Trajectory-Level Credit

Recent VLM-based autonomous driving planners adopt GRPO-style reinforcement learning to optimize driving performance. However, existing GRPO recipes either optimize driving efficiency, risking progress-seeking but unsafe behavior, or enforce early safety constraints, leading to overly conservative behavior; both require lengthy training. To solve these problems, we first reveal two distinct RL regimes: a progress regime (Run-GRPO) that aggressively explores high progress, and a safety regime (Walk-GRPO) that restores safety under stable progress. Based on this finding, we propose $\textit{Run-then-Walk}$, a simple yet effective two-stage reward scheduling strategy for GRPO, achieving both better performance and faster convergence. Unlike one-stage RL, which may focus on progress, safety, or a mixture of both within a single training phase, this schedule explicitly separates progress discovery from safety repair. In the $\textit{Run}$ phase, we focus on progress, allowing the policy to escape the conservative bias and discover high-progress modes. In the subsequent $\textit{Walk}$ phase, we introduce endpoint and safety strategy to repair unsafe behaviors from the Run phase. This reversed schedule overcomes the conservatism of Walk-first methods and the unsafe progress-seeking of joint optimization. We validate it with various VLM-based planners on multiple benchmarks: NAVSIMv1, NAVSIMv2, Navhard, and nuScenes. Extensive experiments demonstrate improved driving performance while requiring 40--50% fewer RL training epochs than the baselines.

Hierarchical planning frameworks combine skills from multiple robot control policies for long-horizon task execution, where determining when to terminate the current skill and advance to the next subtask is essential. Existing approaches often rely on pre-designed completion signal checkers that are hard to obtain in real-world execution. Large-scale vision-language models (VLMs) offer strong reasoning capabilities, but their decision boundaries are not inherently aligned with task completion criteria, while cloud deployment and lengthy reasoning introduce substantial latency, limiting real-time monitoring. We propose StageGuard, an agentic distillation framework for accurate and efficient stage-transition decisions. StageGuard combines teacher-model reasoning with demonstration trajectories to generate structured explanations of subtask completion and policy switching. A lightweight student VLM uses these explanations to generate compact self-explanations, which are used for supervised fine-tuning. We evaluate stage-transition prediction on trajectories from two benchmarks and assess closed-loop task success through integration into hierarchical robot control on BEHAVIOR-1K, with further validation on real robots. Results show substantial improvements in stage-transition prediction while supporting efficient online monitoring.

Spatial intelligence requires more than describing object locations. Under incomplete observation, models must identify and acquire missing evidence, interpret it in a common spatial frame, and act on it. We introduce VA-Bench to evaluate the complete observe-reason-act-revise loop. General-purpose MLLMs learn procedural context from RGB-only demonstrations, actively select camera viewpoints, issue metric Cartesian commands, and revise them from execution feedback. Models receive no privileged object poses, oracle trajectories, or learned action heads. A fixed model-agnostic controller executes only model-specified targets. VA-Bench contains 14 base task families (11 single-arm and three dual-arm), seven held-out geometry/layout variants, and a long-horizon five-object composition track. We evaluate 12 primary model conditions in three independent runs over the same 20 physically verified seeds per base task, reporting terminal success, nine trajectory-level behavioral diagnostics, and subtask progress. First, the best-performing model scores 100.0% on target localization and 78.9% on spatial relations in the annotated run. Its three-run macro-average task success is only 53.93+/-3.17%. Second, active camera control significantly improves task success over passive multi-view observation. In one matched comparison, success rises from 27.86% to 57.50%. Third, held-out geometric transfer can reduce task success by over 30 percentage points. No model completes a strict long-horizon episode, despite substantial partial progress. VA-Bench thus tests whether general-purpose MLLMs can turn visual demonstrations and actively acquired evidence into successful embodied action.

Uncertainty estimation for Vision-Language-Navigation (VLN) models is a critical task since it can help identify ambiguous and unreliable predictions, enabling agents to make safer navigation decisions. As one of the most advanced uncertainty estimation frameworks, conformal prediction (CP) offers a promising approach for uncertainty estimation in VLN. However, given that VLN agent requires a sequence of steps, standard calibration in conformal prediction fails to provide coverage guarantee it promises over a dependent, variable-length VLN episode. To this end, we propose Episode-Normalized Conformal Prediction (ENCP), which rescales a nonconformity score by the policy's residual confidence and calibrates one maximum score per episode. Under exchangeable calibration and test episodes, this construction covers the ground truth at every step with probability at least $1 - α$, while allowing dependence among steps within an episode. Across four VLN policies and three nonconformity scores on R2R and REVERIE dataset, ENCP meets all reported empirical step-coverage targets on the seen-to-unseen evaluation. These results demonstrate that ENCP can provide model-agnostic uncertainty estimates, which might be useful for determining when a VLN agent should defer to a more capable predictor, including human assistance.

Reliable web-agent monitoring is difficult when model-internal uncertainty signals such as token logits are unavailable. In this work, we study prefix-level risk prediction for web agents using observable trajectory signals: given an evolving prefix, estimate whether the current execution remains on track or is tending toward failure. We derive two observable trajectory representations: Macro features summarize cross-step agent--environment behavior and feedback, while Micro features measure the consistency of intention, action, and anticipated state change through repeated black-box queries. Instead of inheriting the final result label, we label the first critical error that remains uncorrected in the observed continuation and is associated with final failure as a key-step boundary, preserving valid early prefixes of failed trajectories as on track. Across WebArena-Lite and Online Mind2Web web agent benchmarks with five open- and closed-source backbones, observable trajectory signals are competitive with internal-signal baselines. The resulting predictors also support early intervention under fixed false-cut budgets and transfer across held-out website categories. These findings show that observable trajectory signals support valuable risk prediction abilities.

Outcome-driven reinforcement learning offers a scalable way to post-train vision-language-action (VLA) policies from sparse task-success feedback. In common GRPO-based VLA post-training, one rollout-level advantage is applied to every action in the trajectory. A rollout that completes several valid stages but fails later can therefore penalize the actions that produced its earlier progress. We call this trajectory-level credit aliasing. Temporal GRPO addresses this problem by constructing detectable task stages, aligning each rollout with stage-specific action intervals, and comparing only rollouts that have entered the same stage. The resulting stage advantages are applied to their corresponding intervals in a single policy update. On RoboTwin 2.0, Temporal GRPO improves task success and sample efficiency, with consistent gains across task horizons. Controlled updates on LIBERO-Long preserve shared prerequisite stages and concentrate improvement at the first stage where rollout outcomes diverge.

Safe offline RL typically assumes access to dense per-step cost annotations, but in practice supervisors provide only trajectory-level stop-feedback: a binary signal at the first unsafe transition, with no per-step attribution. We frame this as a temporal credit assignment problem and propose the Redistribution-based Cost Inference (RCI) framework, which converts sparse stop-feedback into dense per-step costs via return decomposition, then trains a constrained offline policy on the augmented dataset. We show that return-equivalent redistribution preserves the feasible policy set and the optimal Lagrangian in a CMDP, establishing that the transformation is lossless in theory while yielding better-conditioned cost critic learning in practice. Experiments on highway driving and robotic manipulation demonstrate substantially lower violation rates than sparse and classifier-based baselines, with robustness to heterogeneous dataset compositions and label noise.

Flow-matching models are now a mainstream method to image generation, but its adaptation to diverse downstream scenarios typically relies on post-training, which may cause conflicts among task-specific optimization objectives. Reinforcement learning enables direct optimization of task-specific rewards beyond the original models, yet trajectory-level optimization may incur high-variance gradients and cross-task interference. On-policy distillation (OPD) offers dense and stable supervision on student rollouts, but conventional teacher matching remains imitation-based. We propose DreOPD, a Degraded-reference extrapolative OPD method for flow-matching models that bridges these two paradigms. Our DreOPD converts implicit reward extrapolation into closed-form velocity regression, enabling extrapolative post-training with the stability of OPD. It further uses a mildly degraded reference to strengthen the teacher-reference contrast, yielding a clearer extrapolation direction. Experiments on single- and multi-teacher settings show that DreOPD outperforms OPD and multi-task RL baselines in average performance, while surpassing specialized teachers on most metrics.

Credit assignment in multi-turn agent reinforcement learning operates at two levels: assigning trajectory-level credit to actions and distributing each action's credit across its tokens. In this paper, we introduce FACTOR, which separates these decisions. FACTOR uses checkpoint-calibrated TD residuals to assign per-action credits that telescope to the trajectory advantage, and feedback-conditioned teacher-student likelihood gaps to allocate each credit across the realized action tokens. Per-action normalization preserves the action-average coefficient and prevents token-level sign flips. We pair this construction with an action-mean reduction, removing the implicit dependence of an action's scalar surrogate weight on its token length. At the behavior policy and before clipping, each action's inner action-mean surrogate equals its TD credit. FACTOR consistently improves over competitive baselines across ALFWorld, WebShop, and ScienceWorld, with every environment-seed comparison favoring FACTOR and the largest gains emerging on the longest-horizon environment. The same hyperparameters transfer without retuning to a larger backbone and to a different model family. Ablations identify TD action credit as the dominant driver of the improvement, with hindsight token allocation contributing complementary gains.

Web agents observe a browser through text, pixels, or both, and the choice is usually fixed once for all tasks. We measure six observation modes across eight site-model combinations (cells) on VisualWebArena and WebArena and ask what choosing per task would buy. The modes are complementary: each solves tasks the others miss, they fail in structurally different ways, and the best choice reverses between task sets. The obvious prize, an oracle that picks a winning mode for every task, looks large but is inflated by run-to-run noise: rerunning the same mode on the same tasks changes 12-14% of outcomes, so a second run of a mode already in hand gains about as much as adding a new one. What survives is a cost bound: sending only the tasks no mode solves to the cheapest mode cuts cost by 9.5-30.6% in 8 of 8 cells at unchanged success. We then test five routing policies (picking the mode, deciding when to spend on the strong mode, a zero-cost rule read off the task text, a confidence cascade, and pooled cost tiers), and none robustly beats simply fixing one well-chosen mode; the one exception is a fragile result in our sparsest cell. The central obstruction is that routing supervision is produced at the agent's success rate: the weaker the agent, the fewer labels a router gets, exactly where routing would be most valuable. This limit belongs to today's agents rather than to routing itself. Label supply and routing opportunity rise together (correlation 0.95 across cells), so a stronger agent can overturn the result, and we report the rerun noise bands and the full measurement protocol.

Vision-and-Language Navigation (VLN) requires an agent to follow a route-level instruction by executing its constituent steps from egocentric visual observations. Existing VLM-based navigators typically supervise both capabilities through next-action prediction alone, making progress-tracking errors difficult to distinguish from execution errors. When an agent deviates from the route, a corrective action label may recover the next movement but does not indicate whether the agent selected the wrong sub-instruction or failed to execute the correct one. Consequently, the agent may continue making decisions from an erroneous progress state. To resolve this ambiguity, we propose \textbf{Route2Step}, a framework that decouples semantic progress tracking from action generation through an explicit step-level interface. The Instruction Analysis Module ($\mathcal{M}_{\mathrm{IA}}$) predicts this state from the global instruction and visual history. Conditioned on the predicted state and recent observations, the Action Generation Module ($\mathcal{M}_{\mathrm{AG}}$) generates local action chunks. To supervise the progress state without manual temporal labels, E-SPA, a step-alignment procedure, associates sub-instructions with their corresponding portions of route-level demonstrations. These alignments enable state supervision for incorrect progress estimates, while direct action supervision is reserved for rollout groups that repeatedly fail under the correct active sub-instruction. On R2R-CE, Route2Step improves SR from 48.1% to 55.3% and SPL from 43.3% to 48.2%, using 190K state-level corrective samples while requiring only 11.5K directly action-supervised states. Experiments in real-world indoor and outdoor environments further demonstrate the practical applicability of Route2Step. The project page is: https://sisyphus-hxy.github.io/Route2Step/.

Search-augmented LM agents are typically trained with a binary exact-match reward, which throws away most of what a failed trajectory tells us about why it failed. We introduce HindSearch, a hindsight self-distillation procedure for GRPO: after each rollout, a frozen judge writes a short critique of every failed trajectory using the gold answer, and the critique supplies an auxiliary on-policy distillation signal on the student's search actions. On the standard seven-benchmark suite with Qwen2.5-3B-Instruct, HindSearch reaches 39.4% average EM, outperforming prior search-RL baselines. Removing the judge's access to the gold answer erases most of the gain, isolating hindsight as the source of the improvement.

Offline Reinforcement Learning (RL) agents are trained on fixed behavioral trajectories, which makes trajectory-level deletion important when selected data must be removed after training. Evaluating such deletion is difficult because a lower membership score can reflect trajectory removal, residual memorization visible to another attack, or policy collapse that destroys useful behavior. We introduce Trajectory-level memOrization and Unlearning in offline RL (TOUR), a benchmark that combines trajectory-level partitioning, matched non-member controls, retraining references, retained-performance anchors, and multi-attack privacy auditing. Across D4RL locomotion experiments and an exploratory AntMaze extension, TOUR shows that common deletion baselines have environment-dependent privacy-utility behavior. Retraining and fine-tuning often provide stronger retained-utility references than uniform GA+Refit, while TrajDeleter remains a useful comparator but is not uniformly stronger under the same audit. Reference-model, threshold, deviation, equivalence, action-error, representation-based, and query-limited attacks further show that a single likelihood-based membership score can overstate deletion quality. In the evaluated settings, conclusions about offline RL unlearning are therefore not stable under single-score auditing. They depend on matched non-member construction, retraining-relative calibration, attack family, retained utility, and explicit scope for diagnostic architecture or component-level evidence.

LLM preference alignment aims to optimize models toward human preferences across diverse user instructions. Reinforcement learning has become a major post-training approach for this goal, but existing proxy rewards are often outcome-level, mainly evaluating the final response while providing limited guidance for the reasoning trajectory. This can make credit assignment coarse when multiple responses receive similar final scores, leaving trajectory-level preferences under-specified. To address this limitation, we propose Thinking Checklist Reward (TCR), a process-oriented reward for RL-based preference alignment. TCR converts preference pairs into sample-specific thinking checklists and uses them to evaluate whether the generated reasoning trace addresses the preference-implied considerations. To reduce overlap with outcome-level supervision, TCR further introduces an exponential moving average (EMA) residual formulation to isolate a complementary thinking surplus beyond what is predictable from the outcome reward. Experiments on five models from three model families show that TCR consistently improves alignment performance across diverse benchmarks, with ablations further validating the importance of EMA-based residual formulation and sample-specific checklist supervision.

Supervised fine-tuning (SFT) of open-weight LLMs on expert agent trajectories has emerged as a prominent approach to building capable code agents without reliance on proprietary models. A central yet underexplored question is how trajectory quality and quantity jointly shape model performance. We present a systematic empirical study of trajectory data filtering for LoRA fine-tuning of Qwen2.5-Coder-7B-Instruct on the SWE-trajectory dataset (67,074 trajectories, of which 32,161 are resolved). We propose a two-axis quality scoring framework -- Efficiency and Style -- and evaluate it through 16 controlled experiments spanning strategy, scale, and ablation analyses. Since 7B-scale models attain near-zero SWE-bench resolve rates, we adopt cross-entropy (CE) loss on held-out trajectories as the primary metric, validated via first-action generation: CE loss and ROUGE-L are perfectly rank-correlated (Spearman $ρ$ = -1.00), with limited-sample evidence supporting but not conclusively establishing this proxy. Our results reveal a scale-dependent quality-quantity trade-off: at small scales, doubling the dataset (500 to 1,000) yields ~12.7% CE-loss reduction whereas the TopQ-Random gap stays <1% (Mann-Whitney p > 0.10); at 2,000 trajectories this same gap widens to 3.6% (p = 0.016). Ablation further identifies error-retry rate as the dominant sub-dimension, performing comparably to the full composite ($Δ$ < 0.2%). Together, these findings establish trajectory-level quality scoring as a viable but scale-sensitive lever for code-agent SFT and offer a proxy-validated evaluation protocol for the regime where end-to-end resolve rate is statistically infeasible.

Urban traffic congestion significantly increases fuel consumption, greenhouse gas emissions, and commuter delays, resulting in substantial economic losses and environmental harm in modern cities. Traditional traffic signal control strategies such as fixed-time scheduling, actuated control, and reinforcement learning (RL)-based methods, offer different degrees of adaptability; however, RL-based methods can require extensive retraining, careful reward design, and substantial simulation data when transferred across networks or demand regimes. To address these challenges, we propose HiLLTS, an LLM-guided traffic signal control framework that employs a hierarchical three-layer architecture consisting of a central coordination agent, a district layer and multiple cluster-level intersection agents. Experimental results demonstrate consistent improvements in both congestion and environmental performance. Compared with the strongest non-LLM baseline in each scenario, HiLLTS reduces average waiting time by 36.73% under the low-congestion scenario and 14.71% under the high-congestion scenario, while reducing average CO2 emissions by 7.87% and 8.57%, respectively. Larger gains are observed against weaker baselines: under low congestion, HiLLTS achieves reductions of up to 18.00% in emissions and 62.07% in waiting time relative to Fixed-Time control; under high congestion, reductions of up to 28.89% in emissions and 40.36% in waiting time are observed relative to Max Pressure. The ablation study further validates the contribution of LLM-guided coordination over rule-based control

Active visual agents solve fine-grained image tasks by interleaving reasoning with image-grounding actions across multiple turns. However, deployment-time rollout budgets are rarely fixed: some requests permit long rollouts, while others require the agent to act under a tight turn limit. Existing methods train the policy as if the rollout budget were hidden, so when the available budget is smaller than the trajectory the agent prefers, the interaction is often truncated before any valid answer is produced; we term this failure \emph{catastrophic truncation}. To overcome this challenge, we present AdaTurn, a budget-aware framework that conditions the agent on the allowed number of turns and explicitly trains the boundary behavior induced by the budget. Our key component, Forced-Answer DAPO (FA-DAPO), converts the over-budget event from a masked or penalized failure into a trainable final-decision step, teaching the model to synthesize partial evidence when further tool use is no longer possible. We further randomize rollout budgets during both training and inference and introduce a load-balanced scheduler that makes such operations practical. AdaTurn substantially improves low-budget accuracy, for example raising VisualProbe-Medium from 36.7% to 47.6% at four turns, while preserving strong scaling at larger budgets and transferring effectively to multiple backbones and general multimodal benchmarks.

One high-level instruction in long-horizon manipulation can cover several action stages. We use segmented action annotations as an intermediate representation between the full-task instruction and VLA action chunks. A progress module tracks the active stage, while the action policy receives stage information either as current-stage text or as a normalized ordinal stage index in robot state. We compare these interfaces with GR00T N1.6 on LIBERO-10 under direct fine-tuning and continuation fine-tuning from a full-task instruction baseline. Under direct fine-tuning, full-task instruction, current-stage text, and Ordinal Stage-State achieve mean success rates of 57.45%, 50.24%, and 54.36%, respectively, showing that explicit stage information does not automatically improve the policy. Under continuation, the corresponding means are 49.07%, 50.00%, and 53.75%, with Ordinal Stage-State exceeding both alternatives in all three paired runs. The observed benefit differs across interface representations and training arrangements.

Despite progress in Embodied AI, Vision-and-Language Navigation systems remain vulnerable to adversarial visual disturbances. Most existing methods rely on white-box access to target model gradients, which is often unrealistic for real-world deployed systems and computationally exhaustive due to recursive backpropagation for optimization, limiting their applicability. While previous black-box methods predominantly target single-step, instantaneous decision tasks, they struggle to handle the task complexities and temporal dependencies. This highlights the need for a gradient-free attack method that can effectively disrupt the multistep sequential perception-action loop using only observable inputs and outputs. Therefore, we propose AdvNav, a behavior-guided black-box adversarial attack framework that disturbs an agent's first-person views during navigation. To construct an informative surrogate objective for effective optimization guidance in gradient-free search under the black-box setting, we design a dual-granularity behavior-based feedback, aggregating a trajectory-level performance score representing overall navigation degradation, an action-level reward score considering the potential decision risk, and a deviation indicator, all of which are extracted from the agent's self-output behaviors. This feedback guides a hybrid optimization strategy that heuristically tunes perturbation strength via adaptive updates and evolves noise spatial structure genetically, to iteratively discover the most disruptive noise configuration. Evaluated against Transformer-based HAMT and LLM-based MapGPT with two types of backbones on R2R dataset, AdvNav achieves 49.70/65.96/87.30% Attack Success Rate. The result demonstrates the effectiveness and generality of AdvNav, reveals critical perception vulnerabilities and offers insights for the design of future resilient VLN models.

Optimizer experiments observe responses to algorithmic configurations without uniquely revealing hidden mechanisms. We develop a causal optimizer interaction calculus that separates pathwise realization, Mobius decomposition, and experimental identification. Under a fixed innovation coupling, every finite-horizon innovation-driven optimizer admits a behaviorally minimal pathwise realization. For any finite effect support and intervention design, an incidence operator gives the complete observational gauge, exact identifiability, sharp quotient stability, held-out predictions, and exact noiseless configuration complexity. Smooth hidden relaxation generates interactions through inverse hidden-state stiffness. Building on this structural law, we prove an observable-readout transfer theorem: arbitrary smooth update or trace readouts inherit an explicit five-term interaction through first and second hidden responses. Unlike the reduced optimal value, a general readout has no universal interaction sign. Its Boolean effects remain exact integrals of continuous interaction curvature and can therefore be identified by factorial interventions. We also derive Gaussian quotient minimax risk, exact confidence sets and tests, misspecification decomposition, certified downstream decisions, and optimal replication. A controlled real-data experiment on a 65-dimensional strongly convex logistic model validates the complete reduced-value chain. Boolean effects and independently integrated curvature agree within 4.21e-11, while nine held-out continuous intensities agree within 8.88e-13. Gaussian campaigns attain the predicted coverage and power, and 4,500 real-minibatch observations reject an order-two interaction model. Neural trace audits provide complementary evidence that the declared response classes remain informative in nonconvex training.

Computer-use agents must solve long-horizon tasks through repeated interaction with partially observable, multimodal desktop environments. Although imitation learning and offline trajectory refinement provide strong priors, static traces cannot cover the causal feedback loop of real computer use: each action changes the screen state, future action space, and recovery options. EvoCUA-1.5 extends self-evolving computer-use agents from offline experience learning to online reinforcement learning, where policies interact with executable sandbox environments and improve from verifiable task outcomes. Online RL in this setting requires more than directly reusing single-turn language-RL recipes. Multi-turn interaction introduces context-managed observations, sparse terminal rewards, variable-length trajectories, and slow environment feedback. EvoCUA-1.5 addresses these challenges with Step-Level Policy Optimization (STEPO), which preserves trajectory-level advantage balance after decomposition into step-level samples; policy-aware filtering and pass-rate calibration over verifiable synthesized tasks; Dynamic Tri-Adaptive Curriculum (DTAC), which combines learnable tasks, difficult positive replay, and controlled infeasible-task exposure; and a fully asynchronous RL infrastructure with staleness control and mini-group batching. Experiments show that these components improve training stability and downstream performance. EvoCUA-1.5 achieves 63.2% success on OSWorld-Verified, outperforming comparable 32B/35B-scale open-weight baselines and even approaching models with significantly larger parameter counts. Overall, EvoCUA-1.5 provides a practical framework for scaling online RL in multi-turn computer-use agents.

On-policy distillation (OPD) trains a student policy by matching a stronger teacher on the student's own trajectories, offering a promising framework for language agent training. However, its application to long-horizon agentic tasks remains insufficiently explored. We identify two key inefficiencies in vanilla agent OPD: (1) full-horizon rollouts often waste wall-clock resources on tail turns that provide weak and noisy KL supervision, and (2) trajectory-level KL objectives concentrate most of the loss on shallow tokens, leaving deeper decision turns under-trained once initial behaviors are aligned. To address these challenges, we propose TurnOPD, a turn-level budgeting strategy for efficient on-policy distillation of long-horizon agents. TurnOPD consists of two budget controllers: adaptive rollout-depth budgeting, which uses probe-based turn statistics to determine rollout length, and progressive turn-normalized loss budgeting, which gradually shifts KL weighting from token-level to turn-balanced supervision. Experiments on ALFWorld, WebShop, and Multi-Hop Search with task-specialized teacher models show that TurnOPD achieves superior validation accuracy under equal wall-clock training budgets and advances the accuracy--time frontier beyond vanilla OPD.

Reinforcement Learning from Human Feedback (RLHF) for Large Language Models increasingly relies on critic-free methods as a practical alternative to actor--critic training. Despite their simplicity, existing critic-free approaches propagate a trajectory-level learning signal uniformly across all tokens in a trajectory. This requires full-trajectory policy updates for every rollout, leading to substantial optimization cost for long reasoning traces, even though intermediate prefixes often contain enough information to largely determine the final outcome. We propose Prefix-Sampling Proximal Policy Optimization (PS-PPO), a compute-efficient critic-free method for RLHF that exploits this temporal redundancy. PS-PPO introduces a prompt-conditioned cutoff distribution and samples a cutoff timestep for each trajectory. During the update pass, PS-PPO backpropagates only through the sampled prefix of each trajectory and applies an importance-weighting correction so that the resulting truncated gradient estimator remains unbiased with respect to the full-trajectory objective. Experiments on mathematical reasoning and RLHF benchmarks show that PS-PPO achieves large reductions in training compute and peak GPU memory, while maintaining accuracy comparable to strong critic-free baselines.

Despite the rapid advance of Multimodal Large Language Models (MLLMs) in high-level video understanding, a fundamental bottleneck remains: these models collapse complex human motion into coarse semantic labels. Existing benchmarks mostly focus on scene-centric events or local joint articulations, failing to probe global human motion in space over time (trajectory and orientation changes). We introduce HumanMoveVQA, the first comprehensive benchmark designed to evaluate global trajectory and orientation reasoning from an exocentric perspective. Our benchmark utilizes a first-frame anchored world coordinate system, preserving translation and rotation relative to a fixed starting point. We propose a scalable, multi-stage pipeline that lifts 2D video observations into world-consistent 3D motion tracks to generate over 10K structured question-answer pairs across seven reasoning categories, including motion aggregation, sequential ordering, and trajectory-level inference. Our extensive evaluation reveals a critical capability gap in state-of-the-art proprietary models on deep human motion understanding. However, we demonstrate that this is a learnable problem; by fine-tuning an open-source baseline with our targeted, world-consistent supervision, we achieve a significant improvement. HumanMoveVQA establishes a rigorous geometric foundation for developing next-generation, movement-aware video understanding models.

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.

Recently, Muon has gained substantial attention as an appealing alternative to Adam-like optimizers, with many works highlighting its advantages through spectral normalization and improved conditioning. Yet this positive theoretical narrative contrasts with its empirical performance in large language model (LLM) training, where Muon's gains over Adam/AdamW are often mixed, schedule-sensitive, and not uniformly superior. To address this gap, we develop a trajectory-level theory characterizing both the strengths and limitations of Muon. We introduce a mixed-spiked matrix sensing model whose sensing operator decomposes into signal, spike, and bulk components, capturing a mixture of anisotropic structure and long-tail information reminiscent of LLM training. On top of it, we adopted a river-valley perspective in which we view the landscape as composed of a river direction flowing to the desired solution and hill directions encoding nuisance or task-irrelevant information. In the momentum-free setting, we show that Muon moves faster along the information-bearing river direction during early optimization, but can converge much more slowly near the river bottom than gradient descent. We then extend the river-valley perspective to general nonconvex objectives with momentum by studying points on the spectral river. There, while Muon converges faster early on, its orthogonalized update removes residual scale information, making it prone to overshooting and oscillation near the target solution. Together, these results suggest that our characterizations extend beyond spiked matrix sensing and motivate switching to GD-like refinement optimizers in the final phase, rather than relying only on a fixed learning-rate schedule for Muon. We also provide preliminary evidence supporting this two-stage approach in language model training experiments.

Autonomous driving has shifted towards end-to-end policy learning, where reliable, interpretable policy evaluation is a fundamental challenge as driving quality is highly context-dependent. Commonly used rule-based driving metrics like EPDMS are interpretable but lack context-awareness, while recent VLMbased evaluations are context-aware but limited by ambiguous VLM outputs and weak physical grounding. To evaluate driving in a manner that is both interpretable and context-aware, we introduce DriveJudge. DriveJudge is a driving evaluation agent that combines rule-grounded evaluation with Vision-Language Model (VLM) reasoning and selectively invokes physically-grounded deterministic rule functions after interpreting the environmental context. To train and evaluate DriveJudge, we curate a large-scale dataset of 33,577 challenging driving samples with human annotations on whether the driving behavior is reasonable in the given scenario. With this dataset, we address the underexplored problem of driving metric evaluation, and introduce two human-aligned benchmark tasks: Driving Quality Classification and Trajectory Preference Selection. DriveJudge outperforms EPDMS for driving quality classification by 21.23 AUC, and the recent VLM-based DriveCritic for trajectory preference selection by 6.5%, setting a new standard for interpretable and precise driving evaluation.

Safety certification of Vision-Language-Action (VLA) driving planners under ISO 21448 (SOTIF) rests on an Operational Design Domain (ODD) specification that answers two complementary questions: when does the planner start to fail, and how severely does it fail once it does? We evaluate Alpamayo R1, a 10B-parameter open-weight driving VLA, on 15,968 (clip, attack) pairs. We find a conservative-aggregate gap: an aggregate safe threshold of $σ\leq 50$ under a 15% average displacement error (ADE) budget masks well-sampled scenarios that tolerate the top of the tested grid ($σ= 70$). A Gaussian Mixture Model (GMM) on the changed-explanation subset identifies six discrete severity bands (BIC-optimal $k{=}6$), so two perturbation conditions with the same mean error can differ materially in their share of high-severity (C4/C5) failures. Joining the two analyses on the same corpus surfaces a finding neither yields in isolation: the scenarios with the loosest noise thresholds are not those with the lowest high-severity rate: STOP_SIGNAL concentrates roughly $4\times$ the C4/C5 share of LANE_KEEPING despite tolerating a larger $σ$. A deployable SOTIF ODD specification for driving VLAs therefore requires a two-dimensional safety envelope, not a single aggregate value per hazard.

Vision-language-action (VLA) policies bring natural language into closed-loop robot control, enabling robots to execute manipulation tasks directly from text instructions. The same interface gives text a recurring role in control because the prompt is reused at every replanning step, and each prompt-conditioned action changes the future observations on which the policy acts. Existing VLA attacks study adversarial prompts that elicit targeted low-level actions or make such actions persist across changing images. We identify a stronger trajectory-level failure mode: a prompt that still $\textit{appears}$ to specify the intended task but redirects the final physical outcome. We mathematically formalize this setting as $\textit{command-preserving trajectory redirection}$, a prompt-only threat model in which the attacker chooses one prompt before the episode, all policy and environment components remain fixed, and the prompt must stay close to the benign instruction while omitting target words and correction language. To find such prompts, we introduce an on-policy prompt search method that uses rollouts to discover perturbations whose closed-loop behavior tracks a target task while satisfying the command-preserving constraints. Experiments in simulation and on hardware show that near-benign prompt perturbations can redirect VLA rollouts to attacker-specified targets. These results expose a trajectory-level vulnerability in VLA instruction grounding: text that appears to preserve the intended command can still give an adversary control over the robot's final physical outcome. Project website: https://vla-redirection-attack.github.io/

Preference-based reinforcement learning (PbRL) learns policies from human trajectory-level comparisons, avoiding explicit reward design and expert demonstrations. Existing methods typically train utility functions on trajectory or segment-level preferences while relying on per-step utility estimates during policy optimization. This training and inference mismatch induces a distribution shift that severely degrades temporal credit assignment and limits policy learning. We analyze this issue and propose PAWS, a segment-based preference learning method that performs policy updates directly using segment-level advantage functions. By aligning utility training with policy optimization, PAWS preserves trajectory-level preference information and avoids unreliable per-step learning signals. Experiments on simulated robotic manipulation and locomotion tasks demonstrate that PAWS consistently outperforms existing PbRL approaches, highlighting the importance of distribution-consistent preference learning.

Video Text Spotting (VTS) is essential for urban surveillance and intelligent transportation systems, enabling automated reading of street signs, vehicle markings, and scene text in video streams. However, reliable recognition remains challenging due to dynamic video factors common in surveillance scenarios, including motion blur, occlusion, and scale variation, which degrade frame-level recognition. Existing VTS methods typically perform recognition independently on each frame, leading to inconsistent and inaccurate results across sequences. To address these limitations, we propose TraRA (Trajectory-level Recognition Aggregation for VTS), a plug-and-play method that performs trajectory-level text recognition by leveraging temporal and multimodal consistency. TraRA integrates two key modules: (1) the Temporal Clustering and (2) the Vision-Language Aggregation. The former refines noisy trajectories by grouping temporally and visually coherent text instances, while the latter employs a Low-Rank Adaptation-enhanced Vision-Language model to fuse visual cues with linguistic context across frames. By aggregating information over entire text trajectories, TraRA achieves robust text recognition even under challenging surveillance conditions. Extensive experiments on four public benchmarks, including road and urban scene datasets (RoadText, BOVText, ArTVideo, and ICDAR15), demonstrate that TraRA consistently improves tracking and recognition performance over state-of-the-art VTS methods. The source code is available at https://github.com/trid2912/TraRA.

Autonomous LLM agents can pursue hidden malicious objectives through sequences of individually benign actions, making sabotage difficult to detect using standard trajectory-level monitoring. Existing approaches either evaluate complete trajectories in a single pass or partition them into independently scored windows, limiting their ability to connect evidence across temporally distant actions. We propose TRACE, a monitoring framework for long-horizon LLM agent trajectories. TRACE operates through a TIJ (Triage-Inspect-Judge) loop that identifies high-signal regions, performs targeted inspection while maintaining accumulated evidence across reasoning steps, and synthesizes a trajectory-level verdict. We evaluate TRACE on ten task domains from SHADE-Arena against state-of-the-art baselines. TRACE achieves an aggregate F1 of 0.713 and recall of 0.844, with the largest gains on tasks requiring long-range evidence linking.

Reinforcement Learning (RL) has become a promising approach for improving GUI Agents in long-horizon, stochastic digital environments, but trajectory-level success feedback is too sparse to provide reliable credit assignment for intermediate exploration steps. To mitigate this issue, recent studies introduce Process Reward Models (PRMs), which provide finer-grained training feedback through global milestone verification or local step-level evaluation. However, these methods still suffer from two level-specific limitations: global milestone decomposition is subjective and singular, making it difficult to accommodate the multiple valid execution paths in real GUI tasks, while fixed local judging windows may miss long-range key evidence or dilute the decision signal with irrelevant frames. Inspired by stain-tracing mechanisms in network flow analysis, we propose StainFlow, an entity-stain-flow process reward model for GUI Agents. To reduce the subjectivity of global partitioning, we introduce the Global Entity Stain Tracking module, which extracts visually verifiable task entities and tracks how their stain concentrations and states evolve along the trajectory, allowing task phases to be objectively separated by changes in the entity evidence flow. To improve the accuracy of local verification, we introduce the Local Stain Evidence Linking module. Centered on the triggering entities of each candidate key node, it retrieves relevant steps based on their stain concentrations and state changes, and dynamically constructs high-density evidence windows for verifying true key nodes. Extensive experiments on AndroidWorld and OGRBench show that StainFlow relatively improves online RL success by 3.2% and trajectory completion judgment accuracy by 1.8%.

AI assistants in human-AI collaboration often correct suboptimal human actions through behavioral feedback (e.g., alerts or steering-wheel nudges in assistive driving). Such interventions can mitigate immediate errors, but long-term improvement requires addressing the underlying misconceptions that cause repeated mistakes. We introduce SENSEI, a framework that infers user misconceptions from interaction behavior and provides targeted, minimal yet sufficient suggestions to correct them. Our approach departs from action- or trajectory-level interventions by operating over a structured knowledge representation to localize and correct the sources of erroneous behavior. Across three long-horizon tasks with diverse misconceptions and corresponding behaviors, SENSEI demonstrates zero-shot compositional generalization, disentangling multiple overlapping misconceptions despite training only on single-misconception cases. A user study further shows that our method identifies real human misconceptions and provides effective guidance that improves long-horizon task performance, successfully correcting $90\%$ of student misconceptions. Code and project page are available at https://misoshiruseijin.github.io/SENSEI/.

Abundant procedural knowledge on the Web holds great potential for helping agents solve long-horizon tasks. However, such knowledge is often multimodal, heterogeneous, noisy, and implicitly assumes human executors, making it difficult to use directly as the skills required by agents. To bridge the gap between human-oriented guides and agent-executable skills, we formalize this problem as guide-to-skill learning: converting in-the-wild guides into executable skills and continuously improving them from trajectories observable to the agent. To evaluate the capability of existing agents on this task, we introduce MMG2Skill-Bench, the first benchmark designed for this problem. We further propose MMG2Skill, a closed-loop framework that compiles guides into editable skills, conditions a fixed vision-language model (VLM) agent on these skills during execution, and revises the skills from trajectory-level root-cause feedback without using benchmark scores. Across GUI control, open-ended gameplay, and strategic card play with six VLM backbones, MMG2Skill consistently outperforms vanilla baseline agents in every model-domain setting, achieving macro-average gains of +12.8 to +25.3 percentage points across backbones. Ablation studies show that directly prompting agents with raw guides can degrade performance, while both structured skill construction and trajectory-driven revision are necessary for the observed improvements. On success-inferable tasks, analyzer-based early stopping further prevents late-stage performance regressions and saves 25%-53% of attempts when the success signal is properly calibrated.

In Agentic Search, trajectory-level outcome rewards fail to quantify the behavioral contributions of individual steps, while existing step-level reward methods typically rely on costly tree sampling. We view world knowledge as a latent world graph and each IS task as search within a latent task graph, where effective steps should make graph progress toward the answer node. Based on this prior, we propose Graph-Distance Contribution Reward (GDCR), a step-level process reward that scores newly-retrieved and newly-cited entities by their distance to the answer node in a training-time Entity-Relation (ER) graph. We further propose Step Advantage Policy Optimization (SAPO), which converts GDCR into step-level advantages and combines them with trajectory-level outcome advantages. Experiments on four challenging benchmarks validate the effectiveness of our method.

Mastering terminal environments requires language agents capable of multi-step planning, feedback-grounded execution, and dynamic state adaptation. However, training such agents is currently bottlenecked by a reliance on scraped external repositories, which limits domain diversity, environment controllability, and the targeting of specific capability deficits. We introduce LiteCoder-Terminal-Gen, a zero-dependency synthesis pipeline that autonomously generates executable and verifiable terminal training environments directly from domain specifications. Using this framework, we construct two large-scale resources: LiteCoder-Terminal-SFT, comprising 11,255 expert trajectories across 10 domains, and LiteCoder-Terminal-RL, featuring 602 verifiable environments for trajectory-level preference optimization. Supervised fine-tuning of Qwen-family models on our SFT dataset yields agents that significantly outperform their base counterparts. Notably, our 32B variant achieves 29.06%, 18.54%, and 34.00% pass@1 on Terminal Bench 1.0, 2.0, and Pro, respectively. Furthermore, applying Direct Multi-turn Preference Optimization (DMPO) on our RL environments yields additional performance gains. These results systematically demonstrate that fully synthetic, executable environments offer a scalable and verifiable supervision signal for mastering complex, real-world command-line workflows.

Reflexion-style agents rely on self-generated reflections as memory, implicitly assuming that agents can accurately diagnose their own failures. We show that this assumption can fail systematically: across ALFWorld and HumanEval, agents store confident but incorrect interpretations of the task and continue acting on them across trials, even though the environment resets to the correct task each time. We call this failure mode memory confabulation and introduce the Reflection Repetition Rate (RRR), a log-based metric that detects repeated reliance on incorrect reflective content. Using RRR, we identify 16 frozen environments in ALFWorld, where 0 of 121 reflections mention the correct target object, and 4 analogous cases in HumanEval. Our mitigation replaces open-ended self-diagnosis with programmatic extraction of trajectory-level failure signals, increasing correct object mention from 0% to 86%, reducing RRR from 0.64 to 0.10, and solving 3 of 16 frozen ALFWorld environments, suggesting that reflective memory can reinforce false beliefs rather than correct them.

Humans know when to reach for help e.g. $347 \times 28$ warrants a calculator while $2+2$ does not. Language models do not. Prompt-based approaches can instruct a model when to invoke tools, but this scaffolding does not teach it to recognize the boundary of its own knowledge. RL approaches that assign a single outcome reward to the whole trajectory fare no better: trajectory-level credit cannot isolate which tool call in a successful episode actually helped, nor penalize unnecessary calls. We propose \textbf{CARL} (\textbf{C}ompetence-\textbf{A}ware \textbf{R}einforcement \textbf{L}earning), which trains a critic on the model's own rollouts to learn where parametric knowledge suffices and where it needs external help. By decomposing each rollout at natural tool-use boundaries (e.g., code fence delimiters and context block transitions), CARL assigns independent credit to each segment from a single binary outcome, without external judges or step-level annotations. As a result, erroneous tool calls, incorrect extractions, and unnecessary calls each receive appropriately signed advantages. The trained critic captures the model's domain competence: it separates parametrically solvable from tool-dependent questions with AUC 0.93 at 7B. On five benchmarks spanning arithmetic, multi-hop factual QA, and numerical reasoning over financial tables, CARL improves exact-match accuracy by 6.7 points at 7B and 9.7 points at 3B over the best RL baseline, with the largest gain (+8.3 EM at 7B, +9.0 EM at 3B) on Musique. The model issues 53% fewer tool calls on parametrically answerable questions while remaining ${\sim}10$ EM points more accurate on them. Gains are largest at small scale: the 3B improvement is $1.4\times$ the 7B improvement, suggesting that knowing when to ask disproportionately benefits models with smaller parametric memory.

Autonomous driving stacks must pick one trajectory from a multi-modal candidate set; choosing by model confidence ignores safety, traffic-law, and comfort constraints. We present \textsc{RECTOR} (Rule-Enforced Constrained Trajectory Orchestrator), a post-generation reranking layer that scores candidates against a tiered rulebook (Safety~$\succ$Legal$\succ$Road$\succ$Comfort) via differentiable proxies and a scene-conditioned applicability mechanism, then selects with a deterministic $\varepsilon$-lexicographic rule that preserves cross-tier priority by construction -- without retraining the predictor. On the Waymo Open Motion Dataset \texttt{validation_interactive} split (43{,}219 augmented instances, $K{=}6$), under ProtocolB (28-rule proxy catalog, oracle applicability) rule-aware selection cuts Safety+Legal violations from 28.58% to 20.42% and Total from 40.32% to 32.41% versus confidence-only on the same candidates. A uniform-weight weighted-sum baseline matches binary compliance on this benchmark -- the empirical lift comes from rule-aware ranking, while the lexicographic guarantee is the structural differentiator no weight calibration can replicate. Under adversarial confidence corruption, confidence-only selection fails in 100% of scenarios while both rule-aware selectors reject the injected mode in $\sim$96%. All figures are proxy-evaluator results (not a safety certificate), open-loop, 5,s horizon, U.S.\ rules, validation split.

Long-horizon agentic reasoning requires large language models to act over long interaction histories containing thoughts, tool calls, observations, and partial conclusions. The challenge is not merely that these histories grow long, but that information needed for the current decision may be scattered across distant steps and only become relevant later. Existing approaches address this difficulty by truncating the interaction history, compressing it into shorter surrogates, or retrieving selected parts of it for reuse, but they do not explicitly model how access to past interaction should adapt to the agent's evolving state. We instead cast long-horizon reasoning as a problem of state-adaptive memory. To this end, we propose State-Adaptive Memory~(SAM), a standalone framework that consolidates ongoing interaction into compact memory cues while preserving raw trajectory pages for intent-driven recall. These cues are not treated as replacements for history; rather, they serve as lightweight handles that allow the agent to reconstruct temporally distant information according to its current needs, without retraining the underlying backbone. We further optimize the memory module through expert-guided supervision and reinforcement learning, aligning it with trajectory-level utility. Across BrowseComp, BrowseComp-ZH, WideSearch, and HLE, SAM consistently outperforms strong baselines over diverse agent backbones. Our results suggest that explicit memory modeling provides a simple and effective foundation for long-horizon agentic reasoning.

Reinforcement learning with verifiable rewards has become the standard recipe for improving LLM reasoning, but the dominant algorithm GRPO assigns a single trajectory-level advantage to every token, diluting the signal at pivotal reasoning steps and injecting noise at uninformative ones. Critic-free alternatives derived from on-policy distillation supply per-token signals through oracle-conditioned likelihood ratios, yet apply each signal in isolation from the trajectory-level evidence accumulated up to that position. We propose Oracle-Prompted Policy Optimization (OPPO), which rests on a single observation: the oracle signal used by prior distillation-style methods for local discrimination is also the natural Bayesian update of the model's belief about eventual success. Accumulating the signal along a trajectory yields, in closed form and at the cost of one extra forward pass, a running estimate of the success probability at every position, together with a token-level advantage that requires no learned value network and no additional rollouts. A first-order analysis factorizes the advantage into the per-token discrimination signal used by distillation methods modulated by a state weight that concentrates credit on genuinely pivotal tokens, with a directional variance-reduction guarantee. The framework admits two estimators differing only in which model scores the evidence: a \textit{self-oracle} that reuses the student and recovers the on-policy distillation reward as a strict special case, and a \textit{teacher-oracle} that delegates scoring to a stronger frozen model. On two base LLMs across seven mathematics, science, and code reasoning benchmarks, OPPO improves over GRPO, DAPO, and SDPO by up to $+6.0$ points on AMC'23 and $+5.2$ points on AIME'24, with gains that widen monotonically with response length.

Driving Vision-Language-Action Models (Driving VLAs) commonly introduce natural-language reasoning as an intermediate interface for end-to-end planning, but reasoning-centric interfaces face three practical bottlenecks: obtaining high-quality reasoning annotations is difficult, generating and understanding long reasoning chains is challenging for compact models, and inference latency is substantially increased. In this paper, we rethink the design of language interfaces in Driving VLAs and show that concise one-step meta-actions are a simple yet effective alternative to verbose reasoning. Meta-actions provide semantic decision grounding while remaining low-entropy, and being automatically derivable from expert trajectories, enabling scalable supervision and reliable trajectory conditioning. Building on this interface, we propose DriveMA, which combines action-centric supervised training with a turn-level credit-assignment reinforcement learning framework that jointly optimizes meta-action correctness, trajectory quality, and trajectory--meta-action consistency. Experiments show that DriveMA already achieves a new state of the art on the Waymo End-to-End Driving Challenge with a 2B model, reaching a Rater Feedback Score (RFS) of 8.060, while its 4B version further improves the state of the art to 8.079; DriveMA also obtains competitive performance on NAVSIM. Ablations demonstrate that one-step meta-actions offer a better practical trade-off between expressiveness, predictability, and inference efficiency than natural-language reasoning or finer-grained action sequences. Code, data, and models will be released to facilitate future research.

Vision-Language-Action (VLA) models show promising ability in language-guided robotic tasks. However, making VLA policies reliable remains challenging, because a manipulation task is completed through closed-loop interaction, where each action affects subsequent execution. To analyze this problem, we revisit VLA policy during execution and argue that a VLA policy acts both as a planner, which makes task-oriented decisions that change the direction of execution, and as an executor, which realizes these decisions through dense continuous actions. This view suggests that improving VLA reliability requires particular attention to planning actions. Existing optimization methods can imitate actions or improve complete trajectories, but they usually do not explicitly identify planning actions or measure their importance for task success. To address this issue, we propose Planning-Aware Policy Optimization for VLA models (PAPO-VLA). PAPO-VLA first identifies planning actions by jointly considering action variation and trajectory outcome, then estimates their importance through causal sufficiency and causal necessity, and finally incorporates this importance into GRPO advantage estimation. In this way, more important planning actions receive stronger optimization emphasis, while the whole trajectory is still optimized by trajectory-level feedback. Experiments on multiple benchmarks demonstrate the effectiveness of PAPO-VLA.

Large language model-driven multi-agent systems (LLM-MAS) excel at complex tasks, yet unreliable agents remain a key bottleneck to system-level reliability. Automatic failure attribution is therefore critical, but existing approaches, such as direct prediction of agent-error pairs and agent-first failure attribution, rely on local logs of agents and miss global failures that only manifest over full interaction trajectories, such as cross-step inconsistencies and inter-agent coordination errors. Moreover, directly predicting failures induces a large combinatorial search space, hindering fine-grained attribution. To address these challenges, we propose VerifyMAS, a hypothesis verification framework for agent failure attribution. Instead of directly predicting faulty agents and error types, VerifyMAS formulates and verifies failure hypotheses against full trajectories. This verification-based approach decomposes attribution into trajectory-level error validation and fine-grained agent localization, providing an error-first attribution approach that captures global failure patterns while substantially reducing the search space. We further introduce a hypothesis-based data construction strategy grounded in a structured error taxonomy and fine-tune a specialized LLM verifier model for trajectory-level failure verification and agent attribution. Experiments on Aegis-Bench and Who&When show that VerifyMAS consistently improves diverse backbone models, including open-source Qwen and API-based GPT models, outperforming prior methods without sacrificing inference efficiency for long multi-agent trajectories.

LLMs are increasingly deployed in autonomous laboratories, under the assumption that their domain priors and reasoning over iterative feedback let them converge on good designs in fewer iterations than feedback-only baselines. Current iterative scientific design benchmarks, however, score only outcome snapshots at fixed horizons. This leaves the learning trajectory unmeasured, even though the trajectory is what captures learning efficiency, where each iteration saved is a real saving in cost and time. Motivated by this, we examine three evaluation choices that change the conclusions one draws about LLM learning efficiency in iterative scientific design: what to measure, what baseline to compare against, and what to ground against. We introduce LEAPBench, Learning Efficiency in Adaptive Processes, a 55-task framework that pairs a best-so-far area under the curve (AUC) trajectory metric with a classical Bayesian-optimization reference and an audit grounded in published literature. Applied to eight contemporary LLMs, switching from final-outcome to trajectory scoring changes the best-model decision on 53% of tasks at matched horizons, and exposes efficiency gains overlooked by outcome-based scoring. LLMs do not outperform a classical Bayesian baseline. On 16 biology tasks where the oracle's reward signal is aligned with configurations from the published-best design, domain-aware prompting leads to LLM choices that match the published-best's approximately 10 percentage points less often than domain-agnostic prompting at iteration 30. The pattern is sharpest on 6 tasks where the literature-typical and published-best configurations diverge, and domain-agnostic prompting matches the published-best more often on all 6. The trajectory metric also doubles as a tractable training target. Offline reinforcement learning with the metric as a reward improves performance on 14 of 21 held-out tasks.

Recursive learning -- where models are trained on data generated by previous versions of themselves -- is increasingly common in large language models, autonomous agents, and self-supervised systems. However, standard performance metrics (loss, perplexity, accuracy) often fail to detect internal degradation before it becomes irreversible. Here we identify a phenomenon we call silent collapse: under broad recursive conditions, model internal distributions -- predictive entropy, representational diversity, and tail coverage -- progressively contract even as conventional metrics appear stable or improving. We discover that silent collapse is not abrupt. Its onset is reliably preceded by three trajectory-level precursors: (1) contraction of anchor entropy, (2) freezing of representation drift, and (3) erosion of tail coverage. These signals manifest multiple generations before any degradation in standard validation metrics, enabling early warning. Based on these precursors, we propose the MTR (Monitor--Trust--Regulator) framework, a lightweight metacognitive loop that monitors trajectory statistics, estimates a slow-timescale trust variable, and adaptively modulates the effective learning intensity. MTR provides early warning and actively prevents silent collapse without requiring access to pristine real data -- a critical advantage when original data is unavailable, contaminated, or private.

Group Relative Policy Optimisation (GRPO) enhances large language models by estimating advantages across a group of sampled trajectories. However, mapping these trajectory-level advantages to policy updates requires aggregating token-level probabilities within each sequence. Relying on a fixed aggregation mechanism for this step fundamentally limits the algorithm's adaptability. Empirically, we observe a critical trade-off: certain fixed aggregations frequently suffer from training collapse, while others fail to yield satisfactory performance. To resolve this, we propose \textbf{HölderPO}, a generalised policy optimisation framework unifying token-level probability aggregation via the Hölder mean. By explicitly modulating the parameter $p$, our framework provides continuous control over the trade-off between gradient concentration and variance bounds. Theoretically, we prove that a larger $p$ concentrates the gradient to amplify sparse learning signals, whereas a smaller $p$ strictly bounds gradient variance. Because no static configuration can universally resolve this concentration-stability trade-off, we instantiate the framework with a dynamic annealing algorithm that progressively schedules $p$ across the training lifecycle. Extensive evaluations demonstrate superior stability and convergence over existing baselines. Specifically, our approach achieves a state-of-the-art average accuracy of $54.9\%$ across multiple mathematical benchmarks, yielding a substantial $7.2\%$ relative gain over standard GRPO and secures an exceptional $93.8\%$ success rate on ALFWorld.

Agentic memory enables LLMs to persist information beyond a single context window and reuse it in later decisions, but it also introduces a new vulnerability: spurious correlations, where retrieved memory carries miscorrelated evidence and propagates erroneous reasoning into downstream decisions. Despite the widespread use of agentic memory, this risk remains largely underexplored. We address it from two aspects. First, we benchmark several canonical types of spurious patterns identified through causal structure and record them across trajectory-level memory. Diagnosing agentic memory systems on this benchmark reveals that memory improves reasoning on clean inputs but amplifies reliance on spurious patterns when they are present. Second, we propose CAMEL, a plug-and-play calibration method that operates across diverse memory architectures at both write and retrieval time. CAMEL consistently reduces reliance on spurious patterns across all three types while preserving or improving performance on clean inputs and staying robust under adaptive attacks targeting the calibration. Overall, CAMEL offers a principled and lightweight solution toward more reliable agentic memory deployment.

LLM-based multi-agent systems are increasingly deployed on long-horizon tasks, but a single decisive error is often accepted by downstream agents and cascades into trajectory-level failure. Existing work frames this as \emph{post-hoc failure attribution}, diagnosing the responsible agent and step after the trajectory has ended. However, this paradigm forfeits any opportunity to intervene while trajectory is still unfolding. In this work, we introduce AgentForesight, a framework that reframes this problem as online auditing: at each step of an unfolding trajectory, an auditor observes only the current prefix and must either continue the run or alarm at the earliest decisive error, without access to future steps. To this end, we curate AFTraj-2K, a corpus of agentic trajectories across Coding, Math, and Agentic domains, in which safe trajectories are retained under a strict curation pipeline and unsafe trajectories are annotated at the step of their decisive error via consensus among multiple LLM judges. Built on that, we develop AgentForesight-7B, a compact online auditor trained with a coarse-to-fine reinforcement learning recipe that first equips it with a risk-anticipation prior at the failure boundary on adjacent safe/unsafe prefix pairs, then sharpens this prior into precise step-level localization under a three-axis reward jointly targeting the what, where, and who of an audit verdict. Across AFTraj-2K and an external Who&When benchmark, AgentForesight-7B outperforms leading proprietary models, including GPT-4.1 and DeepSeek-V4-Pro, achieving up to +19.9% performance gain and 3$\times$ lower step localization error, opening the loop from post-hoc failures detection to enabling deployment-time intervention. Project page: https://zbox1005.github.io/agent-foresight/

Distributed collaborative intelligence (DCI), encompassing edge-to-edge architectures, federated learning, transfer learning, and swarm systems, creates environments in which emergent risk is structurally unavoidable: locally correct decisions by individual agents compose into globally unacceptable behavioral trajectories under uncertainty. Existing approaches such as constrained optimization, safe reinforcement learning, and runtime assurance evaluate acceptability at the level of individual actions rather than across behavioral trajectories, and none addresses the multi-participant, uncertainty-laden nature of DCI deployments. This paper introduces mechanical conscience (MC), a novel concept and simplified mathematical framework that operationalizes trajectory-level normative regulation for both single-agent and distributed intelligent systems. Mechanical conscience is defined as a supervisory filter that minimally corrects a baseline policy's actions to reduce cumulative deviation from a normatively admissible region, while accounting for epistemic uncertainty. We introduce associated constructs, conscience score, mechanical guilt, and resonant dependability, that provide an interpretable vocabulary and computable governance signals for this emerging field. Core theoretical properties are established: admissibility equivalence, existence of optimal regulation, and monotonic deviation reduction. Illustrative results demonstrate that MC-regulated agents maintain trajectory-level normative acceptability where conventional controllers drift outside admissible bounds, and that the framework naturally extends to suppress interaction-induced emergent risk in multi-agent DCI settings.

Reinforcement learning systems rely on environment interfaces that specify observations and reward functions, yet constructing these interfaces for new tasks often requires substantial manual effort. While recent work has automated reward design using large language models (LLMs), these approaches assume fixed observations and do not address the broader challenge of synthesizing complete task interfaces. We study RL task interface discovery from raw simulator state, where both observation mappings and reward functions must be generated. We propose LIMEN (Code available at https://github.com/Lossfunk/LIMEN), a LLM guided evolutionary framework that produces candidate interfaces as executable programs and iteratively refines them using policy training feedback. Across novel discrete gridworld tasks and continuous control domains spanning locomotion and manipulation, joint evolution of observations and rewards discovers effective interfaces given only a trajectory-level success metric, while optimizing either component alone fails on at least one domain. These results demonstrate that automatic construction of RL interfaces from raw state can substantially reduce manual engineering and that observation and reward components often benefit from co-design, as single-component optimization fails catastrophically on at least one domain in our evaluation suite.

Large language model (LLM) agents have shown strong capabilities in long-horizon reasoning, tool use, and decision-making in digital environments, yet extending them to physically grounded systems remains challenging. Unlike web, code, or game environments, where objectives are often weakly coupled, physical systems evolve through tightly coupled dynamics in which local interventions propagate across interacting subsystems over time. Urban traffic control exemplifies this challenge, as traffic signals, freeways, public transit, and taxi systems continuously interact through shared spatial infrastructure and temporal mobility demand. Existing optimization, reinforcement learning (RL), and LLM-based approaches are largely designed for isolated subsystems, limiting coordinated reasoning and system-level optimization. We propose TrafficClaw, a LLM-based generalizable traffic control agent for physical urban systems. TrafficClaw operates within a unified traffic environment that exposes coupled urban dynamics and feedback, performs executable spatiotemporal reasoning with persistent memory for long-horizon adaptation, and leverages multi-stage agentic RL for coordinated system-level optimization. Experiments across three metropolitan regions and six traffic-control tasks demonstrate strong generalization, robustness, and cross-subsystem coordination. Our project is available at https://github.com/usail-hkust/TrafficClaw.