autonomous system engineers

Abstract: Joint-embedding predictive architectures (JEPAs) learn to predict representations rather than pixels, offering a compact way to model how visual scenes evolve. We ask whether this predictive objective improves closed-loop control when a vision-language-action (VLA) policy already uses V-JEPA features. We call our JEPA-conditioned flow policy family SaiVLA-JEPA. It combines a vision-language model with a diffusion transformer action expert and receives features from a frozen V-JEPA encoder. Two lightweight auxiliary heads predict future representations during training and are omitted at inference, leaving the action sampler unchanged. We compare policies without JEPA features, with current JEPA conditioning, and with additional future supervision on 10,030 paired LIBERO-Plus episodes. Current JEPA conditioning reaches about 82.5% success versus 71.6% without JEPA, while future supervision provides no detectable aggregate gain at the evaluated training budget. We separately evaluate SaiVLA-JEPA-TCA, which adapts latent context, and SaiVLA-JEPA-TPA, which adapts selected parameters. The context-retry gain is a system-level result under a larger attempt budget; parameter adaptation shows no consistent advantage over the frozen policy and target controls. These results motivate assessing latent-prediction objectives through paired closed-loop rollouts rather than latent loss alone.

Abstract: Error-targeted learning gives a legible way to improve a model: identify specific weaknesses of the model, train the model to fix these weaknesses, and check whether the weaknesses are fixed. We study how to apply error-targeted learning to terminal agents, which solve long-horizon tasks specified in natural language by interacting with a command shell. Existing error-targeted methods are designed for single-turn tasks, and existing analyses of agent failures are used to evaluate and debug agents rather than to train them. We propose Misbehavior-driven RL (MiRL), which first profiles the errors of a policy to find generalizable misbehaviors: kinds of wrong steps that recur across many environments. MiRL then trains the policy with RL, starting rollouts at the steps where these misbehaviors occur and rewarding a rollout only if it passes the tests and avoids the misbehavior. On Qwen3.5-4B, eight misbehaviors account for 52.1% of wrong steps. After training with MiRL, the policy repeats the measured misbehaviors 43% to 57% less often on average on unseen prefixes. MiRL outperforms TMax-4B on Terminal-Bench Lite (51.8% vs. 47.4%) and matches it on Terminal-Bench 2.1 (21.3% vs. 20.6%), while using 35% of its rollouts. These results show that error-targeted learning can be applied to realistic, long-horizon terminal tasks.

Abstract: Recursive large language model (LLM) agents solve long-horizon tasks through decomposition, execution, and environmental feedback. Yet after producing an atomic subtask, they often invoke the same large model again to translate a familiar instruction into an executable action. We observe that high-level planning is task-specific, whereas leaf-level action grounding recurs across tasks and can be checked against environment-admitted actions. We introduce LeafSkill, an external reuse layer that retrieves semantically related grounding experience, binds action templates to current objects and admissible actions, and uses a lightweight 1.7B selector to choose a bound action or defer the unchanged subtask to the original agent. Skills remain outside the model parameters, allowing a frozen selector to use post-training additions. Under lexical and syntactic paraphrases, exact matching retrieves none of the reference actions, while semantic retrieval preserves 96.0% and 97.0% candidate coverage; the frozen selector adopts 97.34% of unseen skill types. On an independent confirmation set of 200 Robotouille tasks, LeafSkill completes 137 tasks versus 121 for ReCAP (68.5% vs. 60.5%, exact McNemar p=.0259). On the 106 tasks completed by both, it reduces Qwen3.5-27B tokens by 14.42% and calls by 14.08%. A separate held-out repeat yields a consistent 9.0-point success gain and 18.98% token reduction. Reusable leaf grounding can thus improve task completion while reducing host-model inference without replacing high-level planning.

Abstract: Frontier black-box Large Language Model (LLM) agents are proficient at long-horizon reasoning and precise action generation. However, when it comes to a specific agentic task, they have no internalized knowledge of how that particular environment unfolds. A world model offers a natural remedy, providing the agent with environment specific knowledge by foreseeing how the environment will respond to the action it is about to take. Prior world models for LLM agents mainly predict next observations at each step, which is costly and often uninformative because individual observations provide limited task-relevant guidance and not every step requires call for foresight. This raises a key question: what a world model should predict and when should it be invoked for agentic tasks? Our pilot study suggests that world model guidance is most useful when it anticipates pitfalls within a bounded multi-step future. Motivated by this, we introduce Pitfall Anticipation World Model (PAWM), which selectively intervenes when a potential pitfall arises and predicts the impending failure to steer the agent away from it. PAWM is trained in two stages: Supervised Fine-Tuning (SFT) on synthetic data labeled with agent pitfalls, followed by GRPO co-training with a black-box LLM to further optimize it in the target environment. PAWM improves Claude Sonnet 4.5 performance by 24.3% on ScienceWorld and 7.7% on WebArena-Lite, outperforming a range of baselines. These results demonstrate that PAWM provides an effective form of world model guidance for black-box LLM agents.

Abstract: Generative policies have achieved strong performance in imitation learning tasks, but their success is often constrained by the limited availability of expert demonstrations. Although preference-based fine-tuning can leverage suboptimal rollouts to improve performance, existing approaches heavily rely on expensive human annotations. In this work, we introduce Progress-Guided Preference Generation (PGPG), an unsupervised preference alignment framework that eliminates the need for human-in-the-loop supervision. Our method leverages ordered coverage alignment to measure sequential progress relative to a small set of expert demonstrations, enabling the automatic extraction of pseudo-preference labels from autonomously collected rollouts. Furthermore, we introduce key enhancements to improve the reliability of these progress-guided preferences, enabling effective fine-tuning of pretrained generative policies without requiring manually designed reward functions. Extensive experiments across a suite of challenging long-horizon manipulation tasks demonstrate that PGPG outperforms standard supervised fine-tuning baselines. Moreover, our conceptually simple framework is broadly applicable across multiple classes of generative policies, demonstrating strong versatility when integrated with both diffusion and flow-matching policies.

Abstract: An agent can improve its behavior within an unresolved problem by changing how it acts and which experience it retains. We propose MMLA-RSI, a recursive self-improvement scheme with two implementations in Memory-Mediated Learning Architecture: Reasoning-Time Training (RTT) changes the policy state $Φ$, while Reasoning-Time Memory Adaptation (RTMA) changes authoritative memory $M$. Each route also makes its adaptation method an object of improvement. A current numerical updater trains the policy proposing its successor; a current memory manager chooses the records available to the process proposing its successor. An independently admitted successor takes over the same role in the next generation. We give bounded executable rule languages, shared task/proposal interfaces, and one algorithm for either route or their composition. The RTT instance changes scalar schedules in a clipped-momentum update. The RTMA instance chooses NULL or a complete-row write with possible eviction. A fixed interpreter, trusted committer, and independent evaluator bound what can change. State versions, read closures, and conditional properties describe causal execution, accepted transitions, and finite work. Existing numerical-component studies motivate the construction. In four-candidate program selection, retaining LARC feedback updates reduces expected error by 24.65 and 36.65 percentage points for two trained initializations. In chronological continuous-integration replay, feedback binding is positive while retention increases loss. Completed SQL studies improve substantially over an unadapted substrate but do not establish the required additional online benefit. These results characterize their declared components. The two recursive routes and their joint realization are specified here; their beneficial recursive chains and discovery efficiency remain to be evaluated.

Abstract: Reinforcement learning (RL) has emerged as the primary paradigm for optimizing Large Language Model (LLM) agents, yet efficient critic-free algorithms (e.g., GRPO) suffer from a fundamental uniform credit assignment dilemma that broadcasts trajectory-level advantages equally to all actions. Our analysis reveals that this uniformity inherently reinforces structurally invalid actions in preferred trajectories and ignores the downstream cascading influence of early pivotal deviations in unpreferred ones. To resolve this without introducing critic overhead, we propose Progress-Aware Advantage Shaping (ProAS), a plug-and-play framework that leverages a dynamically evolving Agentic Reference Graph (ARG) to achieve fine-grained credit assignment via Masking Unreasonable Incentives (MUI) and Asymmetric Penalty Redistribution (APR). Extensive experiments across Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct demonstrate that ProAS universally elevates mainstream group-based baselines to new state-of-the-art performance, achieving peak success rates of 91.9\% and 94.3\% on ALFWorld, alongside 66.7\% and 70.8\% on WebShop, respectively. Furthermore, it accelerates training convergence by up to $2.44\times$ while maintaining strong robustness against imperfect or distilled ARGs. Codes: https://github.com/LINs-lab/ProAS.

Abstract: Large language model (LLM) agents excel at complex, multi-step tasks but suffer from catastrophic forgetting during continual learning--acquiring new skills often degrades previously mastered capabilities. Existing post-training paradigms face a fundamental trade-off: Supervised Fine-Tuning (SFT) is sample-efficient but prone to severe forgetting, while Reinforcement Learning (RL) preserves prior capabilities but converges slowly without expert guidance. To bridge this gap, we introduce Trigger-based Supervised Reinforcement Learning (TSRL), a framework that integrates the strengths of both paradigms through an adaptive, state-dependent expert supervision mechanism. TSRL employs RL as the optimization backbone, activating SFT selectively on "hard states" where the policy fails. Theoretically, we show that restricting supervision to failure states derives an upper bound on the cumulative parameter updates, thereby reducing forgetting compared to full SFT. Empirically, TSRL achieves consistent gains across three agent benchmarks (ALFWorld, WebShop, and Deep Search) at both 7B and 32B scales. In continual learning scenarios, our method retains up to 9.6 points more capability on prior tasks than SFT->RL baselines, while matching or exceeding target-task performance.

Abstract: Test-time compute (TTC), which scores multiple candidate actions at inference time, is a powerful way to improve the execution performance of vision-language-action (VLA) models, but existing methods keep the policy frozen, so the gain must be recomputed at every execution. We propose Verify, Act, Distill, a framework that consolidates the gains of TTC into the VLA itself. A fixed verifier scores candidates generated by a flow-matching VLA, the weighted aggregate is executed, and the observations and aggregated chunks from successful adaptation episodes are distilled directly into LoRA parameters with the standard flow-matching loss. This reuses supervision that TTC has already produced during execution, without additional reinforcement learning or rollouts dedicated to teacher generation. On the 40 LIBERO tasks with a SmolVLA 5k checkpoint, distilling from only 10 episodes per task raises the mean success rate of verifier-free single-candidate generation from 54.1% to 70.6%, retaining about 74% of the improvement obtained by aggregating eight candidates. Applying the same aggregation to the distilled policy further raises the success rate from 76.4% to 82.3%, with a particularly large gain on LIBERO-Long (44.0% to 57.8%), and the single-candidate policy also improves on LIBERO-plus, which contains distribution shifts. These results show that using verification as a learning signal, rather than only as an execution-time selector, both retains the improvements of TTC and re-amplifies the effect of subsequent TTC.

Abstract: Language model agents increasingly operate over long interaction streams, requiring online adaptation to distribution shifts through continual learning. However, existing agent harnesses typically either evolve their states only through forward information flow or rely on additional replay or validation data for harness optimization. To address this, we propose ReHarness, a recurrent harness framework for agentic continual learning that complements forward read and write with backward revision under a strictly online streaming protocol. We ask what information could have been captured differently in the harness state to better inform the current instance in the stream. Specifically, after each instance, we leverage the resulting trajectory as feedback to revise the reflections retrieved for that instance, thereby propagating feedback backward through the textual harness state. Across three continual learning benchmarks and three language model backbones, ReHarness consistently outperforms strong raw-context baselines, memory- and reflection-based methods, and prompt optimization approaches. Further analysis shows that ReHarness produces source-grounded and meaningful revisions, corrects overgeneralized or outdated experience under distribution shifts, and improves both stability and plasticity during continual learning.

Abstract: Reinforcement learning (RL) is the central training paradigm for advancing large foundation models towards self-improvement. This report introduces the MiMo-V2.6 series, an omni-modal family that pushes the frontier of model intelligence by scaling RL compute. Prior to RL, we conduct mid-training on a broad multimodal corpus to provide ample exploration space, and build a solid infrastructure on the pretrained hybrid-SWA architecture to support subsequent scale-up. We scale RL compute along three dimensions: (1) larger batches and higher throughput, with an asynchronous training that consumes 1,568 samples and 2.7-3.7B tokens per step at context lengths of up to 1M; (2) more diverse and complex environments, spanning code, general, visual, and cyber domains under a mixture of agent harnesses; and (3) more grader compute, via groupwise agentic grading that yields more accurate reward signals for long-horizon tasks and steers the model towards shorter, more token-efficient solutions. To keep training stable at scale, we freeze the MoE router and establish a multi-layer defense against reward hacking. We further build infrastructure for mixed-task agentic RL, including a unified trajectory representation, high-concurrency multi-framework rollout, decoupled control and data planes, and training-inference consistency. We open-source the training dynamics, RL environments, and RL framework to facilitate reproduction and further research on scaled RL and model self-improvement.

Abstract: We study learning optimal policies in average-reward multichain Markov decision processes (MDPs), where the optimal gain may depend on the initial state and recurrence structures vary across policies, creating challenges for reinforcement learning (RL) methods. We propose an asynchronous value-iteration-based RL algorithm that requires no model knowledge beyond the MDP's transition graph and leverages Bather's decomposition to hierarchically partition the state space into communicating subsystems and transient states. This decomposition induces a recasting of the global decision problem into structured subproblems, which our algorithm exploits. We show that the algorithm converges to the optimal gain and produces gain-optimal policies after finite time. Building on this base algorithm, we develop two further algorithms: one approximately solves the multichain average optimality equations to obtain near gain-optimal policies, and another targets near bias-optimality by approximating the optimal bias function and solving an induced average-reward multichain MDP using the base algorithm. We provide almost-sure convergence guarantees for all three algorithms and empirically compare their tradeoffs, showing that the latter two also consistently improve transient performance relative to the base algorithm. To our knowledge, these are the first essentially model-free average-reward RL algorithms for general multichain MDPs without reductions to discounted problems.

Abstract: Vision-language-action (VLA) policies typically feed dense visual patch tokens into a language-action backbone, preserving scene context but offering no explicit mechanism to regulate how strongly different visual tokens influence policy computation. We introduce DIVA, a Dual-Space Intent-Aware Visual Attenuation module with an anchor-then-attenuate design. DIVA combines high-level task intent with low-level visual evidence to estimate patch-wise relevance anchors, then applies them in two complementary spaces: it reweights projected visual tokens before backbone entry and persistently attenuates low-relevance visual states within the backbone. DIVA preserves the full visual token sequence and requires no external grounding supervision. On LIBERO, DIVA improves OpenVLA-OFT from 96.6% to 98.0% average success and raises its zero-shot LIBERO-Plus score from 69.6 to 72.6. Real-world experiments further show consistent gains under task-irrelevant visual perturbations, supporting the robustness of intent-aware visual attenuation beyond simulation.

Abstract: Action-chunking policies predict multi-step control sequences, but a fundamental question remains: how much of a predicted action chunk should be committed before replanning? Existing systems typically execute a fixed-length prefix, implicitly assuming that the same execution horizon remains trustworthy across states. Some adaptive methods estimate this horizon from the similarity or stability of predicted actions. However, different actions may lead to the same successful outcome, whereas similar actions can produce different futures, suggesting that commitment should be determined by agreement among imagined futures rather than by similarity in action space. To this end, we propose Consequence-Aware Adaptive Action Chunking (CA$^3$C), an inference-time framework built on a simple principle: commit while imagined futures agree, and replan when they diverge. Without modifying or retraining the base policy, CA$^3$C uses an action-conditioned world model to imagine the future consequences of multiple candidate action chunks under the same sampling noise. Using these imagined consequences, we formulate execution-horizon estimation as a Bayesian change-point inference problem and select the execution candidate through future consensus. Across multiple simulation benchmarks and real-world robot manipulation tasks, CA$^3$C consistently improves diverse action-chunking policies, achieving up to a 71.8% relative reduction in failure rate over the corresponding base policies.

Abstract: Memory self-evolution uses task feedback to iteratively improve executable memory programs that store and retrieve information from past interactions. Existing approaches typically adopt holistic evolution, deriving revision directions from mixed feedback and judging progress by overall performance. This can obscure optimization directions and hide capability-specific gains offset by regressions elsewhere, leaving promising directions underexplored. We introduce capability-driven evolution, which extends search guidance from overall performance to individual capability dimensions, preserving promising revisions and expanding exploration beyond the boundaries of holistic evolution. We propose PrisMem, which uses dependency-aware capability selection to prioritize targets with potential cross-capability benefits and history-guided diagnosis to refine capability specialists. Trace-guided integration compares evaluated programs on paired differential cases, using their behavioral differences to consolidate complementary gains into a unified memory program. Experiments show that PrisMem outperforms the strongest baselines by 10.54 and 7.83 percentage points on BEAM-1M and LongMemEval-M, respectively, demonstrating its effectiveness on million-token histories.

Abstract: Vision-language-action (VLA) models often struggle to generalize to skill combinations absent from their fine-tuning demonstrations, even when every constituent skill has been demonstrated. We focus on a vision shortcut as one failure mode: during fine-tuning, visual observations can serve as a proxy for the instruction, so a policy may execute a demonstrated combination associated with similar observations rather than the instructed combination. This motivates training with counterfactual pairs formed by holding a demonstration observation fixed while changing the instruction to specify an undemonstrated combination. These pairs, however, lack corresponding demonstrated action targets. Crucially, the currently required skill has already been demonstrated, but actions from those executions cannot serve as direct targets because the same skill can require different actions across observations. We propose CRAFT, which transfers supervision from demonstrated executions of the required skill to counterfactual pairs using skill representations that can be reused across executions of the same skill. Across three VLA models and two simulation benchmarks, CRAFT improves success on undemonstrated combinations while maintaining high success on demonstrated ones; it also improves compositional generalization on a real robot. Project website: https://taegeunyang.github.io/craft/

Abstract: Benchmarks, audits, and agent protocols describe performance, permissions, and repair, but not how observed evidence should change an agent's authority during a consequential task. We call this the assurance-transition gap. We propose a Runtime Assurance Contract (RAC), a policy-level formal schema binding autonomy boundaries, component eligibility, evidence state, transition policy, human-review capacity, and non-compensatory gates. Under RAC, soft metrics may inform routing, whereas a failed or unknown mandatory gate forces retry, switch, escalation, deferral, or stop; aggregate performance cannot authorize action. We define the contract, an evidence record, a permission rule, and five invariants, and illustrate them in clinical, industrial, and judicial failure probes. We then report a deterministic failure-injection study in agentic coding: 280 constructed cases evaluated by a gate conjunction, a score-only rule, and a restricted protocol baseline. At the published example weights and threshold, the score rule admits 80 of 100 block-required injections and all 40 review-required injections. Tuned in hindsight, it matches the conjunction on this corpus. For positive weights, a positive threshold, binary risk signals, zero-signal controls, and an injected case firing each signal alone, we show that exact agreement holds if and only if the threshold does not exceed the smallest weight. A separate set of 18 hand-authored traces checks version-pinned evidence and review transitions against simpler policy variants. In a further prospective synthetic holdout of 24 episodes, two blinded LLM judges assign identical labels to all 72 action attempts; RAC and a separately implemented full stateful baseline both match these labels. These studies test mechanisms on synthetic cases; they establish neither deployed safety nor cross-domain effectiveness.

Abstract: Residual policy adaptation provides a lightweight way to modify a strong reference policy, but a shared scale and a fixed subgroup partition can hide heterogeneous degradation and become fragile when the deployment mixture of information states changes. We introduce RSD-Poker, a structure-adaptive and shift-robust certification framework that freezes a bank of residual families and scales, learns a policy-visible partition on an independent structure split, and freezes that partition before calibration labels are joined. Each candidate-group pair receives a weighted simultaneous upper certificate for anchor-relative risk and a lower certificate for weak-response utility. A robust group-to-candidate map is then selected over a predeclared uncertainty set of deployment group proportions. Under independent calibration units drawn from each frozen group's law, a candidate bank and partition fixed before calibration, and invariant within-group conditionals, the selected map satisfies its declared mixture-robust risk budget and utility certificate with probability at least $1-ζ_{risk}-ζ_{util}$. The information contract supports both a teacher-backed transform and a teacher-free observation-only student. The retained deterministic 24-state audit remains an exact replay diagnostic: empirical-zero selects $α=0.08$, raising the weak-response proxy from 4.2082 to 4.2889 with $0/12$ held-out threshold crossings. On stratified held-out states, the learned-partition dual selector raises weak utility from 4.4074 under global dual certification to 4.4936 and lowers held-out violation from 0.0215 to 0.0078; its mixture-robust variant reaches violation 0.0059. Across five observation-only checkpoints, risk-calibrated residuals attain weak utility $4.3659\pm0.0177$ and violation rate $0.0178\pm0.0057$.

Abstract: When safety-critical systems are developed from a known internal specification, their correctness can be established by model checking. In the frequent case where such a specification is unknown or inaccessible, runtime verification presents an attractive alternative, e.g., to ascertain that autonomous and agentic systems as well as business processes satisfy desirable properties and/or comply with safety requirements. In this paper we study anticipatory monitoring, an advanced form of runtime verification, where the monitoring state is determined by both the trace prefix seen so far, and all its possible finite-length, future continuations. We focus on monitoring linear-time properties that may involve arithmetic constraints. Automata-based approaches, the de-facto standard in this setting, are notorious for their computational complexity. We propose an alternative approach based on progression and LTLf satisfiability checking, for both propositional and arithmetic settings. We experimentally compare the automata- and progression-based approaches, and a third method that combines the two. Our experiments suggest that the progression-based approach often succeeds in producing a verdict when the automata constructions do not terminate, especially for the arithmetic setting. For the propositional setting, the combined technique provides a good tradeoff.

Abstract: Agentic memory systems reuse past experience to improve future performance, yet most existing designs curate memory at write time: once a task is completed, its trajectory is distilled into a fixed artifact, such as a reflection, workflow, skill, or reasoning strategy, that is later retrieved by similarity. This forces the system to decide what is worth remembering before the future query is known, irreversibly discarding information and producing a query-independent summary that must serve many possible downstream tasks. Learning such a write-time curator is also difficult because the value of a storage decision may only become apparent when a relevant query arrives, potentially many tasks later, creating a long-horizon credit-assignment problem. We instead retain raw trajectories and defer curation until read time, when the current task is known. Given the retrieved traces and the new task, a memory curator synthesizes a compact, task-adaptive payload tailored to the immediate need. Because this payload is consumed on the same task, the curator can be trained directly from immediate task success, avoiding delayed utility signals and the need to artificially group related tasks. Across ALFWorld, WebShop, and $τ^2$-bench, our Just-in-Time Memory (JitMem) consistently outperforms no-memory agents as well as heuristic and learned write-time memory methods, improving over the strongest baseline by 16.2, 16.3, and 3.9 absolute success-rate points, respectively. Notably, even an untrained curator is already competitive with or surpasses these baselines, showing that task-adaptive read-time curation itself is a major source of the gain; training the curator further compounds the improvement.

Abstract: The dominant method of processing sonar data is using image-based representations, requiring the preprocessing of image data on autonomous systems. We propose an alternative data processing method for remote sensing applications via the use of data in Comma-Seperated Value format. Experimentation on our alternative approach shows a reduction of processing time by 91.18%, an improvement in accurate object detection by Machine Learning, and an increase in SNR (Signal-to-noise ratio), PSNR (Peak signal-to-noise ratio), and other evaluation metrics.

Abstract: Deep reinforcement learning (DRL) has achieved strong performance across a wide range of continuous-control problems. These continuous-control policies, however, are often defined as direct mappings from an observed state to an action or action distribution, requiring a single feed-forward network to construct an optimal control decision in one pass. While effective, this formulation leaves little opportunity for the policy to reconsider or progressively improve an action once an initial prediction has been formed. In this work, we explore an alternative approach: rather than learning only to directly predict an action, can a policy learn to iteratively improve one, and can this iterative process provide advantages during policy learning? We introduce Iterative Action Refinement (IAR), an iterative action-construction method that constructs control actions through a sequence of learned residual corrections. Starting from an initial proposal, a shared refinement network repeatedly conditions on the observed state and the current action proposal, allowing each refinement step to revise the action constructed by preceding steps. The final refined proposal is then used to determine the action executed by the agent. We integrate this iterative action-construction mechanism with Proximal Policy Optimization (PPO), yielding REFINEPPO. We evaluate REFINEPPO across 14 benchmark control tasks, complemented by controlled ablations of refinement depth and update schedules and analyses aimed at understanding why iterative refinement is effective. Across these environments, REFINEPPO matches or exceeds the performance of standard PPO while demonstrating faster convergence on several tasks.

Abstract: In recent years, artificial intelligence has made extraordinary progress thanks to large-scale models capable of generalization and the generation of complex outputs. However, transferring this potential into embodied agents reveals a significant limitation: the most advanced systems rely on pre-existing datasets and human feedback strategies that are powerful but insufficient in dynamic or unknown contexts. To adapt, an agent must acquire knowledge through direct interaction with its environment. One strategy to address this challenge involves introducing higher-level mechanisms, such as intrinsic motivations, which leverage curiosity and competence, to guide exploration and learning in complex environments. While this flexibility expands autonomy, it complicates the task of ensuring agents remain aligned with human goals. Alignment, already a challenge for artificial systems in general, becomes even more complex in unstructured and dynamic contexts where predefined rules prove insufficient. To be effective and adaptable, norms must be rooted in experience through an epistemological process that starting from simple, situated principles allows for the gradual construction of more complex rules through experience, autonomous learning, and cooperation with other moral agents. Similarly to children learning social norms by exploring their environment and participating in collective practices, artificial agents must also be educated toward alignment. Following Dennett, the status of a moral agent is not innate but is attributed gradually based on the ability to responsibly manage increasing degrees of freedom. From this perspective, the regulatory sandboxes can be viewed as pedagogical environments for AI: dynamic spaces where alignment develops as a formative process, progressively shaping autonomous behaviors through interaction and cooperation in scenarios of increasing complexity.

Abstract: Agentic AI systems built on large language models fail in two persistent ways that scaling does not fix: they break under distribution shift, and they cannot explain the decisions they make. We argue these are co-symptoms of one structural deficiency in the data lifecycle that governs how agents are trained, evaluated, and deployed. Observational interaction logs record what an agent did, not what it would have done otherwise. They encode spurious correlations without controlled variation, so they lack the counterfactual structure needed to separate causal signal from coincidence or to validate an explanation. No model-centric method can recover invariances the data never contained. We present Data-Centric Anchoring: robustness and interpretability should be engineered into the data environment, not extracted from models after training. Our central contribution is the Data-Centric Agentic Loop, a four-stage framework of Curate, Augment, Constrain, and Attribute. The ordering is structural, not stylistic. Curation precedes augmentation because generative models amplify whatever bias they are trained on. Augmentation precedes constraint because invariance objectives are vacuous without variation across environments to be invariant to. Attribution closes the loop, converting observed failures into targeted data interventions for the next iteration. Each stage manufactures the preconditions of the next, which makes the loop self-correcting rather than merely sequential. We ground the framework in a failure-driven taxonomy that links four core failure modes to the data lifecycle: spurious feature reliance, distribution-shift fragility, uncertainty miscalibration, and explanation unfaithfulness. We close with the limits of this approach and the open problems that stand between it and practical deployment at scale.