differences. Under deterministic dynamics, exact values, and max-backup advantages, the induced search policy's expected return improves monotonically with the budget. We bound the gradient bias from adaptive expansion and show that max backup assigns prefix credit to actions leading to better discovered suffixes. Against a finite chain reference, TTPG's measured bias stays near its no-branching level, while NaivePG's bias grows from 0.1251 to 0.4884. At matched budgets, reward- and value-guided OPTS improve correct-answer coverage and majority-vote accuracy over independent sampling. At matched branch counts, OPTS + TTPG gains coverage with a modest bias increase relative to Fixed-branch + TTPG. Under matched interaction or rollout budgets, OPTS-TTPO improves MuJoCo tail returns over PPO by up to 28.6%, achieves a 34-22-1 win-loss-tie record against PPO on Atari-57 under the last-100-log mean-return metric, and improves micro-averaged avg@32 and pass@32 over PPO across all four Qwen3 models.

Policy mirror descent (PMD) enjoys fast convergence in regularized Markov decision processes (MDPs), but existing guarantees often rely on exact or increasingly accurate policy evaluation. We analyze PMD coupled with a persistent critic advanced by one temporal-difference (TD) update. For finite discounted MDPs, we establish global linear convergence in value for exact coordinate-wise Bellman updates, with any positive constant actor stepsize and arbitrary finite critic initialization. The proof combines a resolvent-based auxiliary distribution with a decaying Bellman-violation correction and a potential weighted by inverse coordinate weights. We then study stochastic TD-PMD with general strongly convex mirror maps under a single off-policy Markov trajectory. With suitably chosen constant stepsizes and a finite-batch TD update, the method achieves an expected value gap of $ε$ after $\widetilde{O}(1/((1-γ)^5 \widetildeσ_b ε))$ transitions. The stochastic analysis relies on the trajectory-wise Lipschitz continuity of the regularizer, derived from uniform bounds on vertex Bregman divergences, together with a visitation-weighted resolvent estimate for signed critic-error propagation that yields an inverse-linear dependence on behavior coverage $\widetildeσ_b$. In contrast to many prior guarantees for regularized policy optimization, our sample-complexity guarantee holds without trajectory resets, generative-model access, or nested policy-evaluation loops. Numerical results are consistent with the theoretical convergence analysis.

Planning with learned world models combines online trajectory optimization with learned value and policy functions for high-dimensional control. Because the planner determines the experience used for learning, while the learned critic and actor in turn score and propose future plans, planning and learning form a closed feedback loop. TD-MPC is a prominent instance of this design. Recent policy-constrained variants strengthen one part of the loop by aligning the learned policy with planner behavior. We introduce PL-MPC (Planning-Learning MPC), which additionally modifies critic supervision and planner terminal-value estimation. Hybrid multi-step TD targets expose critic updates to more realized rewards before bootstrapping; disagreement-aware terminal estimates reduce the influence of uncertain critic values during MPPI planning; and return-weighted actor distillation emphasizes planner-executed actions from high-return episodes. The world-model architecture and MPPI optimizer are otherwise unchanged. On HumanoidBench, the largest gains occur on \texttt{balance-hard}, where Total Average Return (TAR) increases from $98\pm18$ to $387\pm255$, and \texttt{hurdle}, from $199\pm13$ to $466\pm200$; performance across the broader benchmark remains task dependent, and PL-MPC remains competitive on DMControl. Controlled ablations show different component interactions across the two tasks. We further demonstrate zero-shot sim-to-real transfer on wrench-nut alignment with a 7-DoF KUKA IIWA14, obtaining higher observed success than TD-M(PC)^2 on the training object size and two unseen sizes. Code and data will be available at: https://pl-mpc-humanoid.github.io.

Common policy improvement methods, including TRPO, PPO, and GRPO, estimate policy improvement under the behavioral policy's state-visitation distribution rather than the improved policy's own. The substitution makes the objective estimable from the behavioral policy's rollouts but adds a bias growing with policy divergence, hence the trust region or clip, and hence no reuse of a batch far off-policy. We show that under history-injective dynamics, where each state is reached by exactly one history, the dropped state-visitation ratio equals the product of per-step policy ratios along the sampled prefix, on every trajectory and not only in expectation. The ratio is therefore restored exactly, from log-probabilities PPO already computes. Autoregressive generation and canonical-order constructive optimization are both history-injective. The exact correction pays importance-sampling variance that grows with the horizon, so we generalize it to a one-parameter family with PPO ($α{=}0$) and the full correction ($α{=}1$) as endpoints: a single bias--variance knob. A gradient-level analysis of the unclipped surrogate identifies two channels the correction acts through and three conditions under which it carries signal; an enumerable testbed confirms the conditions' predictions. On hard credit-assignment scheduling tasks, a short corrected warmup with aggressive early sample reuse learns faster than PPO and than the same reuse uncorrected; the marginal gain grows with task difficulty ($+0.02$ to $+0.09$ learning-curve AUC), and the early win over PPO tracks the prefix bias that reuse incurs. A correction held throughout, or applied where clipping already contains the reuse bias, is null to harmful.

Reinforcement learning enables large language model (LLM) agents to learn multi-step behaviors through interaction with their environments. However, rewards in many interactive tasks reflect only the final outcome, providing limited guidance on which intermediate decisions advance the task. Successful training trajectories contain intermediate states that can provide supervision for subsequent interactions. We introduce Trajectory-to-Step Policy Optimization (T2SPO), a method that uses past interaction trajectories to provide step-level feedback for policy learning. T2SPO derives remaining-distance targets from successful trajectories and pairs them with representations of the states visited along the way. Conditioned on these examples, a pretrained TabPFN regressor estimates the remaining distance to success at each state of a new rollout. Changes in this distance estimate across consecutive states yield auxiliary credit for agent steps alongside task-level supervision. As training proceeds, newly completed trajectories refresh the estimator's context, incorporating new experience without updating its parameters. Experiments with 1.5B and 7B language models on ALFWorld and WebShop show that T2SPO consistently improves overall task success over GRPO.

Reinforcement learning with a verifiable reward (RLVR) offers a scalable approach to training language-model agents, yet sparse outcome rewards can leave early training with little signal for policy improvement. We identify an On-Policy Acceleration Phenomenon: in our main comparisons, RLVR initialized with on-policy distillation reaches high performance earlier in training and achieves both higher average performance during subsequent RLVR and higher final performance than the alternative baselines. Motivated by this observation, we study On-Policy Warmup (OPW), a teacher-guided stage in which the student trains with teacher supervision on its own interaction trajectories before transitioning to RLVR. Unlike imitation on fixed teacher-generated trajectories, OPW targets states induced by the student's own decisions, including imperfect actions and recovery situations. We provide a theoretical explanation by connecting on-policy reverse-KL distillation to trajectory-level distribution matching. Under a competent teacher and sufficiently small population distillation loss, this connection yields a lower bound on initial verifier success and a corresponding bound on reward-discovery complexity. For group-relative RLVR, we further characterize when increased success probability produces more reward-informative groups. Together, our findings support on-policy distillation as an effective warmup for agentic RLVR and identify initial reward discovery as a mechanism that can contribute to the observed acceleration.

Does memory-dependent control need nonlinear recurrent dynamics? We study simulated air-hockey defence under temporary loss of puck tracking. A DreamerV3 teacher outperforms a memoryless policy under tracking loss, while resetting the teacher's recurrent state sharply reduces performance, which demonstrates that the task requires memory. We distil this teacher into compact recurrent policies with a 64 dimensional state, with a combination of a diagonal linear recurrence and an optional rank-$k$ nonlinear innovation while retaining nonlinear observation encoders and action heads. Across five matched seeds, the purely linear recurrent model ($k=0$) matches both the GRU baseline and the teacher throughout the tested range of tracking loss. Increasing nonlinear innovation rank providing no measured benefits. This result is obtained on a fresh test split, which will be only opened after all models and analyses are frozen. The linear model requires fewer recurrent parameters and less computation than GRU, but performs comparably. These results suggest that, for this memory dependent control task, nonlinear representation learning around a simple linear memory mechanism can be sufficient, and that nonlinear recurrent dynamics are not necessarily required. These conclusions are limited to the simulated task, teacher, state dimension, and blackout horizon considered here, and to policies whose observation encoder and action head remain nonlinear.

Actor-critic methods achieve strong performance in continuous control, but their policies can produce highly oscillatory actions. A common remedy is to add auxiliary smoothness losses. However, their contribution can be negligible when their gradients are small relative to the native actor gradient. Moreover, existing methods often combine multiple auxiliary losses, complicating loss balancing without necessarily improving the return-smoothness trade-off. We introduce DAMPER (Direction-Aware Magnitude-Controlled Projection with Explicit Return Priority), which combines the native actor gradient with a temporal-consistency gradient through conflict-conditioned projection and adaptive magnitude control. It removes the auxiliary component opposing the actor gradient and scales the retained temporal direction relative to the actor gradient norm, preserving positive alignment with the native actor gradient. Experiments with TD3 and SAC on six continuous-control tasks show reduced action oscillation relative to the native agents in all 12 task-backbone pairs and the best oscillation score among the compared methods in eight, with task-dependent return trade-offs.

On-policy distillation (OPD) learns from teacher feedback on student-generated responses and has shown promise in reducing forgetting relative to supervised fine-tuning (SFT). However, its benefits and fragility remain incompletely understood. We study sequential distillation from multiple teachers, where the student minimizes its average divergence from the teachers. Forward Kullback--Leibler (KL) divergence yields a weighted arithmetic mixture, while reverse KL yields a normalized weighted geometric aggregate. We develop algorithms that learn these targets under off-policy and on-policy feedback, respectively, establishing logarithmic regret bounds in the tabular setting and extending the analysis to function approximation. By analyzing these aggregation targets, we identify mechanisms that help explain both the benefits and fragility of OPD. Relative to forward KL, reverse KL can better retain a confident expert's preferences under uninformative feedback, but is more sensitive to teachers that assign very low probabilities to correct responses. Its token-level conditionals also reveal a dependence on continuation distributions that can favor incorrect prefixes over long horizons.

Persistent textual memory allows language models to carry information across long interactions, but learning what to remember is fundamentally a credit-assignment problem. A memory rewrite may only become useful many steps later, while much of the observed utility may be inherited from information already stored before the rewrite. We introduce Memory Gain Policy Optimization (MGPO), which isolates the incremental value of each memory rewrite by crediting it for its marginal contribution to current and future downstream utility. This turns delayed memory utility into a direct learning signal for optimizing what information should persist. We study MGPO on document-level information extraction, where structured supervision makes the effects of individual memory updates directly measurable. MGPO improves extraction while reducing average memory length by nearly 80% relative to the initial memory policy before optimization. The learned memory policy also supports reuse and transfer across domains, downstream models without further training. These results show that effective memory learning depends not only on preserving useful information, but on identifying which memory updates create lasting incremental value.

Reinforcement learning for long-horizon agents typically relies on sparse outcome-based rewards. This leads to a severe cold-start problem, as early-stage policies often fail to solve sampled tasks, leaving little useful reward signal for learning. To mitigate this problem, we use on-policy distillation (OPD) to provide token-level guidance on the student's own rollouts. We find that the benefit of this guidance depends on the performance gap between the teacher and the student. When the teacher substantially outperforms the student, distillation helps guide the student through the early training stage where outcome rewards provide little learning signal. As the gap narrows and eventually reverses, however, continued distillation becomes less beneficial and may hinder further improvement. Motivated by this observation, we propose Gap-Adaptive Teacher Scheduling (GATS), which augments the student's RL objective with an OPD term whose weight adapts to the teacher-student performance gap. Specifically, GATS gradually reduces teacher guidance as the student approaches the teacher's reference performance and withdraws it once that reference is reached. This enables GATS to leverage task-trained teachers smaller than the student, since teacher guidance is primarily needed during early training. Across ALFWorld, WebShop, and ScienceWorld with three Qwen2.5 teacher-student configurations, GATS achieves the highest average success rate among the compared methods in all three configurations, improving over reward-only GRPO by 4.37%-11.87% under matched student rollout budgets. Code is available at https://github.com/Ricardo-H/guide-then-let-go.

Video-based policy learning is particularly promising, as it illustrates target behaviors without requiring action annotations or embodiment-matched demonstrations. A central challenge is deciding what information should be transferred from the video to the robot. Existing approaches commonly convert visual observations into scalar similarity or value signals, or ask foundation models to directly generate reward code. These approaches can make the temporal structure of a task difficult to inspect, ground, and reuse. We present Video2STL, a framework that converts observation-only videos into parametric Signal Temporal Logic (STL) specifications and uses the resulting formal representation for robot learning. A vision-language model extracts an embodiment-independent semantic event trace and constructs a bank of symbolic temporal specifications. The model determines the task structure, while numerical predicate thresholds and temporal bounds are grounded from successful robot trajectories. For policy learning, we separate short- and long-timescale temporal information: short-horizon specifications provide dense rewards through rolling-window quantitative robustness, while a causal monitor over a retained long-horizon specification provides one-time progress rewards for valid temporal prefixes. The same representation supports cross-embodiment transfer from human or animal videos to robot control. Across four manipulation tasks, Video2STL achieves $85.8\%$ average success-once and $67.0\%$ success-at-end, compared with $81.5\%/59.5\%$ for native dense PPO and $65.0\%/42.3\%$ for Text2Reward; in quadruped locomotion, Qwen-3.8 and GPT-5.6-based Video2STL policies achieve $100\%$ success across velocities from $0.3$ to $2.1\,\mathrm{m/s}$ while remaining competitive in high-speed energy efficiency. Project webpage: video2stl.

Recent approaches to reinforcement learning (RL) post-training for large language models increasingly remove the critic to reduce training instability and memory overhead. Even where a critic is trained, it is discarded once training ends, although it has learned to predict outcomes. We revisit this trend and show that a pretrained critic's ability to predict future outcomes can make it a valuable asset for efficient long-horizon reasoning. First, we find that instability in critic-based RL for long chain-of-thought reasoning is largely an optimization artifact: keeping policy updates small and low in variance restores stable convergence. Second, a well-pretrained critic estimates the posterior probability of eventual success from later trajectory states and unfinished prefixes. Its predictions provide outcome-derived, dense, per-prefix learning signals that, during policy optimization, require neither completed rollouts, step-level annotations, nor external reward labels. Building on this insight, we introduce Reward-Free Policy Optimization (RFPO), which repurposes a single calibrated, frozen critic as a rollout-level reward, a value baseline for generalized advantage estimation, and a success forecaster for unfinished prefixes. We further show that binarizing the debiased score stops the policy from exploiting the critic's length bias. Binarized, RFPO matches supervised PPO without a single label in the training loop, while cutting compute and memory overhead. This makes RFPO well suited to long-horizon reasoning tasks, where outcomes arrive late and generation dominates cost: because rollouts can be rewarded before they finish, training no longer has to pay for waiting on every trajectory to complete. Our findings challenge the prevailing critic-free paradigm and establish critic-based, reward-free optimization as a scalable and computationally efficient path for LLM post-training.

Long-horizon LLM agents require reinforcement learning methods that can assign credit to intermediate decisions under sparse and delayed rewards. Existing group-based methods such as GRPO and GiGPO alleviate this issue by comparing rollout returns or repeated anchor states, but they still fail when the compared returns have no variation. We identify this failure mode as zero-credit failure: during early training, many failed rollouts contain useful prefixes, yet existing methods assign them no task-discriminative advantage. To address this issue, we propose Milestone Viability Potential Policy Optimization (MVPO), a potential-routed policy optimization algorithm that learns from viable failure prefixes. MVPO estimates prefix potential over Union-Find viability regions, repairs zero-credit groups with potential-difference advantages, and attenuates the potential branch according to relative performance progress. Experiments with Qwen2.5-1.5B-Instruct show that MVPO outperforms eight strong baselines, including GRPO and GiGPO. Under the same training length, MVPO improves over the GiGPO baseline by +4.4 success points on ALFWorld and +5.3 on WebShop, while adding only 0.16%-0.20% advantage-construction overhead.

We propose STRAT, an auxiliary task that trains deep reinforcement learning (RL) agents to predict a short textual trace of their own state. Inspired by human spatial navigation, the description combines landmark, route, and survey knowledge, tracking the agent's position, inventory, goals, and immediate progress. Environment rules generate this text online without human labelling. Our method adds a single auxiliary head to a standard policy. Across 60 sparse-reward XLand-MiniGrid tasks, STRAT solves complex environments where standard RL fails outright, while compacting state representations and preventing rank collapse. Beyond performance gains, the predicted trace provides a readable account of agent beliefs at every step for no extra cost.