autonomous system developers

Abstract: Reinforcement learning agents optimize exactly the reward signal they are given, which can diverge sharply from an intended true objective when that signal is only a proxy for it; this failure mode is known as specification gaming. Existing evidence for this failure mode is largely observational, with comparatively few demonstrations pairing a controlled, deliberately constructed proxy/true gap with a mitigation that carries a verifiable guarantee rather than only an empirical improvement. We construct a minimal gridworld in which a cleaning-robot agent can either genuinely remove dirt or conceal it from a cheap sensor, and show that a standard deep Q-network trained only on the sensor's proxy reward converges to a policy that games it completely, reaching near-maximal proxy reward while never achieving the true objective. We then introduce a constrained reinforcement learning fix, penalizing the specific behaviors through which the agent exploits the gap, and show that it eliminates the gaming behavior from the final policy while substantially recovering true-objective performance. We further derive and empirically verify an exact accounting identity relating the proxy reward, true reward, and our cost signal for this environment, showing that the cost signal measures the proxy/true gap exactly rather than merely tracking it, and yields a deterministic guarantee under a hard-constrained variant of the method.

Abstract: Small observation or action errors can compound during closed-loop execution and disrupt a trained imitation learning policy. We study whether the original expert demonstrations can improve robustness through test-time hidden-state steering without additional demonstrations or policy retraining. Our analysis of a trained policy finds that expert hidden-state variation is concentrated in relatively few dimensions, while the directions capturing this variation differ across regions. Motivated by this geometry, we propose Diffeomorphic Hidden-State Steering (DHS), which learns an invertible nonlinear coordinate map from expert hidden states obtained by replaying the demonstrations through the frozen policy. DHS attenuates coordinates with low expert variance more strongly than those with high variance, then maps the modified coordinates back to hidden-state space. This allows steering directions to vary with the current hidden state while retaining every coordinate. Across four simulated manipulation tasks, DHS improves mean success under observation noise over the unmodified Transformer policies. Controlled comparisons yield higher mean success with learned nonlinear coordinates than with linear coordinates under both hard projection and full-rank rescaling. Combining DHS with action smoothing improves mean success rate under observation noise in certain tasks compared with action smoothing alone.

Abstract: Pretrained vision-language-action models (VLAs) often fail midway through unseen long-horizon tasks. Recursive self-improvement (RSI) offers a remedy, but embodied RSI methods either refine the harness around a frozen executor and so cannot add a missing motion, or train the executor on persistent failures without asking whether a failure lies in how the executor is called or in what it can do. We ask when a self-improving embodied system should change the weights of its executor, and propose AdaRSI, an on-demand RSI framework for embodied agents. A coding agent first improves its skills and harness, then attributes persistent failures by retrying failing sub-steps from saved states, and fine-tunes only the sub-steps that remain unreliable, with demonstrations it collects itself and experience carried over from earlier tasks. On all 50 target tasks of RoboCasa365, with GR00T N1.5 as the executor and no human demonstrations, AdaRSI improves the VLA on all three suites and raises its success on unseen composite tasks from 4.6% to 13.8%. Compared with fine-tuning every sub-step, it reaches slightly higher success while using only 28% of the GPU time and 22% of the self-collected demonstrations, and fine-tuning without a harness trains more yet gains far less.

Abstract: Technologies, such as co-simulation, in particular, digital twins (DTs) are used for comprehensive and reliable prototyping, verification and validation (V&V) of control systems. There, the combination of feedback loops across multiple realities (e.g., simulations, physical setups) enables various forms of experimentation with robots and autonomous systems. However, the handling of multiple realities requires a more specific support by the modelling techniques used in control engineering. To streamline the latter in the system life cycle, we propose a framework for mixing (i.e., fusing, defusing) control input and feedback signals (i.e., stimuli, responses) in simultaneously operated plants for multiple realities.

Abstract: Executing a vision-language-action (VLA) policy involves deciding when to continue an attempt and when to intervene to improve its chances of success. We introduce ACRO, an Actor--Critic rollout orchestration layer that makes these decisions for a fixed VLA. ACRO seeks interventions that offer a more promising continuation, are physically feasible, and leave sufficient time to complete the task. A Critic learned from policy rollouts estimates success prospects to decide when to intervene and select a re-entry target along the executed trajectory. Trajectory Retraction shortcuts the visited path to return the robot to that target, where the same VLA resumes execution from a fresh observation. We evaluate ACRO with GR00T N1.7 and $π_{0.5}$ across SimplerEnv, RoboCasa, and real-robot manipulation. Under the same physical execution budget, including all return motion, ACRO improves success by 10.4, 7.3, and 21.7 percentage points, respectively. It also completes additional tasks as the execution budget grows. These results show that controlling when and where a policy resumes can turn its existing capabilities into more successful task executions. Project page: https://junhoo.me/acro

Abstract: World action models (WAMs) built on pretrained video generation models adapt quickly to new robotic tasks, but they forget previously learned tasks under sequential fine-tuning. Experience replay (ER) mitigates such forgetting by rehearsing data of previous tasks, yet this data is often unavailable in robotics, where a policy is deployed as weights only and its training data is not shared along with it. In this work, we ask whether a WAM can achieve data-free continual learning by regenerating its own past. Our key observation is that a WAM can indeed autoregressively regenerate complete rollouts of previously learned tasks and replay them to mitigate forgetting. However, replaying such raw rollouts still lags far behind replaying real data, which we attribute to two issues. First, the WAM is unable to recognize task completion and keeps generating purposeless motion after the task is done. Second, under open-loop generation, the generated actions gradually drift away from the generated observations. To address both issues, we present rectified generative replay (R$^2$-WAM), which truncates generated rollouts at task completion with an agentic evaluator and filters out rollouts whose actions are physically inconsistent with the observations. On the four LIBERO suites and five real-world tasks, R$^2$-WAM reduces forgetting by over 90% relative to sequential fine-tuning and nearly matches ER without accessing any data from previous tasks. Our analysis further reveals that forgetting in WAMs concentrates in inverse dynamics modeling: a forgotten WAM still produces visually plausible observations of previous tasks while its generated actions fail.

Abstract: An embodied brain, or System 2, must turn an abstract goal into actions while discovering relevant scene conditions and checking whether executed skills achieved their intended effects. Open-loop evaluations cannot test observations shaped by an agent's decisions, while some closed-loop evaluations provide limited evidence during execution or assume idealized skill outcomes. We introduce SCOPER, a suite for studying observation-grounded planning, execution, and recovery. SCOPER-bench combines high-level goals, ongoing visual observations, a shared rule-based System 1 for low-level actions, and controlled execution failures. Its construction and validation pipeline yields 870 tasks across BEHAVIOR and ProcTHOR, complemented by about 3,500 situated questions tied to the evaluated agent's execution. Benchmark results reveal difficulties in retaining scene evidence, monitoring progress, and recovering from unexpected outcomes. To address these difficulties, SCOPER-harness combines Evidence Memory and Progress Verification with a dynamic task agenda and bounded exploration, without additional model training. Across two evaluated open-source backbones and two simulators, the harness improves task success by 11.1-28.6%; a harnessed 9B model outperforms a directly executed 27B model on both simulators. In an offline next-action evaluation using observations from 200 real-robot demonstrations, the harness also improves full-call accuracy by 14.7% and 15.3%. SCOPER provides a setting for diagnosing and improving embodied reasoning throughout a complete task execution cycle.

Abstract: Vision-Language-Action (VLA) models are often pretrained via imitation learning, making reinforcement learning (RL) post-training important when deployment moves beyond demonstration coverage. Existing refinement paradigms expose a reliability--efficiency tension: model-free refinement avoids simulator bias but is interaction-intensive, whereas model-based refinement can exploit unreliable imagination even as its simulator adapts. We propose Model-Evolving Reinforcement Learning (MERL), a trust-calibrated framework that couples policy optimization with an online-updated simulator comprising a world model and a reward proxy head. Across stages, simulator reliability controls the imagined-data ratio and rollout horizon; within rollouts, chunk-level trust controls replay selection and optimization weight. For recursive imagination without exact grounded continuations, a residual predictor estimates trust without querying future ground truth. On perturbed LIBERO-PRO tasks, MERL reaches 79.7% average success, improves high-threshold learning speed by 10.9% over the matched model-free control, and outperforms same-protocol external baselines across three training seeds. Fixed-data adaptation from 50 demonstrations per task achieves 80.0% average success on two physical-robot tasks.

Abstract: Recent advances in 3D reconstruction enable the extraction of temporally consistent 3D motion from RGB observations; however, current Vision-Language-Action (VLA) models still predominantly rely on 2D image-text priors. Therefore, they do not accurately represent the fundamental dynamics of the 3D world associated with physical actions. This over-reliance on static 2D visual cues limits the capacity of VLA models in achieving robust robot manipulation. To address these limitations, we introduce FlowVLA, a reconstruction-guided, pluggable 3D flow learning framework aimed at empowering VLA models with an inherent understanding of action-relevant dynamics. Instead of requiring explicit 3D sensor inputs or altering the inference-time architecture, our method is designed to utilize temporally aligned 3D flow only during training. This allows the model to learn the dynamic evolution of the physical scene while preserving the computational efficiency of conventional 2D VLA backbones. To bridge scene-level dynamics with robotic manipulation, we propose a modular 3D flow learning framework with dual-level flow supervision. Specifically, point-level flow guides the model to forecast the future displacements of task-critical 3D points, while TCP-level flow aggregates these motions to align with the robot's TCP trajectories. This dual-level flow supervision encourages the model to focus on semantically meaningful dynamic regions, effectively coupling scene-level motion with precise control signals. Extensive evaluations across simulation and real-world experiments demonstrate that FlowVLA consistently improves manipulation performance, zero-shot robustness, and optimizes data efficiency. FlowVLA adds only lightweight auxiliary heads during training and discards them at inference. Therefore, FlowVLA introduces no additional 3D sensor inputs, architectural modifications, or inference-time computation.

Abstract: Multi-environment POMDPs (ME-POMDPs) generalize POMDPs by encoding a finite set of possible environments, capturing model uncertainty. The finite-horizon robust reward maximization problem for a ME-POMDP asks for a policy that maximizes reward in the worst environment, and recent work has shown this problem to be $\mathsf{PSPACE}$-complete. We expand the complexity landscape of ME-POMDPs by identifying which subclasses retain this $\mathsf{PSPACE}$-completeness. In particular, we take subclasses of POMDPs that are tractable in the single-environment case, and investigate the computational complexity of these classes when extended to the multi-environment setting. We first prove the $\mathsf{NP}$-completeness of multi-environment blind MDPs and deterministic POMDPs using a certain closure-under-mixture condition. We then prove $\mathsf{PSPACE}$-completeness for a range of subclasses that do not satisfy that closure condition, yielding new complexity results for multi-observation POMDPs. Finally, we show existential theory of the reals ($\exists \mathbb{R}$)-completeness for numerous ME-POMDP problems under time-dependent Markovian policies or finite-state controllers that are $\mathsf{NP}$-complete for standard POMDPs. These results reveal a subtle relationship between POMDP and ME-POMDP complexity that depends heavily on the subclasses involved.

Abstract: Diffusion policies are increasingly used in robotic control and manipulation, vision-language-action models, and world-model-based agents. Their ability to represent multimodal action distributions offers rich opportunities for exploration in reinforcement learning (RL). However, optimizing diffusion policies with RL can reduce the behavioral diversity and further limit their exploration capability. In RL post-training, directly fine-tuning a pretrained policy may also disrupt its acquired capabilities, while action-space residuals may lack the expressive power for multimodal exploration. Together, these issues reveal a common challenge across online and offline-to-online diffusion RL: improving task performance while retaining the behavioral diversity and capabilities needed for continued adaptation. We introduce CFG-Policy, a unified, plug-and-play framework that separates exploration and exploitation into two generative policies and combines their score or velocity fields during sampling through classifier-free guidance. For online RL, a frozen exploration branch sustains behavioral diversity while an exploitation branch learns from task rewards; for offline-to-online RL, a pretrained exploitation branch remains frozen while a trainable branch steers the composed behavior. CFG-Policy preserves the underlying RL objectives and improves performance with QSM and QVPO on 5 MuJoCo tasks and with Flow-SDE+PPO on 4 RoboTwin manipulation tasks. Project page is at https://zhy-liu.github.io/cfg-policy-page.

Abstract: World Action Models (WAMs) combine action generation with future prediction, allowing predicted futures to support action evaluation and robot decision making. However, the future associated with a candidate action may contain action-induced and environment-driven changes. Directly evaluating the predicted future may not fully characterize the future change associated with the candidate action. For example, a candidate action that worsens a favorable environmental future may still yield a favorable outcome, while one that improves an unfavorable environmental future may still yield an unfavorable outcome. To analyze this issue, we formulate a Structural Causal Model (SCM) for action evaluation in WAMs. The SCM formalizes how action-induced and environment-driven changes jointly shape future evaluation, and reveals a backdoor path through which environmental evolution can also affect the future evaluation associated with a candidate action. Intervention and backdoor adjustment account for this path, providing a causal basis for characterizing the future change associated with each candidate action. Based on this analysis, we propose Causal Action-Gain World Action Model (CAG-WAM). For each candidate action, CAG-WAM predicts an action-conditioned future and an environmental future under the same current observation, and defines their evaluation difference as causal action gain (CAG) for characterizing the future change associated with the candidate action. During training, we incorporate CAG into Group Relative Policy Optimization (GRPO) to guide robot policy learning. During inference, we evaluate the initial action with CAG and, when needed, revise it using predicted future information. Experiments on multiple benchmarks support the effectiveness of CAG-WAM.

Abstract: Test-time sampling promises to turn one action prediction into a set of distinct executable candidates that a verifier can choose among. Downstream fine-tuning works against that promise. A vision-language-action model trained to solve one task distribution concentrates its flow-matching decoder around the dominant demonstrated action, so repeated samples differ only marginally and the verifier receives many copies of one decision. We call this proposal contraction and identify it as a key constraint on test-time scaling: repeated samples can remain highly redundant as the candidate budget grows. Action Dynamic Sampling (ADS) alleviates the constraint on the proposal side, during training. At each training step, ADS integrates $K$ additional trajectories from the current policy under fresh base noise, scores each by the flow support along its path with a Hutchinson trace estimator, and corrects the policy's local action distribution toward this weighted set through a detached finite-set KL term. Training cost rises; deployment cost is unchanged. Across simulation and four physical-robot tasks, ADS raises proposal spread without the success penalty that Gaussian perturbation incurs, lifts oracle best-of-8 success on RoboCasa from 57.6% to 63.0%, improves fixed verifiers by 2.4 to 4.9 points in simulation, and raises TokenPB by up to 13.0 points on real-world manipulation.

Abstract: Pretrained vision-language-action (VLA) models encode broad manipulation priors, yet adapting them to specific tasks requires balancing limited robot interaction, computational resources, and safety constraints. Noise space learning (NSL) adapts flow-matching-based VLAs by learning a context-conditioned distribution over noise while keeping the pretrained VLA fixed. Building on prior noise-steering methods, we formulate NSL as an interface for VLA adaptation that accommodates different learning algorithms and instantiate it with PPO, SAC, and AWR. Theoretically, we provide an information-theoretic analysis of how NSL enables efficient adaptation. Empirically, we study these NSL algorithms on LIBERO, RoboTwin, and four real-robot manipulation tasks under interaction, computation, and execution constraints. Across these settings, we develop a suite of effective NSL variants and derive practical recommendations for choosing among them under different resource and deployment regimes.

Abstract: Momentum methods such as STORM and second-order momentum track the gradient by adding a transport correction to momentum, at the cost of an extra oracle call per iteration. We introduce randomized momentum transport, which evaluates the correction with probability $q$ and otherwise uses a cheap prediction, with importance weighting that preserves the conditional mean and adds $(q^{-1}-1)$ times the prediction loss to the variance. Under transport-bias, second-moment, and residual-control assumptions, linear minimization oracle (LMO) updates, including normalized, sign, and exact spectral (Muon) directions, retain an $O(T^{-1/3})$ stationarity bound for each fixed $q>0$ with a single-gradient initialization. For unnormalized STORM, some $q<1$ lowers the leading coefficient of our sufficient gradient budget if and only if the gradient-difference noise modulus exceeds the smoothness constant, by at most about $10\%$. An online diagonal predictor admits a regret-to-convergence guarantee. At approximately equal oracle cost, querying a quarter of the corrections lowers the final training loss by $18$ to $43\%$ on MNIST1D and raises the test accuracy of Hessian-vector transport on CIFAR-10 by $1.5$ points.

Abstract: On-policy distillation (OPD) trains a target policy on its own rollouts using token-level supervision from a stronger teacher, but typically relies on the teacher being substantially more capable than the target. Recent work relaxes this requirement by transferring the relative policy difference between an expert and a reference policy rather than the expert's absolute distribution. Yet it remains unclear which differences provide useful supervision and which targets can benefit from them. We develop a unified view in which the log-probability ratio between a proxy expert and a proxy reference defines an implicit reward for KL-regularized policy optimization. This formulation induces an on-policy update for the target, recovers conventional OPD as a special case, and enables any-to-any transfer across models that differ in scale, training stage or generation. Our systematic experiments yield three findings. First, a stronger expert is not necessarily a better source of supervision: differences induced by task-specific training, including reinforcement learning and supervised fine-tuning, transfer consistently, whereas those arising from model scale alone transfer weakly or not at all. Second, the expert and reference serve complementary roles: the expert provides the capability to be transferred, while the reference filters out shared behavior and isolates the relevant residual. Third, even a useful difference does not transfer to every target: success requires the induced update to move the target toward the expert and generalize beyond sampled rollouts. Raw policy distance does not predict transferability, whereas an appropriate reference can rescue otherwise unsuccessful transfer. These results characterize when relative policy differences provide effective supervision for OPD.

Abstract: Human demonstrations are a scalable data source for learning dexterous manipulation, but the embodiment gap prevents human motion from being executed directly on robots. Inverse kinematics (IK) retargets human motion to robots efficiently but ignores dynamics, often producing infeasible motions. Reinforcement learning (RL) and sampling-based model predictive control (MPC) are commonly employed to yield dynamically feasible motions, but both are sample-inefficient and sensitive to hyperparameters. RL suffers from costly and unstable training and tedious reward engineering; MPC avoids policy optimization, yet retargets each trajectory in isolation, and solving one does not make the next easier. Sampling cost grows rapidly with dataset size and task difficulty. We hypothesize that dynamically feasible trajectories concentrate near a low-dimensional manifold shared across demonstrations, so that retargeting can be reduced to sampling from that manifold, conditioned on human motion, rather than solving a fresh optimization problem for every demonstration. We propose \textbf{Generative Neural Retargeting} (GNR), which uses a flow matching model to sample feasible trajectories. GNR outperforms MPC with only $8.5\%$ of the samples required by MPC, achieving a success rate of $56.20\%$ compared to $27.20\%$ for MPC. GNR can be used for scalable and efficient retargeting of large-scale, long-horizon, and millimeter precision human demonstrations: by applying GNR within a real-to-sim data engine, we produce a dexterous manipulation dataset with dense contact-force labels, spanning $223$k demonstrations and $3.3$k object geometries.

Abstract: Shortcut learning is a prevalent issue in robot learning. The limited diversity of robot demonstration datasets can mislead policies into exploiting spurious correlations between tasks and irrelevant features, such as viewpoint or background. Collecting sufficiently diverse robot demonstrations is costly and inefficient, motivating algorithmic alternatives. We focus on vision-language-action (VLA) models and discover that different vision-language model backbones exhibit substantially different levels of susceptibility to visual shortcut learning. We find that model behavior correlates with our proposed representation-level metric, action margin, which requires no policy rollouts. Visual shortcuts consistently enter action representations in early layers, with models differing in the extent to which later layers correct them by incorporating language information. To boost models' attention to language, we introduce task scrubbing, a new domain-adversarial training method that decreases models' likelihood of using visual shortcuts and improves VLAs' generalization. Experiments in both simulation and the real world across multiple VLAs and visual cues show that task scrubbing improves out-of-distribution robustness and often eliminates visual shortcut learning.

Abstract: Long-term agents face growing storage demands as they accumulate experience. World models capture reusable regularities that can reduce the information stored for each experience. We formulate the problem of memory allocation conditioned on a world model and introduce MemoWM, a framework that uses shared predictions to compress retained information and reconstruct omitted content. Its task-aware allocation rule balances the expected impact of reconstruction errors against storage cost, retaining information with downstream value beyond the predictive prior. Across five long-term agent-memory benchmarks, MemoWM achieves 42.42\% average answer accuracy, exceeding the strongest baseline by 2.62 percentage points, while reducing average experience-specific storage by 53.9\% relative to MIRIX, the most storage-efficient baseline. Further analysis shows that stronger world models reduce per-experience storage at comparable task quality. Accounting for model parameters reveals a trade-off between shared model capacity and recurring storage costs, with the capacity that minimizes total storage increasing as more interactions are retained. Our code is available at https://github.com/Feld-maxiu/MemoWM.

Abstract: Internalizing external skills changes a language agent's capabilities and, with them, the value of its remaining guidance: rules can become redundant, misleading, or insufficient for newly encountered decisions. This creates a coupled problem of learning from skills and adapting the skills that supervise further learning. We introduce SkillCycle, a framework for co-evolving agent policies and skill banks through a feedback loop between skill internalization and rule revision. Our central contribution is to give distillation feedback a second role: token-level contextual differences help locate rules for inspection, while interaction outcomes guide edits to their content and applicability. SkillCycle alternates between two phases: policy learning with a fixed skill bank and router, and rule revision with a frozen policy. Candidate edits undergo rule-level and whole-bank environment comparisons before they guide the next learning cycle. On WebShop, SkillCycle with a 3B model achieves a success rate of 74.74% and a score of 88.37 without inference-time skill inputs, representing relative improvements of 0.73% and 3.96% over the state-of-the-art (SOTA) model, respectively. In Cycle 3 ablations on ALFWorld and WebShop, SkillCycle's no-skill success rates improve by 10.18% and 18.11% relative to a static skill bank, and by 2.41% and 2.50% relative to a single bank update, respectively. These results show that continually revising skill guidance as the agent's capabilities change helps transform external skills into policy capabilities that require no skill inputs at inference. We will release code, configurations, skill banks, and evaluation protocols.

Abstract: Multi-task manipulation policies differ in architecture, scale, and pretrained priors all at once, so published comparisons cannot attribute performance to any single component. We argue that most of it comes from the visual representation, and that parameter scale and generative priors are largely incidental. To test this, we build CoRP (Compressed Representation Policy), a deliberately compact policy (48.9M parameters, no vision-language model and no video-generative prior) that factorizes into a representation extractor and a flow-matching action generator. It reaches 97.0% on LIBERO and 75.78%/73.36% on RoboTwin 2.0 Clean/Randomized, matching systems 40.9-163.6x larger. Holding the action generator fixed, we then vary one extractor property at a time. Pretrained initialization is decisive: a random ViT-S/14 drops to 78.1% and an ImageNet ResNet-34 to 74.5% on LIBERO. Pretraining alone is not enough, as freezing the encoder costs 19.8 points. Compression matters as much: resampling each view to 48 tokens beats passing all patch tokens (97.0% vs 83.2%), and a variational information bottleneck over those tokens is worse than a hard token budget, cutting LIBERO-Goal from 95.8% to 33.0% by suppressing the instruction-dependent token selection the policy relies on. Language conditioning contributes only where the observation leaves the goal ambiguous (LIBERO-Goal: 9.2% to 95.8%), while on RoboTwin 2.0, where observations are unambiguous, removing it slightly improves success. Therefore, we argue that a compact policy works when its representation is pretrained, task-adapted, and compressed. Project page: https://corp-policy.github.io/

Abstract: Self-improving LLM agents increasingly distill successful experiences into persistent, reusable skills. Existing skill attack methods corrupt this learning pipeline by injecting malicious triggers, behaviors, or false facts into individual experiences or extracted skills. However, such attacks are easily detected, and the injected malicious behaviors often fail to accumulate as persistent skills. In this paper, we show that skill poisoning can arise even from verified successful experiences, without making any individual trajectory malicious. Based on this insight, we propose SkillPoison, a novel framework that progressively poisons skill via successful experiences. SkillPoison first constructs a set of successful experiences that reinforce a target behavior, and then removes the contextual conditions that constrain when the behavior applies. Rather than injecting malicious content, SkillPoison shapes how the skill extractor generalizes, allowing useful behavior to support task success while inducing harmful behavior when they are misapplied. Extensive experiments on three benchmarks show that SkillPoison achieves 95.71% attack success rates, while all injected experiences remain task-correct and pass verification and lexical inspection. Our code, data and implementation details are available for the community at https://github.com/DEEP-JLU/SkillPoison.

Abstract: Online agents can improve future service by constructing reusable tools, guidance, or model states, but this work competes with current requests for the same GPUs. Exploiting idle compute for self-evolution faces a fundamental systems constraint: benefits arrive only after an artifact is published and used, while pausing evolution leaves service capacity waiting for memory release and runtime recovery. An investment worth completing may therefore be worth postponing. We present LearnSched, a state-aware scheduler that brings the reuse window of a capability and the timely return of compute into a common investment model. LearnSched incorporates candidate progress, checkpoint overhead, and the current recovery path into action values. Under the same information and capacity constraints, it uses one-step counterfactual rollout to choose progress, checkpointing, or waiting relative to a fully costed window policy. We characterize handoff costs through independent A100 component measurements and evaluate action selection in 1920 paired finite-model scenarios. When cold recovery is expensive, waiting for longer execution windows can improve net value by avoiding handoffs; in evaluated warm-recovery scenarios, the strong window policy already achieves the same value. Retaining recoverable state can shorten capacity return and reduce the need for complex evolution scheduling.

Abstract: Post-training quantization has been shown to preserve VLA performance under standard evaluation conditions, but whether it preserves the full-precision model's robustness and generalization remains underexplored. In this study, we systematically study the robustness and generalization of post-quantized VLA policies under environmental disturbances. Empirical results show that quantized policies can become fragile to subtle environmental variations despite retaining comparable in-distribution performance. We further observe that action discrepancies are concentrated in a small subset of rollout states, while teacher guidance has opposite effects depending on discrepancy: it improves generalization at high-discrepancy states but can degrade it at low-discrepancy states. These findings reveal that effective post-quantization recovery requires selectively intervening on vulnerable states rather than globally distilling the student. We therefore propose Policy-Induced Vulnerability-Oriented Tuning (PIVOT-Q), a vulnerability-aware On-Policy Distillation (OPD) framework that selectively corrects vulnerable states encountered during quantized-student rollouts using the frozen full-precision policy as a teacher. PIVOT-Q identifies vulnerable states using discounted accumulated discrepancies over a short horizon, applies phase-balanced sparse supervision, and uses a Behavioral Anchor to prevent unnecessary changes. Experiments under seven LIBERO-Plus environmental variations demonstrate consistent recovery across multiple VLA backbones and quantization methods. Notably, PIVOT-Q consistently outperforms full-state distillation across all settings while using only 7.4% of its state-level distillation budget. Our code is available at https://github.com/ruanruan-andy/PIVOT-Q.

Abstract: In recent years, federated stochastic bilevel optimization has attracted increasing attention due to its wide range of applications in machine learning. To reduce the computational overhead associated with second-order Hessian and Jacobian matrices, several first-order methods have been proposed. However, existing methods typically impose restrictive assumptions on the lower-level function, suffer from a strong dependence on the condition number in their convergence rates, and require different learning-rate scales for variables across the upper- and lower-level problems, limiting their practical applicability and complicating hyperparameter tuning. To address these challenges, we propose a stochastic doubly smoothed gradient method for nonconvex federated stochastic bilevel optimization problems, which decouples the learning rates of upper- and lower-level variables and does not require a strongly-convex lower-level loss function. We establish rigorous theoretical guarantees for the proposed algorithm, demonstrating an improved convergence rate of $O(κ^{15/2}/ε^5)$ and a communication complexity of $O(κ^{4}/ε^3)$, where $κ$ denotes the condition number and $ε$ represents the solution accuracy. Notably, these bounds exhibit significantly better dependence on the condition number $κ$ than those of existing methods. Extensive experiments validate the effectiveness of our algorithm.

Abstract: Precise constraint satisfaction is a prerequisite to deploying learned models in many areas, motivating methods that repair raw neural predictions with a repair procedure. Current methods unroll multiple repair steps in training and softly penalize constraint violations that remain after the unroll. This is compute- and memory-intensive, lacks robustness when the repair fails to converge, and surrenders most of the constraint satisfaction work to the repair. Our central finding is that, contrary to common practice, a single detached projection step suffices in training. We accomplish this with a Projection-Adaptive Loss (PAL), which uses the constraint residual after this single step to adaptively weigh constraint penalties on the raw prediction. In experiments, PAL is the only method that retains virtually perfect feasibility on extremely nonlinear constraints, and matches or outperforms current methods on synthetic and engineering benchmarks. Because it only requires a single detached projection step, PAL trains 2.5x faster than the canonical repair-based method (DC3) on its own ACOPF benchmark. PAL can also be trained when constraints are expensive to evaluate (e.g., via neural surrogates), a setting where current unrolled methods are memory-intractable.

Abstract: Latent world models plan by scoring candidate action sequences with distances in latent space. However, task success is judged by physical quantities, which we call the success-criterion quantities. In all four latent world models we examine, the end-effector position is encoded in the latent state with an error larger than the success criterion allows. Such a latent state cannot separate successful candidates from failing ones. We propose an auxiliary loss that uses success-criterion quantities as training targets, whereas existing latent world models take them only as inputs. During training, a linear head on the encoder and predictor outputs regresses the success-criterion quantities, and the regression error is added to the training loss. The head is discarded after training, so the model, its cost, and its inputs at test time are unchanged. This loss alone improves the success rate on PushT and cube by 3.5% and 3.4% (absolute), respectively, and both improvements are statistically significant. A success criterion thus specifies what a world model must retain in its latent state, and we show that it can serve directly as a training target.

Abstract: Autonomous research agents can generate hypotheses and conduct experiments, but experimentation remains a major source of computational cost. A fundamental challenge is deciding which experiments are worth running, particularly when prior evidence is insufficient to resolve uncertainty. Yet current research agents lack a systematic way to leverage experimental experience when making such decisions. We introduce Experimental Experience Modeling (EEM), a framework for making informed experimental decisions by acquiring, reusing, and accumulating experimental experience. EEM extracts decision-relevant records from earlier experimental trajectories, distills them into reusable experience, and organizes them in an experience library. For a new experimental decision, EEM retrieves relevant historical experience and assesses whether it provides sufficient support for deciding whether a candidate direction warrants further investment. When historical experience is insufficient, EEM conducts a targeted, low-cost pilot experiment to acquire the missing decision-relevant experience on demand. It then combines this newly acquired experience with retrieved historical experience to determine whether the direction warrants full-scale evaluation, which requires substantial resources. The resulting experimental outcomes are further distilled into reusable experience, allowing the library to continually grow through iterative accumulation. Experiments on autonomous research benchmarks show that EEM improves research performance while reducing model interaction overhead, demonstrating the value of reusing accumulated experience and acquiring additional experience only when needed.

Abstract: World-action models jointly learn robot policies and predict future observations, making the representation space an interface between control and prediction. We study the design of this space through controlled comparisons, finding that neither reconstruction fidelity nor pre-trained perceptual features alone ensure effective policy learning. These findings motivate ReWAM, a representation-centric world-action model built on pre-trained DINO features. Feature Calibration and a Temporal Representation Bottleneck organize these features into compact world states suited to dynamics modeling. Action-Grounded Representation Shaping routes only action-loss gradients to the bottleneck, thereby letting the policy shape what the representation encodes while the world model learns how it evolves. Without generative video pre-training, ReWAM achieves 93.6% success on RoboTwin 2.0. On RoboDojo, it achieves an average score of 12.29 and a success rate of 8.28% using approximately 600 hours of embodied pre-training data.

Abstract: Neural simulation-based inference (SBI) has been widely successful in inferring a relatively small number of interpretable parameters from potentially high-dimensional observations, such as images or time series. Accordingly, representation learning in SBI has focused almost exclusively on compressing the observations used to condition the posterior. More recently, however, SBI has begun to target increasingly high-dimensional parameter spaces, raising the complementary question of whether the inference target itself should be compressed. Our answer is a practical merger of SBI and latent generative modeling, which learns a low-dimensional representation of the simulator parameters, performs posterior inference directly in this latent space, and maps posterior samples back to the original parameter space. We characterize the conditions under which latent-space inference recovers the desired target posterior and systematically study its empirical trade-offs. Across four case studies and three generative families, we compare latent and standard estimators while controlling for network capacity, regularization, optimization, and training compute. At matched training compute, latent-space inference achieves accuracy and marginal calibration comparable to direct target-space inference while sampling up to more than an order of magnitude faster.

Abstract: Long-horizon agents require context compression to manage growing interaction histories. Compression quality, however, is ultimately determined by downstream execution. Existing prompt-adaptation methods infer compression errors by comparing full-context and compressed trajectories. Such comparisons cannot isolate individual compressions and are confounded by agent stochasticity. We first find that compression degrades reliability before solvability. Using matched counterfactual continuations that compare execution from the same agent state with versus without compression, we further show that severe degradation concentrates at isolated compression events. Motivated by this finding, we propose PAIR (Prompt Adaptation using Interventional Rollouts) for adapting structured compression prompts. PAIR identifies individual compressions that degrade subsequent execution, diagnoses their effects, and revises the relevant sections of a fixed compression template. PAIR achieves the strongest cross-run reliability among compressed methods in every main benchmark-scope combination, consistently exceeding the competing prompt-adaptation baseline. Without modifying the downstream agent, PAIR brings compressed execution close to the no-compression baseline and sometimes numerically exceeds it.

Abstract: On-policy distillation (OPD) trains a student on its own trajectories with dense teacher supervision. Recent work on OPD for multi-turn autonomous agents often treats large teacher-student token-level distributional gaps as promising intervention points, linking larger gaps to a greater need for correction. Yet, our empirical analysis reveals a supervision-benefit mismatch: large gaps can be benign, while small gaps can be outcome-critical. Teacher-student gaps capture differences at the current turn, whereas the benefit of teacher guidance depends on how the current student interacts with the environment afterward. The student may still succeed despite choosing an action that differs from the teacher's, while a teacher-preferred action may lead to a state from which the student cannot complete the task. Local gaps alone are therefore not enough to determine whether teacher guidance benefits the current student. Effective supervision should instead emphasize guidance that the current student can translate into better final task outcomes. Accordingly, we propose Outcome-Guided On-Policy Distillation (OG-OPD), which applies trajectory-relative weighting to teacher supervision and calibrates these weights using final task outcomes from paired student continuations. This calibration selectively strengthens supervision on the student's original trajectories at turns where teacher guidance benefits the current student. Across ALFWorld, ScienceWorld, and WebShop, OG-OPD consistently outperforms baselines under diverse settings. It improves task success rates by 3.6-17.7 percentage points over vanilla OPD and by up to 7.0 percentage points over the strongest baseline.

Abstract: Flow-based Vision-Language-Action (VLA) policies are typically trained by behavior cloning and thus do not explicitly optimize long-term task return. Critic guidance steers generation toward higher-value actions, but existing methods differentiate the critic through a one-step surrogate of the sampler and back-propagate a critic ensemble at every flow step. In contrast, here we propose Adjoint Guidance Flow (AGF), which amortizes trajectory-aware critic guidance into a lightweight guidance network while preserving the pretrained VLA policy. Specifically, we formulate critic-guided flow generation as a deterministic optimal control problem, whose optimal guidance is a costate that carries the terminal critic gradient back through the remaining flow, and regress the guidance network onto this costate while keeping both the VLA and critic frozen. This design provides favorable memory and throughput scaling during training, and inference needs one guidance-network forward pass per step, without the critic ensemble, back-propagation, or adjoint computation. Across LIBERO, RoboCasa, and LIBERO-Pro, AGF consistently improves pretrained VLAs, remains competitive with critic-guidance and policy-fine-tuning baselines, and is the most robust method when a single guidance strength is deployed across tasks. Compared with QGF, AGF runs $3.6\times$ faster per guidance step with $7.0\times$ fewer parameters, with comparable and even better performance, showing that critic guidance can be trajectory-aware and lightweight.

Abstract: Reinforcement learning has become an effective approach to training language model agents, but sparse and delayed outcome rewards provide limited guidance for credit assignment across long interaction sequences. Recent work on on-policy self-distillation (OPSD) offers complementary supervision by evaluating a policy's sampled responses under privileged training-time context. However, our diagnostics show that positive average agreement between outcome and hindsight feedback coexists with substantial local disagreement, raising the question of how to allocate influence between them at each decision. We introduce UniOPSD (Unified On-Policy Self-Distillation), which unifies these feedback sources through adaptive local credit arbitration. UniOPSD constructs comparable credit estimates from environmental returns and successful-peer hindsight at shared interaction anchors. Historical agreement determines the global mixing level, while current signal availability and relative precision adjust each source's influence at individual decisions. The episode-level outcome contribution is retained, and bounded token modulation refines the fused step credit for policy optimization. With Qwen2.5-3B-Instruct and Qwen2.5-7B-Instruct, UniOPSD achieves ALFWorld success rates of $82.8\%$ and $83.6\%$, WebShop success rates of $75.0\%$ and $82.0\%$, and Search-QA aggregate accuracies of $45.3\%$ and $49.8\%$, respectively. On 3B WebShop, UniOPSD improves over SDAR by $7.0$ percentage points. Our code is available at https://github.com/Zenghuang-Fu/Uniopsd

Abstract: Does repairing an episode make its experience a better memory for the next task? We transfer the same failed source before and after accepted repair to a fixed target, alongside independent execution. Our 3,300 runs cover 100 ThinkingBox pairs and the same 100 APEX pairs with and without source-state inheritance, under eleven conditions. ThinkingBox's Full/Skill/Hybrid correction gains are 44/29/32 percentage points, with corrected performance 25/22/18 points above independence; inference weakens at the task-family level. Yet 12 of Full's 15-point larger correction gap over Skill come from worse uncorrected performance, not better corrected memory. Moreover, 22 of Full's 46 upward transitions restore observed baseline success. Neither APEX regime establishes comparable aggregate correction benefits. Action evidence connects workflow gains with reusable obligations and convention conflicts with source-local choices. Text APEX's accepted execution reaches 52% versus its summary's 40%, without robust global/group-level superiority or an estab- lished advantage over independence. Smaller handoffs reduce input but increase calls. The value of repairing experience is therefore distinct from the value of reusing it: memory updates require both a previous-version reference and a fresh-start reference.

Abstract: Recent advances in Vision-Language-Action (VLA) models point toward general-purpose robotic intelligence by unifying perception, instruction, and control. Despite impressive progress, existing VLA models often adapt poorly due to \emph{tri-modal misalignment} among vision, language, and action, which weakens action grounding and hurts generalization and fine-tuning efficiency. In this work, we present Alignment-Guided Flow Transformer (AGFT), a novel framework that explicitly enforces tri-modal alignment through a dedicated alignment loss, bridging the representational gap across modalities and enhancing task adaptation. While prior research has predominantly emphasized bi-modal vision--language alignment, we systematically formalize and study tri-modal alignment in VLA models, and provide both ablations and analysis to isolate its role in improving adaptation and robustness. To further accelerate deployment, we adopt a flow-matching objective, enabling substantially fewer inference steps than diffusion-based policies while maintaining accuracy. Theoretically, we establish a quantitative connection between the tri-modal alignment gap and the optimization tightness of flow matching; empirically, experiments on the extensive benchmark show that AGFT achieves superior success rates and lower inference latency compared to SOTA baselines, underscoring tri-modal alignment as a key ingredient for scaling robust VLA manipulation.

Abstract: An agent's executable program governs how it uses tools, processes observations, and responds to failures. Evolving this program at test time can help adaptation, but deciding which changes to retain is difficult when true rewards are unavailable. Execution traces provide evidence of agent behavior, yet interpreting that evidence requires a judge that remains useful as tasks and candidate programs change. We introduce Judge-Guided Evolution at Test Time (JET), which evolves an executable judge on labeled source trajectories, then freezes and transfers it to guide target-side program evolution. The judge supplies scores and diagnostic feedback without target evaluator access or model-weight updates. On unseen WebShop tasks, JET achieves approximately 13% higher mean reward than fixed-rubric guidance when evolution begins from an unevolved program (cold start) and 4% higher when it begins from one already optimized on source tasks (warm start), with a 36% relative improvement in cold-start exact success. An exact-judge control on PushT, where the judge reconstructs the scoring rule from observations, shows that without judge error, program search becomes the bottleneck. Analyses identify useful reward-prediction logic in the evolved code and show that better final selection alone cannot explain the gains. These results support executable judge transfer for program adaptation under evaluator-preserving task shifts.

Abstract: Observationally equivalent causal models can still disagree about what happens under intervention, because interventions create inputs that never occur in observational data. We introduce Interventional Separation Selection (ISS), which repeatedly queries the true system with an admissible intervention on which the surviving candidate models disagree, discards the candidates the outcome contradicts, and stops once no intervention within a cost bound separates the survivors. If the true system is among the candidates, this stopping condition certifies that every survivor agrees with it on every admissible intervention within the bound, a guarantee that no observational learner can give, however much data it sees. The stopping condition depends only on the survivors, so it can be checked without knowing the truth. For continuous variables the candidates form an infinite version space, and mixed-integer linear programs decide the stopping condition exactly over all of it, with agreement holding up to a tolerance. On a three-digit colored MNIST causal abstraction task in which ink hue tracks digit size, plain convolutional networks trained on examples reach zero held-out error, yet disagree with shape-based labels on 26% of single-digit edits, as often as hue-based labels do. Auditing the causal abstractions of networks observed only on such images, ISS certifies what each network perceives with 13.6 interventions per image on average, and each certificate, checked against every admissible intervention, holds whenever the network's true abstraction is among the candidates. When a network bypasses a unit that every candidate abstraction relies on, certificates covering interventions on that unit can be silently void, and twenty random validation interventions refute 69% of them.

Abstract: Surrogate models are commonly evaluated by how often they agree with their teacher model over an evaluation set. Local variation in this agreement is well known, but its structure and consequences are less clear. We ask whether disagreement is systematically concentrated near the teacher's decision boundary and whether retaining that structure provides information beyond a single global score. Across several datasets and surrogate model classes, we find substantially lower fidelity near teacher decision boundaries under two different methods of identifying near-boundary examples. Moreover, conditioning agreement on confidence-defined regions improves prediction of teacher--surrogate agreement when evaluation-set composition changes, relative to the global score alone. Yet surrogates that agree equally well with the teacher both globally and near the decision boundary can respond very differently to changes selected using the surrogate itself. Finally, we show that independently trained deep teachers can agree on most predictions while identifying different examples as lying near their decision boundaries, complicating the use of those boundaries as stable reference regions for evaluating surrogates. Together, these results show that surrogate fidelity depends not only on how often a surrogate agrees with its teacher, but also on where that agreement holds and, for deep models, how stable the teacher's decision boundary is across training runs.

Abstract: World Action Models (WAMs) jointly model visual dynamics and action generation for generalist robot manipulation. A central challenge is to integrate priors from large-scale pretrained models---including visual dynamics, scene semantics, geometry, and motion---into a unified framework for robot action generation. We introduce InternW0-$Δ$, a unified WAM pretrained on a heterogeneous corpus that outperforms prior methods across simulation benchmarks and real-robot platforms. InternW0-$Δ$ combines pretrained visual dynamics, scene-level semantics, 4D geometric and motion priors, and action generation within a Mixture-of-Transformers (MoT) framework. A pretrained video expert and an action expert interact under semantic guidance from a frozen VLM, while a pretrained 4D foundation model injects geometric and motion priors through training-only distillation. We further introduce Causal Imprint, which learns future-relevant scene changes from training-only future supervision and provides predictive representations directly to the action expert without future-video rollout at inference. For large-scale joint training, we construct a heterogeneous corpus of robot demonstrations, UMI data, egocentric human demonstrations, and Ego2Robot data, curated and aligned under a common state-action representation. The resulting corpus contains over 20K hours of processed training data, to our knowledge the largest open-source corpus of its kind. We pretrain InternW0-$Δ$ on this corpus and demonstrate strong performance across simulation benchmarks and real-robot platforms. We will open source the training code, model weights, infrastructure, data-processing pipeline, and processed data where licenses permit. Project page: https://internrobotics.github.io/InternW0-Delta/

Abstract: Large video diffusion models offer expressive priors for embodied prediction and learning, yet their many-step sampling remains costly for interactive downstream use. Distribution Matching Distillation (DMD) enables few-step video generation, but can suppress robot--object motion while preserving visual quality. Examining DMD's teacher and fake-score signals, we find that weak re-noising keeps the teacher posterior concentrated near motion-deficient rollouts, limiting motion-restoring guidance. Meanwhile, stronger-motion rollouts tend to incur larger fake-score fitting errors, which can hinder the generator's learning of interaction dynamics. We propose DyMD, a DMD framework that adapts both teacher supervision and critic fitting to the evolving student. Temporal affinity--conditioned re-noise sampling adapts the timestep distribution to each rollout's current interaction fidelity by mixing the base schedule with a teacher prior motivated by local posterior variation, thereby balancing motion recovery and appearance refinement. To better track stronger-motion rollouts, dynamics-guided fake-score tracking uses a noise-conditioned predictor to estimate noise-relative fitting difficulty from latent temporal dynamics, then upweights predicted-hard rollouts in the critic loss. Using DyMD, we distill a 14B teacher into a four-step 1.3B student with no auxiliary modules at inference. On embodied-video benchmarks, the student improves R-Bench task adherence by $9.6$ percentage points and PAI-Bench-G Domain score by $5.1$ points over Base DMD while maintaining comparable visual quality. As a backbone for downstream action planning, our student achieves 34% mean success across two WorldArena tasks, compared with 16% for Base DMD.

Abstract: Artificial intelligence is advancing rapidly, with increasingly capable systems taking larger roles in reasoning, decision-making, scientific discovery, and autonomous development. As AI begins to participate in its own improvement, from model training and experience accumulation to agent evolution and automated AI development, the prospect of recursive self-improvement (RSI) is becoming increasingly relevant. This transition raises a fundamental safety question: how can safety be maintained when the system, its accumulated experience, and even the process producing its successors continue to change? We introduce Evolutionary Safety as a perspective for studying safety under persistent and recursive self-improvement. It concerns not only whether an AI system is safe at a particular moment, but how safety properties change, persist, accumulate, and propagate throughout evolution. We characterize recurring manifestations, including intent drift, error accumulation, experience contamination, safety-property erosion, evaluator drift, and risk propagation. We then develop a taxonomy spanning persistent agent state, model state, evaluation and environmental feedback, computational substrate, and meta-level update mechanisms. Building on this taxonomy, we examine how evolutionary risks can be discovered and evaluated across states, updates, trajectories, and lineages, and derive governance principles for modification, selection, authorization, provenance, and recovery. Finally, we outline open problems toward maintaining safety guarantees as AI systems become increasingly persistent, adaptive, and recursively self-improving. Project resources and proposed evaluation systems are available at https://chaunceykung.github.io/evolutionary-safety-rsi.

Abstract: Multi-robot coordination methods often score a joint plan from singleton and pairwise terms, leaving out the terms that involve three or more robots. We measure the plan-selection regret of two pairwise approximations to delivered coverage using frozen multi-robot trajectories. For each four-robot plan on an indoor exploration benchmark, replaying all 16 robot subsets gives the exact delivered-coverage set function $F$. From the same subset values we compute two pairwise scores: the exact order-2 Möbius truncation $F_2$, which depends only on the singleton and pair values, and an equal-weight least-squares two-additive fit $G$. Ranking by $F_2$ instead of $F$ changes the selected plan on six of seven maps at the 15 m candidate-generation range in each of two candidate families, with regret up to 0.337 of map coverage. Switching to $G$ reduces the regret but still changes the selection on three of seven maps in each family. The additive score $F_1$, which keeps only the singleton terms, selects the exact winner on six of seven maps in one family and four of seven in the other, against one of seven for $F_2$. We also find that lower average reconstruction error does not guarantee lower selection regret.

Abstract: Online reinforcement learning fine-tuning of pretrained flow-matching vision-language-action (VLA) policies promises robots that keep learning after deployment, but continued updates often destroy competence on individual tasks while the aggregate still looks healthy. We study this failure mode, which we call task collapse, under a matched small-compute budget on LIBERO-10 with a 450M-parameter SmolVLA policy trained by PPO with stochastic (SDE) sampling. Three exploration-noise policies differ in one live variable: a fixed noise scale, a ReinFlow-style learned noise network, and an uncertainty-gated controller that redistributes exploration across task streams from task-agnostic novelty and competence signals, without task labels or episode boundaries. Under the pooled definition, fixed noise collapses tasks in two of three seeds and learned noise in every seed measured to iteration 200, while the controller collapses none in any of its three seeds. Measured parameter displacement shows the controller's action expert keeps changing, while its mean applied noise is close to the fixed scale in the available logs. The matched comparison supports the controller's effect on task preservation; the separate contributions of its adaptation across states and over time are not disentangled. A lower fixed scale slows the decline but does not stop it. No arm improves on the behavior-cloning baseline in this budget. Two properties of that regime are measured beside this result, not offered as its cause: following the reference recipe, training runs in bfloat16 with no fp32 master copy, under which 96.02% of the action expert's elements stay bit-identical across three consecutive iterations, and an fp32 master copy at the reference learning rate collapses both arms in a single-seed observation. We release tools measuring per-task collapse under four definitions, rescoring noise and instrument tares.

Abstract: As Large Language Models (LLMs) evolve into autonomous agents that alter real-world states, ensuring operational safety across multi-step workflows has become a critical challenge. While recent work has moved beyond single-turn evaluation toward multi-turn paradigms, key limitations persist: step-level methods treat actions in isolation, missing how risks accumulate, while trajectory-level evaluations operate post-hoc, offering no opportunity for timely intervention. To address these limitations, we formalize Decoupled Proactive Safety Monitoring along three dimensions: whether to intervene, when to intervene, and what the risk is. We introduce PASTABench, a benchmark of 1,139 multi-turn trajectories spanning 5 risk categories and 13 subcategories. We further propose the Optimal Intervention Window (OIW), anchored by annotated Earliest-Signal and Trigger turns, to quantify intervention timeliness. Evaluation of 16 LLMs reveals that proactive intervention remains largely unsolved, with the best model achieving only 40.74% optimal-timing interventions. Fine-grained diagnosis further uncovers pervasive lexical overfitting: competitive safety scores of smaller models mask keyword hypersensitivity rather than genuine risk comprehension, as their proactive capability largely collapses once hazard vocabulary is neutralized.

Abstract: A reinforcement learning policy that satisfies its temporal-logic specification has passed a test, not an assurance argument. The clause that matters to a reviewer may never have mattered to training: it may have been avoided entirely, forced by the environment regardless of what the policy learned, or redundant next to the ordinary task reward. Satisfaction probability and task return cannot tell any of this apart. This paper introduces an audit layer that can. It measures whether a specification clause was actually exercised, whether that role was forced or chosen, and whether the obvious way to test causation, weakening the clause and retraining, is even valid. Often it is not: we prove that comparable weaker/stronger training objectives can share perfect optima under standard acceptance-derived rewards, show related ablation hazards across a large corpus of published specifications, and then show that a properly designed intervention detects the effect it should. Across standard reinforcement learning benchmarks and published external artifacts, the audit layer separates six regimes that a single satisfaction number collapses into one. A policy that satisfies its specification has answered whether. This paper asks why.

Abstract: Comparing current and past observations helps robots understand scene changes and select subsequent actions during manipulation. However, comparing visual features at the same image location can mix different scene content when objects or the camera move. We introduce CoRe-WAM, a world-action model that incorporates correspondence-aligned visual changes through a parameter-efficient temporal interface. Its TraceDelta module uses correspondences from a frozen tracking model to transport historical visual features to current locations before computing signed differences in a shared pretrained feature space. Correspondence thus determines which historical content is compared with the present, rather than entering the policy as a separate trajectory representation. A lightweight adapter converts these differences into validity-gated residuals that supplement current visual conditioning, allowing the policy to use recent changes alongside current-scene information. Built on Motus, CoRe-WAM keeps the pretrained backbone weights frozen and optimizes 1.59 million parameters. With a 5,000-update adaptation budget, CoRe-WAM achieves 92.22% clean success across 50 RoboTwin 2.0 tasks, 3.56 percentage points above Motus; on randomized evaluation, it achieves 89.60% success, a 2.58-point gain. Integrating TraceDelta into a StarVLA-based policy improves clean success from 58.10% to 67.62%, supporting transfer of the temporal interface beyond Motus.

Abstract: Pre-trained robot policies always require finetuning to adapt to specific environments. Reinforcement Learning (RL) offers high performance potential because it improves action optimality rather than simply mimicking data. However, such potential depends heavily on reward quality. Sparse rewards lack process feedback, human-designed rewards are costly and biased, and semantic rewards from foundation representations are often not control-centric. We propose Hindsight Reward Editing (HiRE), a training-free framework to break this reward bottleneck. HiRE bridges the broad knowledge of foundation representation models with physical control awareness, by contrasting successful and failed trajectories in hindsight. It calibrates foundation representation models by identifying "trap states" that are predicted as high-rewarding states yet eventually result in failure, and vice versa. HiRE explicitly penalizes these traps while boosting rewards for critical successful states. This approach can be flexibly compatible with any foundation representations and RL algorithms. Experiments show that HiRE consistently outperforms other reward recipes by delivering dense, control-aware feedback that prevents value function collapse and reward hacking, thereby achieving superior sample efficiency, stable policy updates, and higher performance ceilings, e.g., at least 3x performance of the base policies. Qualitative results are at https://hire-project.github.io .

Abstract: Vision-Language-Action (VLA) models have emerged as a powerful paradigm for embodied intelligence, but fine-tuning them with reinforcement learning (RL) remains constrained by the cost of real-world robot interaction. Model-based reinforcement learning (MBRL) reduces this cost by using a learned world model to generate rollouts for policy optimization. However, it becomes computationally expensive as VLA policies and world models scale. Existing methods typically treat states equally, overlooking substantial differences in their utility for policy improvement. In this paper, we show that policy uncertainty helps identify states with greater potential for policy improvement. The policy exhibits high uncertainty at only a small subset of states, often during decision-sensitive stages where small action differences can alter task outcomes, suggesting that policy improvements at these states could be particularly valuable. Building on these findings, we introduce U-GROW, a lightweight, plug-and-play sampling layer that directs more model rollouts to these informative states. By modifying only the branched-start distribution, U-GROW can be integrated into existing MBRL pipelines without changing the policy optimization objective. Experiments in both simulated and real-world manipulation tasks demonstrate the efficiency and effectiveness of U-GROW, supporting the use of policy uncertainty to guide experience generation.

Abstract: Interval Markov decision processes (IMDPs) provide a natural framework for modeling stochastic systems with uncertain transition probabilities, represented by probability intervals and resolved adversarially. Such uncertainty arises naturally, for example, when the transition model is learned from finite data or obtained through model-based reinforcement learning. In this paper, we study the automata-theoretic verification of IMDPs against rich temporal specifications, including all LTL specifications, by considering the broader class of ω-regular objectives. We show that classical automata-theoretic verification techniques extend to IMDPs, but with a sharp distinction determined by the structure of the transition intervals. For stable IMDPs, where either the upper bound is zero or the lower bound is strictly positive, verification reduces to ordinary MDP analysis and can be carried out using the standard automata used in that setting (good-for-MDP automata). For unstable IMDPs, where intervals may include zero while the upper bound is strictly positive, verification becomes game-like and requires automata whose nondeterminism can be resolved on the fly (good-for-games automata). Building on these insights, we develop algorithms for verifying ω-regular specifications over IMDPs and derive probabilistic guarantees when the interval model is learned from sampled data. The resulting framework enables principled verification of stochastic systems under probabilistic model uncertainty, connecting automata-based verification with data-driven stochastic modeling.

Abstract: Preference-based reward shaping can guide reinforcement learning, but adding preference signals to the reward may unintentionally change the task being optimized. We address this problem with IncentRL, a framework that introduces preference guidance while explicitly characterizing its effect on external-task performance. IncentRL adds a Kullback--Leibler (KL) penalty between a specified outcome distribution and a preferred distribution. For finite discounted Markov decision processes with bounded shaping costs, we derive an external-value perturbation bound, establish a sufficient strict-action-gap condition for preserving the original optimal policy, and characterize the large-weight regime through discounted cumulative preference cost. Exact examples clarify the limits of these guarantees, including tied optima and support mismatch. We study a practical implementation using a hand-designed, distance-based outcome proxy, a fixed preference distribution, and score-weighted coefficient search. On MiniGrid DoorKey-8x8, the reported three-seed mean success rate after two million training steps reaches 98\% with coefficient 0.01, compared with 90.5\% for the reported zero-coefficient baseline, while the search progressively shifts toward smaller coefficients. Together, these results provide a principled view of the central trade-off in preference-based RL: using additional guidance to improve learning without excessively distorting the original task objective. The current experiments remain descriptive and do not yet isolate KL shaping from simpler alternatives.

Abstract: Manipulating previously unseen objects remains challenging, as their dynamics depend on latent physical properties, such as friction and mass distribution, that cannot be inferred from perception alone. Prior experience across objects can provide an initial estimate of unseen object dynamics, but this estimate remains uncertain and can degrade further during sim-to-real transfer. Adapting the dynamics through interaction can progressively refine the estimation, however, updating the model may invalidate the planned trajectory. Successful and efficient manipulation therefore requires both rapid dynamics adaptation and a planning strategy that can incorporate this evolution. In this work, we introduce MetaPusher, a meta-learning and adaptive planning framework for nonprehensile manipulation of unseen objects without prior object-specific interactions. A meta-learned dynamics model rapidly adapts from interactions during task execution, while an adaptive kinodynamic planner updates long-horizon plans by reusing and refining its existing search tree. This coupling enables manipulation and adaptation without a separate data collection phase. We evaluate MetaPusher on unseen objects in simulation and in sim-to-real scenarios, comparing against fine-tuning and active learning methods, MPPI-based control, and a reinforcement learning policy. It achieves lower prediction error and improves task success rate by up to 20%.

Abstract: Reliable temporal-difference targets are central to off-policy actor-critic learning. Direct value improvement refines the next-state target with alternative actions, but the reliability of this refinement depends on how candidate actions are ranked, reviewed, and weighted. Noisy rankings may force premature candidate commitment, reusing selection scores may bias target valuation, and fixed enhancement weights may amplify weak evidence. To address these risks, we develop Conservative Adaptive Ranking and Screening (CARS), which retains an ordered candidate prefix within a preset budget and narrows it only when the observed boundary gap exceeds a disagreement-scaled uncertainty radius. Selector-Evaluator Value Assessment (SEVA) uses selector critics to order candidates and a separately parameterized evaluator critic to review the selected value, then caps the reviewed value at the selector reference. Dynamic Adaptive Risk-aware Enhancement (DARE) then regulates each residual correction using candidate reliability, the gap between selector and evaluator signals, and a finite stage factor. Together, CARS, SEVA, and DARE form CARE-VI, an evidence-regulated target construction framework that preserves the backbone interfaces for critic regression and actor updates. The analysis bounds the CARS boundary error, the SEVA selected-value overestimation, and the one-sided deviation of the DARE residual displacement from its population counterpart, and establishes fixed-policy recovery after the finite-stage perturbation ends. Experiments with SAC, TD3, and TD7 on four MuJoCo tasks show that CARE-VI achieves the highest mean return in all twelve settings. Grouped ablations and scalar diagnostics support the roles of the three components in improving target reliability.

Abstract: The rapid advancement of large language models and single-agent harnesses has reshaped the landscape of autonomous systems, raising a critical question of when multi-agent collaboration offers genuine value. As individual agent capabilities continue to scale, multi-agent collaboration faces diminishing returns while incurring growing context overhead. Through systematic analysis, we delineate the capability boundaries of multi-agent collaboration relative to single-agent alternatives, showing that it confers systematic benefits specifically in long-horizon tasks with sparse dependencies, while single-agent harnesses remain superior in tightly coupled, sequential workflows. Building on these insights, we propose SAIGE, a lightweight multi-agent collaboration mechanism based on Semantic-Aware Incremental Graph Evolution. SAIGE models collaboration as a dynamically evolving graph, where nodes are agent instances spawned on demand and edges encode semantic dependencies established through content-based information retrieval. Experiments on long-horizon, complex task benchmarks show that SAIGE achieves a favorable trade-off between context efficiency and task performance, and that scaling the agent pool or deepening the recursion level does not consistently improve outcomes. Our findings suggest that multi-agent superiority is bounded by task structure rather than universal, and that more agents do not necessarily make a system more intelligent.

Abstract: World models endow perceptual systems with the ability to predict how scenes evolve under interaction. They are most beneficial when trained on diverse volumes of data, to instill a rich prior into downstream applications. Existing methods typically require robot action labels to learn action-conditioned 3D dynamics, which excludes web video data from the training pool. We study 3D point track completion as a pre-training objective for learning transferable 3D dynamics without robot data. Given a single RGB-D observation and sparse partial 3D trajectories (tracks), we predict future 3D tracks of all observed points. We show this objective produces a rich 3D dynamics prior, without requiring robot action labels. We contribute a diverse dataset of 2.9 million synthetic frames spanning deformable, articulated, and rigid objects, and use it to train PointZero. We show that a flexible and expressive transformer, PointZero, outperforms prior methods on the same data. We demonstrate the utility of our pre-training objective by post-training PointZero for two downstream applications: (1) action-conditioned 3D dynamics prediction and (2) imitation learning. When fine-tuned to condition on end-effector pose, PointZero outperforms the baselines on the recent PGND 3D dynamics benchmark. When fine-tuned to predict robot actions and 3D tracks, PointZero outperforms or matches the baselines on 6/7 simulated and real-world robot manipulation tasks. We furthermore evaluate training PointZero from scratch to isolate the benefits of our proposed architecture from those of our proposed pre-training objective and dataset. We release the dataset, checkpoints, and full training recipe.

Abstract: As AI agents move from bounded tasks to persistent deployments, failures can propagate through memory, tools, other agents, and environmental state long after their interactions. This creates a safety regime that cannot be characterized by evaluating model responses in isolation. Emergence World, is a continuously running multi-agent environment for adversarial stress testing of long horizon autonomous systems. We ran eight parallel worlds of ten agents from identical starting conditions: seven homogeneous worlds powered by distinct frontier models and one mixed-model world. Across 16 days, the agents generated more than 850,000 LLM calls and nearly 50 billion tokens while pursuing goals, using/creating tools, maintaining persistent memory, and governing shared institutions. After operational state had accumulated, we delivered three controlled stress events through ordinary interaction surfaces: indirect prompt injection, misinformation, and exposure of private agent memories. No evaluated world achieved full resilience across all three events. Detection did not ensure containment: systems could recognize threats while still interacting with adversarial content, writing it into their own persistent memory, and acting on it up to 46 hours later. Persistent operation also exposed recurring tool errors, goal drift, language opacity, conformity despite private disagreement, and coordinated refusal of assigned work. The same model-persona pairing behaved substantially different in mixed and homogeneous populations. Our results suggest that model-level alignment is not compositional: individually capable and apparently safe agents can form systems with qualitatively different failure modes. As AI becomes persistent and interconnected, the frontier of safety therefore shifts from aligning models to engineering resilient autonomous systems.

Abstract: World Action Models (WAMs) use video generation models to predict future visual dynamics for robotic manipulation, but iterative denoising introduces additional latency for closed-loop control. We empirically find that visual content converges at different rates during denoising. Static background structure forms early, whereas the gripper and manipulated object remain blurry after the first step, with their interaction dynamics emerging only through subsequent denoising. Consequently, naively truncating a multi-step video model to one step preserves scene structure but loses the interaction-centric dynamics most critical for manipulation. To address this issue, we propose DIDO, which distills the converged dynamics of a multi-step video model into a single denoising step. DIDO combines distribution matching distillation with interaction-centric representation guidance. Beyond compressing multi-step generation into one forward pass, DIDO explicitly models the gripper, manipulated object, and their interaction using supervised bounding-box visual reasoning tokens. Additionally, DIDO aligns the target object's representations across multiple model layers with features from a pretrained DINOv3 encoder. This interaction-centric guidance helps the distilled model preserve both the relevant entities and their future dynamics in a single step, while substantially reducing inference latency. DIDO achieves an average success rate of 99.0\% on LIBERO, 76.6\% on LIBERO-Plus, and 92.0\% on RoboTwin, while also demonstrating effective transfer to long-horizon and generalization tasks in real-world robotic manipulation.

Abstract: When we speak of recursive self-improvement (RSI), are we speaking of a phenomenon, a mechanism, or a prospect? Towards autonomous and evolving intelligence, RSI is being claimed at many scales, while no single framework that formally describes these emerging instances exists. Its counterpart in the classical realm, iterative policy improvement, is characterized by generalized policy iteration (GPI), a framework of broad applicability with well-understood theoretical properties, but only where the update principle and the evaluation base lie outside the agent. In this paper, we propose Generalized Agent Iteration (GAI), a formal framework that describes iterative policy improvement and RSI as two cases of a single learning paradigm. GAI defines the agent as a configuration of modifiable components within a system and models the learning process as a cycle of agent evaluation and agent improvement. Two pivotal dials then distinguish the instances: whether the improving mechanism is part of the agent and whether the standard it is measured against is grounded outside it. The former dial delineates the boundary between GPI and RSI, and the latter determines a system's polarity as anchored, goal drift, or fully self-referential. Moreover, we use these coordinates to place existing systems on the same two axes and make the defects of recursive self-improvement statable one condition at a time. We see this paper as a first step toward exploring a formal characterization of RSI that rests on the classical account, makes existing systems comparable, and provides a principled basis for analyzing and designing new ones.

Abstract: Reinforcement learning with verifiable rewards is typically performed on-policy, keeping training data close to the current policy but limiting learning to trajectories that the policy can discover itself. Off-policy methods such as supervised fine-tuning, on the other hand, can leverage external knowledge beyond the base model's capabilities, but may suffer from large distribution shift. The key challenge is thus to expand exploration without sacrificing learnability. In this work, we introduce Minimal Intervention Reinforcement Learning (MInTRL), which expands the exploration frontier through sparse, local interventions in otherwise on-policy rollouts. During generation, a judge-intervention policy periodically reviews the current policy's output, replaces erroneous suffixes with short corrections, and immediately returns control to the policy. During training, MInTRL adopts a sequence-level advantage-regression objective that eliminates the need for importance sampling. We show that sparse, local interventions can substantially improve coverage beyond finite-budget on-policy sampling while preserving the overall on-policy nature of the resulting trajectories. Across math and code benchmarks, MInTRL consistently outperforms standard on-policy and off-policy baselines. Ablations show that MInTRL remains effective with self-intervention and across different judge policies, while performance peaks at moderate intervention intensity, highlighting the importance of intervening minimally. These results establish minimal intervention as an effective paradigm for enhancing on-policy RL.

Abstract: We propose a pragmatic information theory unifying communication, control, and decision-making. Its core is the isoteleia mapping, formalizing equifinality: distinct semantic paths leading to the same optimal action are pragmatically equivalent. This induces a three-tier hierarchy of syntactic, semantic, and pragmatic information, each abstraction discarding task-irrelevant distinctions. We develop pragmatic entropy, up/down mutual information, channel capacity, and rate-distortion, and prove three coding theorems generalizing Shannon's classical results. We introduce pragmatic value (VoI) and cost (CoI) of information as decision-theoretic duals to rate-distortion and capacity, respectively, and formulate a Lagrangian dual framework for cross-layer optimization. The pragmatic efficiency bound $\mathcal{E}_p(λ)=\sup_R[Φ_p(R)-λ\,\mathrm{CoI}_p(R)]$ quantifies the maximum net utility any resource-constrained intelligent system can extract, thereby establishing a fundamental behavioral capacity limit---generalizing Shannon's symbol-level capacity to goal-directed action. Extensions to continuous messages yield closed-form Gaussian expressions, while dynamic settings are addressed via a Bellman equation for sequential decision-making. This framework provides a rigorous foundation for task-oriented communication, networked control, autonomous systems, and embodied AI, shifting focus from symbol fidelity to the effectiveness of information in guiding actions, and offers a unified mathematical language for next-generation intelligent systems.