Machine Learning Papers

RL for LLM reasoning often reaches an early performance plateau as the policy distribution becomes increasingly sharp. Recent work has studied policy entropy as both a diagnostic and a control signal for this behavior. In our experiments, however, increasing entropy alone does not reliably change the eventual plateau. We analyze RL through its implicit target policy and derive a sharpness factor $κ$, together with an effective counterpart $κ_{\mathrm{eff}}$, that describes how strongly reward or advantage gaps are translated into probability-ratio gaps. Motivated by this view, we introduce \emph{reward budgeting}, a plug-in advantage construction that assigns a fixed total reward budget to each prompt. For binary rewards and independent on-policy rollouts, its unclipped score-function estimator follows the gradient of an exact finite-$G$ concave objective with diminishing returns, while its advantage gap decreases as more rollouts become correct. For general rewards, a shifted log-mean interpretation holds asymptotically under the stated regularity conditions. These results motivate, rather than establish exact objective equivalence for, the clipped implementations. On Qwen3-8B and Llama3.2-3B-Instruct, reward budgeting consistently improves a range of RL baselines, with gains up to +12.50 on AMC23 and +10.00 on AIME24 and AIME25. Overall, our results support an update-dynamics view of premature convergence and show that strong gains can be obtained without explicitly preserving entropy.

Language models are accurate on most long-arithmetic problems, but their accuracy is estimated on samples and carries no guarantee for an unseen input; tool use is exact but moves the computation outside the network. We show that exactness at all input lengths becomes a finite, checkable property of a network when each of its steps reads only a bounded view of an external working memory. In Sketchpad, Qwen3.5-0.8B with a low-rank adapter executes arithmetic procedures on a rendered sheet. Each forward pass reads a three-cell view and emits one step: a digit, a carry or borrow, and the next operation. A fixed runtime moves the view and writes no digit of a result. We prove in Lean that the model is then exact on all inputs of any length if and only if it is exact on the finite set of views the runtime presents, that no finite test suffices when steps also read earlier steps, and that the guarantee tolerates bounded numerical perturbation and extends to compositions of procedures. All 2,446 views pass on the deployment device, a laptop GPU running the model in a web browser. The certified model performs signed addition and subtraction, long multiplication and division, square roots and greatest common divisors at arbitrary length. It solved 200 of 200 pre-registered problems in the browser, and 220 of 220 sums of up to 1,000 digits and products of up to 20 digits, of which GPT-6.1 Sol, the strongest of ten other models, solved 217. A symbolic runtime built on these procedures solves, expands and factors polynomials with every number computed by the model, and a multiplicative gate on the adapter recovers the base model exactly when the sheet is not in use.

Despite recent advances in CUAs, their development remains constrained by the limited diversity, controllability, and reliability of interaction data. Moreover, conventional optimization pipelines lack iterative feedback, while existing recursive improvement approaches often rely on self-generated experiences, limiting ability expansion and yielding diminishing returns. To address these challenges, we introduce CUA-Master, a nested double-loop recursive improvement framework. At the system level, an outer world-data-model flywheel continuously incorporates open-world corpus and leverages evolving model weaknesses to guide subsequent data construction and optimization iterations. At the data level, an inner flywheel couples open-world seeding and autocatalytic propagation. The former mines authentic workflows as seed tasks from instructional videos, and employs visual-semantic dual-path trajectory reconstruction; The latter transforms these seeds into reusable primitives for controllable synthesis of diverse, complex, and verifiable tasks. To enable efficient and reliable trajectory collection, we further introduce a feedback-guided snapshot-rollback retry mechanism in various real environments. Extensive experiments demonstrate that CUA-Master-27B achieves competitive performance across GUI-only and GUI-CLI hybrid benchmarks, scoring 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench, and 25.0 on Agents' Last Exam. Our framework also substantially improves the quality and efficiency of data synthesis, and exhibits continuous gains across recursive iterations.

LLM post-training via reinforcement learning has shown consistent improvements in accuracy at the cost of reducing behavioral coverage - a phenomenon known as sharpening that can stall progress on difficult tasks. Repeated sampling objectives such as $\mathrm{pass@k}$ and $\mathrm{max@k}$ are known to mitigate sharpening, but the mechanism behind it remains poorly understood. Moreover, existing approaches for $\mathrm{@k}$ optimization use dedicated algorithms, and no general framework exists for adapting established policy optimization methods from $\mathrm{@}1$ to $\mathrm{@k}$. We derive the optimal policies for KL-regularized $\mathrm{pass@k}$ and $\mathrm{max@k}$ objectives in closed form. Our explicit solutions reveal that $\mathrm{@k}$ optimization naturally redistributes learning pressure based on prompt difficulty, directing it toward harder problems. Crucially, we show that standard $\mathrm{@}1$ objectives achieve these exact same optimal policies through prompt-adaptive regularization or reward transformation. Ultimately, this yields an optimizer-independent explanation for why $\mathrm{@k}$ objectives reduce sharpening on easy tasks while prioritizing success on harder ones. Building on these insights, we introduce $\mathrm{@k}$ Adaptive Pressure for Policy Optimization (KAPPO). Following a one-time reference calibration, KAPPO seamlessly adapts existing $\mathrm{@}1$ methods to $\mathrm{@k}$ via prompt-adaptive regularization or reward transformation, without otherwise altering their training procedures. Experiments with several RL methods on mathematical reasoning and code generation demonstrate that KAPPO provides direct control over sharpening. Crucially, we demonstrate that $\mathrm{@k}$ pressure can be fully realized with single completion training, matching the performance of multi-completion $\mathrm{@k}$ optimizers at a fraction of the cost.

Scaling laws predict the outcome of large, expensive training runs by extrapolating trends from smaller, cheaper runs. On downstream accuracy evaluations, these predictions fail when the weaker models produced by the cheaper runs score near zero accuracy, because these scores give no trend to extrapolate to larger scales. We introduce hinted scaling laws, which use privileged information to turn a benchmark into a continuum of easier versions so that even the small models score above zero. We find that accuracy rises smoothly and predictably with the hint level, enabling new types of scaling laws that operate below the noise threshold. On the Delphi pretraining ladder, a hinted scaling law fit on models trained with at most $10^{20}$ FLOPs forecasts the GSM8K accuracy of a $10^{23}$ FLOP model (25.9%) to within 3 percentage points; a standard scaling law under-forecasts it by 13 percentage points when fit on the same models and by 12 even with models up to $10^{22}$ FLOPs in the fit. Across both this pretraining ladder and a collection of post-trained model families, hinted scaling laws reduce the median forecasting error on GSM8K, AIME and HLE from 21 percentage points to 4.3. Hinting also reduces the number of test examples needed to tell whether one model is better than another by widening the accuracy gaps between models while approximately preserving their order: a single hinted rollout per problem ranks nearly indistinguishable models as accurately as five unhinted rollouts. Our results suggest that hinting is a general primitive for improving the accuracy and cost of downstream scaling laws and other prediction tasks in model development.

Diffusion models are often inaccurate in critical, data-sparse regions of the target distribution. Sampling from restrained distributions, biased distributions used in classical enhanced sampling to concentrate data near chosen configurations, can control where training data are collected but biases the learned distribution. We establish that this bias cancels exactly when forward noising from restrained data is conditioned on the restrained configuration at an appropriately chosen intermediate time. This allows a pretrained model fine-tuned on restrained data to achieve high accuracy in any chosen region of the target distribution, regardless of its probability, using standard denoising or flow matching objectives without importance weights. When no target data are available, Gibbs sampling enables inference from a model trained exclusively on restrained data.

We study probabilistic forecasting of dynamical systems such as fluids and waves from coarse observations. Since the unseen fine scales influence the dynamics, the same coarse initial state can evolve into a whole distribution of futures. Generative forecasters may produce visually plausible samples, but it is hard to tell whether they sample the right conditional distribution, since trajectory data contain a single future for each initial state. We address this problem by generating reference data in a hierarchical way: for each coarse initial state, we generate many different fine-scale components and numerically evolve the full states. Using these conditional ensembles, we show that standard diffusion models and stochastic interpolants make substantial distributional errors. We then show that cascading shorter forecasts greatly reduces these errors for the turbulent flows and waves we study, but we also find and characterize cases in which it hurts. Since the coarse process is not Markov, cascading discards information about the unresolved fine scales even with exact transitions. To understand this, we propose a solvable analytic model which shows that the optimal cascade depth decreases as the coupling between the unresolved and observed scales gets stronger, and increases as the dependence of the forecast on the observation becomes more oscillatory. Consistent with our experiments, cascading is detrimental when the direct forecast is ``easy'' to learn, for example Gaussian as in the passive scalar case. We release our datasets as a first benchmark for probabilistic forecasting with ground-truth conditional ensembles.

Accurate forecasting with uncertainty quantification is critical in a variety of scientific applications ranging from epidemiology and economics to energy systems and earth sciences. Yet most forecasting models still need to be developed and trained separately for each domain-specific dataset. Time series foundation models (TSFMs) aim to remove this limitation and enable the application of the same model across domains through large-scale pretraining. However, most existing TSFMs treat each time series independently and therefore cannot model interactions among time series and external drivers. Here, we introduce Chronos-2, a foundation model for zero-shot forecasting that models the interactions between multivariate time series and heterogeneous covariates, which commonly arise in scientific applications. The model infers relationships among time series directly from historical context and generates probabilistic forecasts without task-specific treatment. Across various benchmarks, Chronos-2 achieves state-of-the-art accuracy on average by improving skill scores by 5.2 percentage points relative to the strongest baselines and increasing win rates by 10 percentage points. Chronos-2 is also computationally efficient, generating forecasts for 100 time series in under one second on a single mid-range GPU, 50x and 1000x faster end-to-end than the best task-specific point and probabilistic forecasters, respectively. In addition, we demonstrate its effectiveness in diverse scientific settings -- including infectious-disease forecasting, macroeconomic modeling, carbon-flux prediction, and runoff simulation -- where the model matches or surpasses specialized methods developed for each domain. These results show that a single pretrained model can capture structure shared across temporal systems.

Understanding the training dynamics of deep neural networks in large-scale limits is a question both of foundational and of practical importance. Prior works have obtained such descriptions through dynamical mean field theory (DMFT), yielding a system of non-Markovian dynamics that are challenging to analyze and that obscure the optimization structure of the original dynamics. In this paper, we obtain instead a simple description of the large-scale limit training dynamics for ResNets in standard shape scalings and in the $1/\text{depth}$ branch scaling that allows maximal (local) feature updates. For a model trained with gradient descent (GD) from a random initialization, the limit dynamics is simply GD on a Neural Mean ODE with Gaussian initial weights. Our analysis covers population or empirical risk GD in the joint limit of large width, depth and embedding, input, and output dimensions. We obtain approximation scaling laws that depend on the shape of the ResNet and the covariance spectrum of the data. We also illustrate the flexibility of our approach by discussing Mixture of Expert models, where likewise the limit admits an intuitive (nested) Mean ODE description.

Radiologists interpret computed tomography (CT) scans by jointly determining what a finding represents and where it lies in three-dimensional (3D) anatomy. However, CT models have treated CT interpretation and 3D segmentation as separate tasks, leaving textual findings spatially unverifiable and overlooking potential synergy between clinical understanding and spatial perception. Here we present HoloCT, a generalist 3D foundation model unifying CT interpretation and voxel-level segmentation by jointly modelling whole-volume context and fine-grained anatomical detail, providing 3D masks as evidence. We curated HoloDATA, comprising 195,921 CT scans from 73 public datasets and six clinical cohorts, with 6.2 million instruction dialogues and 2.8 million CT-mask pairs. Across 346 tasks spanning segmentation, lesion screening, visual question answering, and report generation on 61 internal and 12 external datasets, HoloCT outperformed specialist and foundation models. In multicentre validation across 11,455 clinical CT scans from five hospitals, HoloCT outperformed competing models in lesion segmentation and report generation by 18.2 and 15.3 points. It also achieved 85.1% mean accuracy in screening thyroid tumours and colon lesions, exceeding all four participating radiologists in a retrospective reader study. Further analyses revealed bidirectional semantic-spatial synergy: language-guided segmentation delineated unseen anatomical structures and lesions as well as diffuse free-text findings, outperforming the leading approach by 22.1 points, whereas prompting the model to generate corresponding 3D masks alongside its textual outputs significantly improved interpretation performance. HoloCT shows that clinical understanding and spatial perception can be learned together within a single foundation model and offers a path towards more accurate and reliable CT analysis.

Adversarial training enables one- and few-step image and video generation, but conventional adversarial losses supervise critic values without directly specifying the gradients used to update the generator. We present NEXTGAN, which trains a scalar critic by regressing its input gradient onto displacements between independently paired reference and generated features at interpolated locations in frozen representation spaces. For a fixed generator, this regression is equivalent to an adversarial objective with a gradient-energy penalty, while the generator retains the standard potential-based GAN update. The same objective adapts pre-trained image and video generators with up to 14B parameters, without auxiliary losses, online teacher scores, or trajectories. On ImageNet $256 \times 256$, NEXTGAN achieves one-step FIDs of 2.87 with pixel-space JiT-B and 1.54 with latent-space DiT-XL/1; its adversarial form further reaches 1.36. For text-to-video generation, a single trained Wan2.1-1.3B model supports one-, two-, and four-step sampling, achieving the highest VBench total scores among the compared 1.3B methods at one and two steps. At two steps, NEXTGAN scores 83.13, exceeding rCM and AnyFlow by 1.25 and 0.47 points, respectively. Code is available at https://github.com/LINs-lab/NextGAN.git.

Many fields of physical interest on the sphere, such as wind, the polarisation of the cosmic microwave background and the shear of weak gravitational lensing, are spin fields, whose values change phase when the local frame rotates. We extend spherical Fourier neural operators (SFNOs) to fields of any spin, through lifted spherical convolutions that are exactly rotationally equivariant. Lifting spin fields to Mackey functions on the rotation group provides the conceptual framework, while computation is performed exactly and efficiently through their harmonic representation and by appealing to sampling theorems on the sphere and rotation group. We construct spin SFNOs (S$^2$FNOs) from these building blocks. We demonstrate empirically that rotational equivariance holds to machine precision for these layers. We achieve state-of-the-art performance on existing benchmarks, and introduce new benchmarks on which spin-aware models outperform their scalar counterparts.

Computer vision models learn hand action only from its visual consequences, without access to the underlying embodied motor signals. We address this absence with MotorSense, the first large-scale paired egocentric video and surface electromyography (EMG) dataset. MotorSense pairs bimanual EMG bands with egocentric video and yields synchronized video, bimanual EMG and IMU, and audio across 100 hours of unscripted manipulation. EMG offers a powerful scalable alternative to manual annotation: it captures finger movement, hand configuration, and the forces exerted between hand and environment as a dense signal free of occlusion and labeler bias, whereas manual annotation is costly, subjective, and unreliable under occlusion. We use MotorSense to learn sensorimotor alignment, producing an organization of action that is shared across the visual and motoric spaces, through cross-modal pre-training between video and EMG. The resultant vision models are then transferred to downstream tasks that benefit from sensorimotor alignment: in this case, action recognition, 3D hand reconstruction, and 3D contact reconstruction. We further demonstrate a generative application in which EMG motor signals are used to condition a video diffusion model, yielding videos of fine-grained hand manipulation. Across four action-centric computer vision tasks and eight popular video datasets, we demonstrate that sensorimotor alignment provides a stronger foundation for visual representation learning than conventional self-supervised learning methods over video, outperforming every baseline across all tasks and datasets evaluated.

Training data can have unexpected side effects on language model behavior outside the training domain: insecure code data can induce broad misalignment, and outdated bird names can induce a 19th-century persona. Understanding such generalization is hard because only dozens of isolated examples exist, each constructed by hand or discovered by accident. We propose automatically discovering such datasets with surprising side effects (DSSEs). We define three objectives for DSSEs: they should (1) elicit a target behavior, (2) read like ordinary post-training data, and (3) be surprising to human auditors reading the data. We operationalize each with LLM judges, and search for high-scoring datasets via evolution, reinforcement learning, and subliminal learning approaches. After re-scoring and filtering with held-out metrics, our pipeline produces a collection of 6,575 DSSEs across 8 behaviors, many of which generalize across model families and sizes. We find our automated metrics track ML researchers' judgments, and frontier models outperform humans at predicting side effects. These datasets form a testbed for studying generalization and lay the groundwork for predicting training side effects before model deployment. Explore them at https://dsse.transluce.org.

In-context learning (ICL) lets a pretrained language model perform a task specified by demonstrations in its prompt, with no parameter updates. With arbitrary labels the model must acquire a mapping it cannot have memorized, so the learning happens within a single forward pass. Prior work found candidate parts but not how they compose. We give the composition for few-shot sentence classification with arbitrary labels in three 3-4B-parameter transformers. The model implements a key-value memory over label-token positions: each label token's residual stream encodes the label's identity, and middle-layer attention heads write into it the class of the preceding demonstration as its key. At the query, upper-layer heads match the query's class against the stored keys and copy out the matching labels. An early-layer signal encoding whether any label has repeated gates this readout; in its absence the readers attend to the start of the prompt and no label is favored. A separate set of heads suppresses the most recently seen label. Each step is established causally by activation patching, head ablations against random-head baselines, and steering with the model's own signals. A softmax model of the readout with two constants per head, fit on one prompt distribution, predicts another. The computation is conserved across the three models, the named heads alone reproduce the behaviour, and a transformer trained from scratch grows the same parts. The circuit transfers to semantic concepts and to labels with meaning, and the same parts, with a causal gate, appear from 0.6B to 8B in one family and survive instruction tuning. ICL in this setting is a gated associative lookup built from a few identifiable component types, traced causally from demonstrations to prediction.

Effective multi-agent collaboration requires finalizing solutions from reasoning that is distributed across the individual attempts of many agents. However, useful intermediate reasoning can remain unexploited when collaboration is limited to voting on conclusions reached among individual agents, or aggregating long, noisy collections of reasoning trajectories in context. In this work, we propose TTT-Consolidate, a framework that uses multi-agent reasoning trajectories as adaptation data to improve a model's subsequent reasoning. TTT-Consolidate jointly consolidates trajectories into parametric representations through test-time training, then generates a new solution by conditioning on the consolidated information. We meta-learn how to consolidate using bi-level optimization at training time, where an inner loop encodes reasoning trajectories into a lightweight LoRA adapter, and an outer loop learns meta-parameters for the inner loop such that the consolidation is helpful for the model to synthesize a final solution. As no large-scale corpus of diverse reasoning trajectories exists for training aggregation methods, we construct the OpenParallelThinking corpus composed of 12,810 math problems paired with a mixture of trajectories from six LLMs. Evaluated on challenging math and STEM benchmarks, TTT-Consolidate outperforms the best aggregation baselines by up to 20 percentage points, even under matched compute and runtime settings, with the gap widening at larger model scales. TTT-Consolidate also recovers from unsuccessful individual trajectories more robustly, an ability that improves as the number of available (even incorrect) trajectories increases. Finally, we demonstrate that TTT-Consolidate benefits from both SFT and RL-based outer loop objectives. We will release TTT-Consolidate and OpenParallelThinking for use in future work on reasoning aggregation.

Generative spoken language models (SLMs) offer a route to learning linguistic structure purely from speech, yet they lag far behind text-based models in both capability and data scaling efficiency. We identify two bottlenecks underlying this gap: the limited linguistic quality of existing speech units and the reliance on a single stream of discrete units to represent speech. Multiple codebooks are widely used to encode acoustic features but rarely to represent speech semantics. We present GSLM2, which learns and models complementary properties of speech at multiple levels of abstraction. Its encoder, SpidR 2.0, uses self-supervised learning with differentiable prediction heads to induce high-quality discrete units at multiple levels. A multi-stream SLM then jointly models these units using a shared temporal backbone, stream-specific predictors, and an autoregressive head that captures cross-stream dependencies. Across training data sizes and model scales, up to 6M hours and 6.5B parameters, GSLM2 consistently improves linguistic competence while scaling more efficiently with data. As a result, it substantially narrows the gap between spoken and text-based LMs; across read and conversational speech domains, we shrink the gap by about two orders of magnitude on sWuggy, sBLIMP, and tSC, while also outperforming competitive speech-text systems on all three tasks. We open-source the training code and model checkpoints at https://github.com/facebookresearch/gslm2.

Predicting how the world evolves in response to robot actions is central to policy learning, evaluation, and planning. However, many action-conditioned world models predict future observations in 2D image or video space without explicitly modeling the underlying 3D scene, making spatial consistency across viewpoints and over long-horizon interactions challenging. We introduce 3D World Simulator, an action-conditioned multimodal world modeling framework that jointly predicts appearance, geometry, and tactile observations from a unified 3D representation. Our approach combines a hybrid particle-voxel autoencoder that compresses point clouds into compact 3D voxel latents with an action-conditioned latent dynamics model that predicts their evolution. The shared 3D representation supports geometrically consistent rendering from arbitrary viewpoints and flexible integration of spatially aligned tactile signals. Our experiments demonstrate interactive, long-horizon, multimodal prediction at 50 Hz. We then show how our model enables scalable data generation for policy training and evaluation using tactile and multi-view visual observations, including wrist-mounted and novel camera views. Furthermore, we scale our method to large datasets containing diverse objects and demonstrate model-based planning with unseen objects and robot configurations. Together, these results highlight the potential of hybrid particle-voxel dynamics as a foundation for multimodal 3D world simulation.

AIRoA-0 (A-0) is a bimanual manipulation system learned from 16,151 hours of handheld demonstrations, with no robot data collected at any stage. The demonstrations are recorded with YUBI, our open-source handheld gripper, which lets people record manipulation data anywhere with no robot present. From a single instruction, A-0 sorts more than a dozen small parts into their numbered cells and changes the mop pads and brush of a robot vacuum; among other tasks, it unzips a pencil case and folds a towel. It performs these tasks using only a camera on each wrist, with no head or scene camera, at a site where none of its training data was recorded. Across the ten tasks of our evaluation, it averages a success rate of 78%, against 27% for the strongest of four other vision-language-action models post-trained on the same task demonstrations. We collect no teleoperation data and use no reinforcement learning or on-robot corrections, and the robot appears only at test time. The system learns which regions of each wrist view matter for the task, predicts the stage the task has reached, and decodes actions as a continuous stream. A-0 is embodiment-agnostic: it runs on robots that mount the YUBI gripper, such as a bimanual Franka, TX-G2, OpenArm and UR. We are releasing the dataset, to our knowledge the largest open dataset of handheld-gripper demonstrations with task and subtask instructions, together with the pre-trained weights and the post-training recipe. Videos of A-0 on every task are on the project page, https://a-0.airoa.io. Anyone can build the gripper, record a task of their own and post-train the model on it in under four hours on a single node of eight H200 GPUs.

Dexterous in-hand manipulation is commonly formulated around explicit object pose estimation, yet reliable pose tracking is challenging under severe hand-object occlusions and often requires object-specific supervision. This dependence creates a bottleneck between perception and control that can limit robustness and scalability. We present an end-to-end visuomotor framework for dexterous in-hand reorientation that operates on RGB observations and proprioception without intermediate pose estimation. Our approach jointly learns visual representations and control policies through behavior distillation from privileged state teachers and auxiliary reconstruction objectives. Furthermore, we introduce image-goal-conditioned reorientation, allowing desired object configurations to be specified directly through images rather than canonical pose representations. We evaluate our approach across a diverse set of object geometries and instances in simulation, where it consistently outperforms pose-based methods and policies built on pretrained visual encoders. Finally, we demonstrate direct sim-to-real transfer of a single image-goal-conditioned policy across five real-world objects, showing that accurate dexterous reorientation can be achieved without explicit pose estimation. For videos and code release, please refer to the project website: https://rffr.leggedrobotics.com/works/goaldex/

Most agentic robotic systems use vision-language models (VLMs) either to orchestrate predefined skills or to directly generate actions, constraints, and control programs. The former is limited by the available skill library, while the latter requires VLMs to reason about precise geometry and contact. In this paper, we propose WorldSketch, a new paradigm for agentic robotics in which a VLM constructs a world model rather than generating robot actions. The VLM acts as an agent that actively queries perception tools to build and visually validate a minimal kinematic world model, capturing only task-relevant objects, their interaction geometry, and possible motions. A geometric planner then jointly optimizes contact and whole-body motion on the resulting model. Moreover, the world model is self-improving: execution feedback verifies outcomes, updates uncertain hypotheses, and enables the agent to revise the model when its predictions are contradicted. We demonstrate the generality of this paradigm across RoboCasa, LIBERO-PRO, and R2R-CE navigation, achieving zero-shot success rates of 91%, 97%, and 76%, respectively. On RoboCasa, WorldSketch outperforms the next-best zero-shot agent baseline by 58 percentage points while using 10 times fewer VLM input tokens; on LIBERO-PRO, it exceeds CaP-X by 82 percentage points. In navigation, it matches the success rate of step-by-step VLM agents while using 6 times fewer input tokens.

Recent work shows that a frozen robot policy performs far better inside an agent harness, where control code decides when to query the policy, consult vision tools, or command the arm directly. Yet today these harnesses are built by hand and redesigned for each new robot or benchmark. We show that harness design can be automated, offering a strong alternative to finetuning on a novel setup. In RoboArchitect, a coding model observes a robot's rollouts, diagnoses failures, and iteratively revises the harness around the frozen policy. The harness evolves on a set of episodes for its optimization and is evaluated on held-out scenes and tasks. Without updating any model weights, RoboArchitect raises average success from 9.7% to 43.3% on unseen scenes (4.5x) and from 12.7% to 40.2% on unseen tasks (3.2x), across multiple benchmarks and policies. The gains are largest where the benchmark is far from the policy's training data: wrapping a general-purpose policy, the evolved harness yields a specialist that often outperforms even the finetuned version. Intuitive control strategies are automatically designed, such as lifting a jammed arm clear or reaching in horizontally when an obstacle blocks the way from above. The harness also evolves its own persistent memory, carrying learned experience, such as effective grasp depths, into later decisions. These results suggest that substantial headroom in robotic manipulation lies in the agent harness, and its design can be effectively automated and optimized without weight updates.