Graph-based policy optimization improves credit assignment for long-horizon LLM agents by organizing rollout trajectories into state-transition graphs. However, existing methods construct graphs independently within each policy update, discarding transitions discovered by earlier policies and limiting advantage estimation to small, batch-local rollout groups. We propose \emph{Temporal Instance-Graph Policy Optimization} (TIGPO), which extends graph-based credit assignment across policy updates. TIGPO maintains a persistent transition graph for each task, allowing valid transitions discovered by different policy versions to jointly determine credit for current rollouts. To actively reconnect current exploration with historical experience, TIGPO allocates a fixed rollout budget between Exploration slots for ordinary task sampling and Revisit slots for delayed reattempts of previously explored tasks. For each revisit, TIGPO pairs the current rollout group with its corresponding earlier Exploration group to construct a cross-temporal reference. The enlarged reference is designed to stabilize relative advantage estimation under small rollout groups, while comparison on the same task directly captures policy improvement across training stages. Historical transitions and scores serve only as structural and detached statistical references and are never replayed in the policy loss. Experiments on ALFWorld and WebShop demonstrate that TIGPO consistently outperforms prior group-based and graph-based policy optimization methods.
Reinforcement learning typically optimizes average reward. For generative policies, the average can hide an important distinction: two policies can achieve the same mean reward while having very different chances of producing a rare but high-reward rollout. This matters as sampling increases during training and inference, since its benefit depends on retaining probability mass on high-reward outcomes. We propose to optimize this coverage directly. Rather than considering only expected reward, we consider all of its upper tails: for each reward threshold, how likely is the policy to exceed it? This turns a continuous reward into a family of binary success events. We introduce Tail-Likelihood Reinforcement Learning (TailRL), which maximizes the log-probability of exceeding a randomly chosen reward threshold. Its gradient gives more weight to rare, high-reward rollouts and can be interpreted as a mixture of Best-of-(k) gradients. TailRL requires only a simple modification to the advantage function, making it compatible with existing reinforcement learning pipelines. Across object localization, maze navigation, GUI grounding, and code optimization, TailRL leverages rare high-reward training samples to avoid suboptimal solutions and yields models that benefit more from additional samples at inference time.
Modern large language models (LLMs) rely on reinforcement learning to build strong capabilities in individual domains, but integrating those capabilities into a single deployable model remains challenging. By routing each sample to the teacher whose domain matches it, existing approaches let a domain label decide which teacher provides supervision. However, domain expertise holds only on average: the matched teacher is not always correct on a given sample, while a teacher from another domain sometimes is. The reliable teacher therefore has to be identified per sample, not per domain. In this paper, we introduce Multi-Teacher Self-Distillation Policy Optimization (MT-SDPO), an on-policy distillation method that unifies several frozen teachers into one student model. MT-SDPO consists of three components: (1) self-anchors, where a rollout is supervised by a correct rollout from its own group; (2) answer-verified eligibility, where a teacher may supervise a sample only if its own answer passes a verifier; and (3) privileged distillation, which merges the anchor and all verified feedback into one context that an exponential moving average self-teacher reads and the student does not, thereby keeping one policy at deployment. Across five students from three model families, MT-SDPO lifts the weakest domain of Qwen3-8B by 14.79 points and narrows its domain gap by 74.7%, a better balance than serving one matched teacher per domain. Verified reliability, not domain membership, should decide who teaches. Code is available at https://github.com/hexixiang/MT-SDPO.
Advertising recommendation requires continuously tuning complex system parameters while balancing commercial returns and user experience. Recent work has introduced large language models (LLMs) with skill documents to assist this labor-intensive process, but skill optimization remains largely prompt-driven, lacking a principled mechanism to attribute rewards to specific document edits. To address this limitation, we propose Document-Mediated Reinforcement Learning (DMRL), a skill self-evolution framework that models skill document optimization as a sequence of structured editing actions. In DMRL, an upper-level agent performs controlled document edits, while a frozen lower-level task agent evaluates their effects through A/B testing. To address credit assignment and long-term outcomes, we introduce two key components: (1) Dual-Relative Policy Optimization (DRPO), a post-training policy optimization method for robust and risk-aware advantage estimation; and (2) Long-term Reward Predictor (LRP), which estimates long-term outcomes by modeling population heterogeneity with disentangled representation learning and cross-attention transfer. DMRL was deployed on a large-scale short-video ads platform and extensive empirical evaluation shows that DMRL outperforms state-of-the-art baselines across key advertising metrics
Large-scale video retrieval requires embedding models to encode long and diverse videos under tight visual-input and inference budgets. Existing methods typically sample a small, fixed set of frames at their original resolution, limiting temporal coverage and ignoring frame importance. Our empirical analysis shows that expanding temporal coverage improves retrieval even under a fixed visual-input budget. Gains are larger when the original per-frame resolution is preserved, highlighting the complementary roles of temporal coverage and spatial fidelity. Motivated by this finding, we propose AllocEmbed, an allocate-then-embed framework that reallocates a fixed visual-input budget across more frames. A lightweight allocator uses low-cost previews to assign frame-wise resolutions before the embedding backbone, preserving more detail where it most benefits retrieval while reducing visual cost elsewhere. We further introduce Retrieval-Driven Policy Optimization (RDPO), which learns the allocator directly from retrieval feedback using a rank-validated similarity gap and a confidence-guided efficiency incentive. Operating entirely before the backbone, AllocEmbed integrates with existing retrieval systems without modifying the embedding model or downstream pipeline. Experiments on the MMEB-V2 V-QA and V-RET tasks and our LongRet benchmark show that AllocEmbed achieves the best overall retrieval performance among the evaluated budget-matched methods and transfers across embedding backbones. Our code is publicly available at https://github.com/jinsong8/AllocEmbed.
Baha Zarrouki, Arslan Thobani, Jasper Hoffmann +6cs.RO cs.LG eess.SY
In Model Predictive Control (MPC), cost-function weights shape closed-loop behavior, yet changing conditions often make fixed parametrizations suboptimal and motivate context-dependent online adaptation. Learning such policies is difficult because behavior depends implicitly on numerical MPC solutions, producing nonlinear, potentially nonsmooth, long-horizon dependencies on policy parameters. This creates a bias-variance tradeoff: Reinforcement Learning (RL) optimizes realized closed-loop return from environment samples but is sample-inefficient, whereas Gradient-Based Policy Learning (GB-PL) uses low-variance solver gradients from differentiable MPC to optimize surrogate losses on predicted trajectories but can be biased under model mismatch. We propose Solver-Gradient Guided Reinforcement Learning (SG-RL), a solver-sensitivity augmentation for RL-based online MPC cost-weight adaptation. SG-RL keeps sampled closed-loop return as the objective and uses bounded solver-derived gradients as auxiliary guidance to improve stability and sample efficiency. We instantiate SG-RL in Proximal Policy Optimization (PPO) with four modular algorithms that inject solver-gradient guidance into actor-update scaling, policy loss, advantage estimation, and value-function learning. On two full-scale autonomous racing platforms with intentional model mismatch, SG-RL reaches PPO's best closed-loop return with up to 70.6% fewer samples, outperforms GB-PL baselines by at least 54% in closed-loop return, and generalizes zero-shot to unseen environments.
As Vision-Language Models (VLMs) tackle dynamic 3D spatial reasoning, ego-motion perception becomes essential to resolve monocular scale ambiguity. However, current models often overfit to smooth trajectory priors rather than genuinely understanding physical motion. Consequently, their spatial reasoning degrades severely under large displacements, a phenomenon we term Kinematic Collapse. This failure stems from spurious visual-motion correlations in natural videos and a lack of explicit physical supervision. To evaluate this, we introduce Dyn-3D, a benchmark using counterfactual 3D rendering to rigorously decouple visual changes from true kinematic properties. Furthermore, we propose the TempoVista framework, featuring the Kinematic-GSPO algorithm. By embedding metric physical ground truth into policy optimization, TempoVista explicitly grounds visual representations in 3D space. Experiments demonstrate that our approach significantly improves both motion estimation and robust spatial reasoning by utilizing camera dynamics as an effective geometric calibration signal.
On-policy distillation (OPD) offers dense token-level supervision as an alternative to the sparse outcome-level advantages of reinforcement learning with verifiable rewards (RLVR). However, the teacher scores student-generated trajectories that are inherently off-policy for it, so the reliability of its supervision, and hence the source of the student's improvement, remains unclear. We quantitatively analyze teacher supervision during OPD training and find substantial noise whose prevalence increases with teacher scale. Surprisingly, the student policy is insensitive to such noise, converging to comparable performance regardless of whether noisy supervision is retained or removed. Does OPD distill at all? By analyzing what drives its gains, we find that learning concentrates on low log-probability tokens, and using a single fixed negative advantage matches the performance of teacher-provided ones. This suggests that OPD works largely by suppressing low log-probability tokens, which requires no teacher. These findings motivate On-Policy Self-Adaptation (OPSA), a supervision-free method using entropy-adaptive negative advantages. It assigns stronger learning signals to high-entropy positions, suppressing tail tokens, and evenly redistributing probability mass among head tokens. Compared with the base \texttt{Qwen3-1.7B}, OPSA improves Avg@32 by 35.41 points on AIME24, corresponding to a 263\% relative gain, and more than doubles Pass@32 across all three benchmarks. It also outperforms OPD by 16.77 points in Avg@32 on AIME24. Extensive experiments and analyses across model families and tasks further demonstrate its effectiveness and generalizability.
Expected-cost constraints can still permit rare, high-cost events. Monte Carlo conditional value at risk (CVaR) gradients can be noisy at high confidence, whereas critics that model an outcome distribution add complexity. We propose BCPPO (Bachelier-Inspired Constrained Proximal Policy Optimization), a proximal policy optimization (PPO) method. Separately initialized cost-prediction networks (critics), trained with random sample masks, produce disagreement that marks predictions sensitive to which state-action regions occur in the training data and to critic training. A Bachelier formula for the expected amount above a reference level converts this disagreement into a smooth policy-update penalty. Gradients from this penalty do not alter the critics, so temporal-difference (TD) critic learning is unchanged. A saturation-aware controller adjusts the mean-cost penalty and stops accumulated error from growing while that penalty is clipped. Deployment retains only the policy network. The disagreement penalty is neither a tail-event probability nor a guaranteed error bound, and it provides no safety guarantee. Across 175 runs with shared tasks, costs, budgets, training steps, and evaluation seeds, no comparator attains both higher mean return and lower mean CVaR than BCPPO in any task. On Push1, BCPPO has no lower return and no higher CVaR than every comparator, with at least one strict gain. These results support a practical balance among reward, caution around cost predictions that vary across trained critics, and policy-only deployment.
Large language model (LLM) agents are increasingly deployed in long-horizon, interactive, and stateful environments. In these settings, a single wrong action, such as refunding the wrong purchase, can cause irreversible task failure and must be intercepted before execution. Such failures may not appear in every single run, but can emerge across repeated trials, making reliability across steps and trials critical. However, ensuring agentic reliability is challenging: even frontier LLMs struggle to explain why an action may be wrong, especially in long, intertwined trajectories governed by domain-specific policies. Much recent work relies on prompt-based critique agents, while optimization-based methods lack a systematic way to produce rich verification rationales for training. We address this gap with CAST, a critique-aware training framework that converts sparse task outcomes into action-level supervision for critique learning and policy optimization. CAST analyzes agent trajectories to synthesize structured rationales explaining action validity under partial observability. The resulting critique model is used to construct critique-aware training data for optimizing the policy model. Fine-tuning Qwen3-family models on dynamic tool-calling benchmarks, CAST improves reliability across domains, outperforming GPT-OSS-120B by over 10% pass^4 on Retail tasks and yielding an additional 9% improvement on Telehealth in an out-of-domain setting. These results demonstrate that critique-aware training improves the robustness of LLM agents in realistic dynamic environments.
Tian Zhang, Zhuo Huang, Hongrui Ye +3cs.CV cs.AI cs.LG
Vision-language-action (VLA) driving methods increasingly combine multi-trajectory imitation learning with group-relative policy optimization (GRPO), making trajectory selection critical to final performance. However, some high-scoring trajectories that improve imitation can degrade subsequent GRPO by inducing advantage estimates misaligned with the current policy's feasible behavior distribution, driving updates away from safe and compliant behaviors. To address this, we propose a novel framework that aligns multi-trajectory supervision with policy optimization. To address the policy gradient bias induced by infeasible noisy trajectories outside the feasible region, augmented trajectories are constrained to a neighboring manifold of the ground-truth feasible region, and a Pareto-optimality criterion is adopted in place of the conventional aggregate score, retaining only non-dominated candidates and thereby filtering out conflicting samples at the source. To ensure that expanded trajectory supervision is effectively absorbed during policy optimization, we introduce two complementary mechanisms: feasibility-first advantage assignment and dynamic distillation. The former adapts Pareto credit to the feasibility composition of each rollout group and guides fully infeasible groups toward safe references. The latter updates teacher trajectories across refinement rounds to continually transfer useful supervision. Together, they progressively translate the benefits of expanded supervision into policy improvement. On NAVSIM v1 and v2, our method achieves 91.4 PDMS and 89.1 EPDMS, respectively, under single-trajectory inference, and recovers 440 of 658 initially failed scenes, 11.1\% higher than the original GRPO baseline.
Decoupled Clip and Dynamic Sampling Policy Optimization (DAPO) is a prominent variant of Group Relative Policy Optimization (GRPO). DAPO introduces several improvements over GRPO. Among these, Dynamic Sampling contributes the most to DAPO's accuracy gains relative to GRPO. To improve accuracy, Dynamic Sampling enhances training stability by eliminating zero policy gradients from zero advantages. Specifically, it avoids such zero gradients by filtering out prompts where sampled responses are either entirely correct or incorrect. However, our theoretical analysis shows that Dynamic Sampling decrease training efficiency as it cannot effectively utilize hard-to-sample correct responses on hard prompts. Formally, it asymmetrically amplifies the advantages of distinct responses to the same prompts. On hard prompts, incorrect responses undergo greater amplification than correct ones. This leads the model to avoid generating the observed incorrect responses rather than capitalizing on the hard-to-sample correct ones on hard prompts, resulting in low training efficiency. To improve training efficiency, we propose Direct Advantage Amplification (DAA), which amplifies the advantages of hard-to-sample correct responses on hard prompts, as obtained by Dynamic Sampling. This ensures that, when Dynamic Sampling is used, these hard-to-sample responses can be effectively capitalized on, implying higher training efficiency. By integrating DAA into DAPO, we obtain Difficulty-aware Advantage Amplification Policy Optimization (DA3PO), which is implemented with fewer than 30 lines of code from DAPO. Experiments show that DA3PO significantly outperforms GRPO and other classical GRPO variants.
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.
Recent prominent post-training methods, such as Reinforcement Learning (RL) and On-Policy Self-Distillation (OPSD), have driven rapid progress in mathematical reasoning for large language models, yet their reliance on ground-truth labels precludes test-time training (TTT). Replacing ground truth with majority-vote pseudo-labels is a natural alternative, yet it is fragile: an incorrect vote corrupts the teacher and misleads every token. We observe that this failure mode is asymmetric: rollouts that disagree with the pseudo-label are typically wrong regardless of whether the vote itself is correct. Building on this observation, we propose Test-Time Policy Optimization (TTPO), an asymmetric objective that distills agreeing rollouts via OPSD and penalizes disagreeing rollouts with Grouped RL. Token-level selection further refines both branches: distillation down-weights already-converged positions, while RL penalizes only confident errors. Both updates remain well-grounded even under frequent pseudo-label errors, and majority-vote routing yields tighter self-supervision as the model improves. Without any labels, TTPO matches label-supervised OPSD on five competition-level benchmarks, raises Qwen3-1.7B from 38.0% to 45.2% in TTT, yields +25.2% to +36.4% without thinking, and shows strong cross-task generalization.
Group-relative reinforcement learning waits for sibling rollouts of the same prompt, which is costly for long and variable tool-use trajectories. Single-stream Policy Optimization (SPO) removes this dependency with a persistent prompt-level value estimate, but its recipe whitens one advantage per trajectory before optimizing a token-mean actor loss. We show that trajectory centering generally does not center the token-weighted quantity consumed by the actor, and fix the mismatch by standardizing terminal-outcome advantages under the action-token measure. We additionally organize prompt evidence by the policy event that generated it rather than learner receipt order. Across matched runs on ALFWorld at two model scales and on Math-TIR, SPO++ improves online learning efficiency over SPO. A paired ablation identifies action-token-measure normalization as the strongest tested component.
Large Language Model (LLM) agents increasingly solve long-horizon tasks through multi-turn interactions with users and external tools. In these settings, relevant task information often unfolds over time rather than being fully specified at the initial prompt. Service agents make this challenge especially concrete: users may clarify or revise their goals, while tool responses provide information needed for subsequent decisions. Thus, a final reward alone cannot indicate which actions contributed to resolving the task. Recent methods rely on comparative evidence from other trajectories or resampled continuations, or on separately constructed step-level learning signals, to refine credit. However, a completed rollout already records how information and errors flow between agent actions. We introduce Influence-Aware Policy Optimization (IAPO), which represents each rollout as a typed influence-dependency graph over trainable agent actions, with user and tool observations serving as evidence. IAPO converts support-use and failed-use structure into routing weights that redistribute the same trajectory-level advantage. Experiments with Qwen3-4B and Qwen3-8B demonstrate superior performance over multi-turn reinforcement learning (RL) baselines across three service-agent benchmarks: {τ^2}-Bench, UserBench, and AgentChangeBench. BFCL-v4 Multi-Turn further shows that these gains do not compromise multi-turn function-calling performance. This work advances the understanding of credit assignment in multi-turn user interactions and provides a principled approach to training service agents from sparse outcome feedback.
Reinforcement learning with verifiable rewards (RLVR) enables language models to learn multi-turn interaction with external tools, yet its sparse outcome rewards provide no signal for identifying which intermediate decisions are responsible for success. Branch sampling induces local comparisons among alternative continuations, but existing methods tend to conflate two distinct problems: allocating a fixed rollout budget and translating branch outcomes into token-level credit. We introduce Contrastive Branch Policy Optimization (CBPO), which disentangles these two problems and assigns a dedicated mechanism to each. Generation entropy screens candidate branch positions across the entire response, while path-level and node-level decay distribute a fixed budget across trajectories and positions to prevent exploration from collapsing onto a few paths or adjacent tokens. A parent trajectory together with the branches that share an identical token prefix forms an exact-prefix group, and the reward variation within this controlled group defines the Contrastive Branch Value (CBV), an outcome-based estimate of local decision sensitivity that rescales continuation advantages without altering their sign. When multiple nodes are selected along the same trajectory, CBPO partitions it into non-overlapping credit segments, thereby avoiding duplicated gradients on shared tokens. Requiring only outcome rewards and no process-level annotation, CBPO provides a practical solution for fine-grained credit assignment in tool-integrated agent training. Extensive experiments on ten benchmarks, including five for mathematical reasoning and five for knowledge-intensive search, show that CBPO consistently outperforms state-of-the-art policy-optimization and branch-based methods, attaining the highest macro-average accuracy in both domains and across two model scales.
Agentic retrieval-augmented generation (RAG) requires language models to decide when to continue searching and when to answer. Existing RL-based methods rely on external supervision and overlook the agent's internal belief about whether the current evidence is sufficient. To address this problem, we reformulate the search decision quality as belief-action alignment and propose MetaRAG, a belief-action aligned policy optimization framework for agentic RAG. MetaRAG uses Verify-first Action Generation to elicit an explicit verification process before each actual action, and Internal Belief Probing to estimate the policy model's own answerability belief from the same question-history context. Based on these, MetaRAG derives a consistency reward that is further gated by answer correctness, avoiding reinforcement of internally consistent but incorrect trajectories. The belief probe is used only during training and introduces no inference-time overhead. Experiments on seven public QA benchmarks show that MetaRAG consistently improves the accuracy-efficiency trade-off over strong RL-based agentic RAG baselines, with gains that transfer to deep research settings, different optimizers, and multiple model backbones.
Self-reflection is a powerful mechanism for credit assignment in human learning, converting sparse outcome feedback into actionable guidance. However, its potential for post-training Large Language Models (LLMs) remains underexplored. We propose Self-Reflective Policy Optimization (SRPO), a framework that internalizes this capability. SRPO enables LLMs to analyze their own completed trajectories, synthesize errors into concise "reflection patches," and use reflection-conditioned teacher scores on student on-policy rollouts as dense token-level training signals. This process effectively transforms sparse terminal supervision into dense, token-level learning signals without requiring external critics, separate reward models, or larger teacher models. We demonstrate that SRPO achieves state-of-the-art performance across mathematical reasoning and long-horizon agentic benchmarks with exceptional data efficiency. Using a Qwen3-8B base model, SRPO attains 73.3% on AIME'24 using only 8% (0.08x) of the training FLOPs required by scaled supervised fine-tuning, while significantly improving success rates on WebShop (64.7%), ALFWorld (76.8%), and SWE-Bench-Lite (31.2%). Code is available at https://github.com/Galleons2029/SRPO
Policy optimization (PO) for Large Language Models faces a stability--exploration trade-off, currently mediated by an action-side Policy-KL regularizer. This puts practitioners in a double bind: keeping Policy-KL constrains response behavior and consumes the action-side exploration budget, while dropping it leaves the optimization without an explicit drift control. We argue for an alternative that breaks the dilemma by moving regularization to the input side. As training progresses, the distribution over training queries induced by the current policy drifts unchecked from its pre-RL reference distribution. Concretely, Environment-Regularized Policy Optimization (ERPO) introduces a Query-KL (QKL) term that bounds this query distribution shift, together with a dataset-static reference-derived per-query weight that biases each per-query update toward queries typical under the reference. The QKL gradient flows strictly through the query likelihood; the response score function used by policy-gradient estimators does not appear in the QKL term, so QKL exerts no direct gradient pressure on the response distribution---exploration is preserved. ERPO plugs into GRPO/PPO/REINFORCE-style pipelines without additional forward passes. On six mathematical reasoning benchmarks, ERPO replaces the standard Policy-KL regularizer while achieving effective control over query distribution drift, delivering stronger accuracy and substantially more stable behavior under high-temperature decoding and long-horizon training.Our source code are available at https://github.com/alibaba/ERPO
Multi-agent debate can improve large language model reasoning by eliciting diverse hypotheses and critiques, yet its performance is often constrained by weak moderation. Common pipelines rely on fixed budgets, agreement-based stopping, or untrained judges, leading to redundant deliberation and unreliable evidence aggregation. We cast moderation as a meta-cognitive process, monitoring debate utility, controlling deliberation, and adjudicating a final answer, and introduce Meta-Moderator, a learnable framework that dynamically regulates debate and decides when to finalize an answer. Meta-Moderator is trained independently of the debaters via outcome-driven policy optimization, making debate regulation an explicit capability rather than an incidental effect of prompting. Across five benchmarks, Meta-Moderator outperforms widely used decision layers and transfers across tasks and system configurations. Further analyses show that it allocates debate more selectively and reduces mis-aggregation after informative hypotheses appear.
Tool-Integrated Reasoning (TIR) is a fundamental capability for LLM agents to solve complex tasks by interacting with external tools iteratively. Reinforcement Learning (RL) has become the dominant paradigm for enabling this capability. However, existing approaches typically assign uniform trajectory-level advantages and treat all correct tool calls equally, ignoring the varying difficulty and learning value across trajectories and reasoning steps. This can lead to imprecise learning signals that do not adequately distinguish between trivial and challenging tool-use patterns. To address this limitation, we propose HiDiffTIR, a Hierarchical Difficulty-aware policy optimization framework for multi-turn TIR. HiDiffTIR performs difficulty-aware credit assignment at both trajectory and turn levels, enabling the policy to focus on more informative trajectories and harder reasoning steps. Notably, this fine-grained optimization is achieved without additional supervision, relying solely on group-level statistics derived from standard RL rollouts. Extensive experiments on three tool-using benchmarks demonstrate that HiDiffTIR consistently improves multi-turn TIR performance and tool invocation accuracy over strong RL baselines, highlighting the necessity of difficulty-aware credit assignment for effective policy optimization in tool-integrated LLM agents.
Agentic reinforcement learning (RL) has become a critical stage in the post-training of large language models. Existing critic-free, group-relative methods estimate policy advantages from multiple rollouts, avoiding the substantial memory overhead of conventional proximal policy optimization (PPO) and achieving strong performance on long-horizon interactive tasks. Despite their success, recent studies revealed three limitations: (1) Lack explicit value generalization and effective temporal credit assignment; (2) Suffer from potential advantage collapse in long-horizon complex tasks; (3) Require a costly trade-off between sampling budget and policy performance. In this work, we propose Single-rollout Autoregressive Policy Optimization (SAPO), a low-memory and compute-efficient framework in which the policy and value functions share a single autoregressive backbone. SAPO exploits the autoregressive structure of LLMs to produce policy and value predictions at distinct causal boundaries with shared parameters, while independently optimizing the PPO objectives and auxiliary on-policy SARSA objectives. To robustly estimate the contribution of each turn, we further introduce a trajectory-level generalized advantage estimator that combines lambda-returns with batch normalization. Experiments across ALFWorld and WebShop with Qwen2.5-1.5B/7B show that SAPO trains stably and outperforms PPO and GRPO by mean +15.1 and +12.1 percentage points, respectively, while eliminating the memory cost of a separate critic model and reducing per-iteration runtime by 33.2% over PPO.
Bo Qian, Yuting Wu, Shuang Zeng +3cs.LG cs.AI cs.CL
Credit assignment is challenging in long-horizon agentic reinforcement learning, where supervision often comes only from final rewards. Existing methods refine trajectory-level signals into step-level credits through step grouping or graph-based advantage estimation, but can overlook meaningful intermediate milestones. We propose MileGPO (Milestone Inference with Local Evidence for Graph-Based Policy Optimization), which derives process-level credit from grouped on-policy rollouts through three designs. Milestone Discovery identifies candidate milestones on successful rollouts and recurring traps on failed ones. Reliability-Calibrated Shaping (RCS) weights these candidates by outcome-based confidence, strengthening reliable milestones and traps while down-weighting uncertain ones. Progress-Contrastive Calibration (PCC) further tests whether a candidate reflects local progress and whether its incoming ansition outperforms observed alternatives from the same state.MileGPO requires neither auxiliary models nor additional environment interaction. Experiments on ALFWorld and WebShop show state-of-the-art performance and a small in-distribution to out-of-distribution gap on ALFWorld. Ablations and credit diagnostics indicate that reliability weighting, local progress, and same-state branch evidence complement milestone discovery and resolve ambiguous intermediate credit.
Medical image captioning requires translating heterogeneous visual evidence into concise clinical descriptions, where errors in findings, assertion states, or anatomical relations can alter clinical meaning despite surface-level fluency. Sequence-level policy optimization can directly optimize complete captions, but common rewards rely on global text similarity, direct image-caption compatibility, or unordered concept overlap, leaving visual neighborhoods and clinical-claim structure implicit. We propose a clinically structured surrogate reward framework for post-SFT medical image captioning. The framework combines biomedical semantic and short-range lexical fidelity with two structured rewards: distributional image-neighborhood alignment, which matches the medical-image-bank distributions induced by reference and generated captions, and clinical graph consistency, which applies maximum-weight one-to-one matching to entities, assertion states, and typed relations. The four rewards are independently normalized within each rollout group, combined with fixed relative weights, and optimized with GDPO. Across organizer-evaluated hidden test sets for the Standard and Synthetical ImageCLEFmedical Caption tracks and three vision-language backbones, the method improves Overall, Relevance, and Factuality over matched SFT baselines in all six backbone-track combinations, with average relative gains of 3.4%, 2.1%, and 5.8%, respectively. Ablations and paired diagnostics indicate that the structured rewards provide complementary signals, reducing image-neighborhood divergence and improving entity-assertion-relation consistency.
GRPO is increasingly used for reinforcement learning of vision-language-action (VLA) policies because, unlike PPO, it does not require training a critic. This simplification comes with a sampling cost: group-relative advantages require multiple rollouts from each scene. Under binary success rewards, groups whose rollouts all succeed or all fail have zero advantage and are discarded by dynamic sampling. These groups are especially common early in training, when most rollouts fail, wasting much of the expensive robotic rollout budget. We introduce Prism-GRPO, which augments binary outcome reward with a weighted trajectory-level execution-quality score. By splitting same-outcome groups into a quality spectrum, Prism-GRPO recovers training signal while ensuring that every success still outranks every failure. Quality scores can be derived from simulator contacts, executed actions, or visual observations, avoiding task-specific progress rewards. We prove that Prism-GRPO never increases the probability that a sampled group is discarded for having zero advantages, and derive a gradient-alignment condition under which its combined update remains a local ascent direction for task success. Across four RoboTwin tasks spanning different horizons and coordination patterns, Prism-GRPO improves success and quality at matched rollout budgets and reaches target success rates with up to 56% fewer rollouts. It also suppresses a reward-hacking shortcut, with the cleaner behavior transferring under direct deployment to a real robot. Through ablations, we show consistent gains across contact-, smoothness-, and VLM-derived quality signals.
Group Relative Policy Optimization (GRPO) has become a widely used approach for post-training Large Language Models (LLMs) for reasoning. In GRPO, the group gradients induced by different queries within the same mini-batch are directly averaged to form the policy update. However, these group gradients can point in conflicting directions. Our empirical analysis suggests that group-gradient conflicts tend to be associated with less effective policy updates, motivating the need for a reliable aggregated update direction under such conflicts. Standard GRPO aggregation treats the realized group gradients as deterministic contributions and does not account for differences in their reliability during aggregation. To address this issue, we propose Gradient Uncertainty-Aware Policy Optimization (GUPO), which models each group gradient as a random variable under a Bayesian formulation and estimates its probability distribution. GUPO then derives gradient uncertainty using a Dirichlet-based formulation and uses it to calibrate the contribution of each group gradient during aggregation. Extensive experiments on multiple benchmarks demonstrate the effectiveness of GUPO.
Multimodal harmful meme detection is typically formulated as image--text harmfulness classification. A model may correctly predict harmfulness while misidentifying the attacked target or its supporting evidence. We therefore extend harmful meme detection with fine-grained target identification, asking what type of target is attacked, who is targeted, and where the target appears in the meme. The model predicts harmfulness for every meme and, for harmful memes, outputs the target category, target entity, textual mention, and visual region. To support this task, we introduce Meme3W, which unifies multiple public harmful meme datasets and provides human-verified annotations for harmful instances. We further introduce Joint Record Accuracy (JRA), a strict record-level metric requiring the harmfulness label and all target-identification fields to be jointly correct. Experiments with representative multimodal large language models reveal a substantial gap between harmfulness accuracy and JRA. To narrow this gap, we propose HarmTrace, an anchor-calibrated decoupled optimization framework. HarmTrace strengthens target-entity supervision through entity-aware supervised fine-tuning. It then applies Conditional Target-identification Policy Optimization (CTPO) to decouple harmfulness and target-identification advantages, restricting target-identification optimization to label-correct responses for harmful examples. CTPO uses a Virtual Positive Anchor (VPA) as a fully correct reference for target-identification advantage normalization. HarmTrace improves both JRA and harmfulness accuracy across the evaluated backbones, with JRA on the Qwen3-VL-8B backbone increasing from 17.58\% to 52.51\%. Our code is publicly available at https://github.com/llly1234/HarmTrace-for-Harmful-Memes.
Reinforcement learning (RL) with group-relative advantages has become the de facto standard for post-training language model reasoners. However, when optimizing multiple reward objectives, existing methods typically scalarize the reward vector with a fixed weighted sum before group-wise standardization. We show that this design leads to two fundamental problems: rollouts with distinct reward profiles can receive identical advantages, and all objectives are optimized with fixed relative weights regardless of their current level of saturation. As a result, training continues to allocate gradient budget to already-solved objectives instead of focusing on those with greater remaining headroom. We introduce \textbf{Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization} (SA-MRPO), which standardizes each reward objective independently and adaptively discounts its contribution according to a batch-level estimate of objective saturation. This dynamically reallocates optimization effort toward under-optimized objectives while empirically maintaining performance on those that are already well satisfied. We further show that saturation-aware reweighting can reverse the sign of an update, rather than merely rescale its magnitude. Across mathematical reasoning with two- and three-objective reward combinations, SA-MRPO improves the harder correctness objective over GDPO in 12 of 15 benchmark comparisons, with gains of up to $5\%$ on AIME24. On adaptive reasoning it improves accuracy on all five benchmarks, by $3.8\%$ on average and up to $9.2 \%$ on AMC23, and on coding benchmarks it improves pass rate by up to $2.3\%$, while in all settings maintaining the easier objectives near their already satisfied levels.
Group Relative Policy Optimization (GRPO) learns from reward differences within a rollout group, but receives no useful relative signal when every sampled response is incorrect. Privileged self-distillation can fill this gap with dense token supervision, yet applying it throughout training creates a different failure mode: the teacher is a biased, low-variance surrogate for the reward objective, so persistent imitation can oppose reward-improving updates after the policy becomes capable of producing successful trajectories. We introduce I-SDPO (Instance-Level Adaptive Self-Distillation Policy Optimization), which treats teacher reliance as capability-dependent. I-SDPO makes one routing decision per input instance and shares it across that instance's rollout group: all-incorrect groups use a privileged self-distillation objective, whereas any-success groups remain intact for GRPO. This design uses imitation only where group-relative rewards are uninformative. A local analysis characterizes when teacher and reward directions align and shows that a non-vanishing biased distillation weight induces an optimization bias floor. The routing rule automatically reduces the expected distillation rate as success probability rises, withdrawing teacher influence without a hand-designed schedule. On SciKnowEval, I-SDPO obtains the best result in all four scientific domains and improves average mean@16 accuracy from 56.67% with GRPO to 70.31%, with a maximum domain gain of 18.24 points.