A driving foundation model must learn reusable decision knowledge from rare critical events and apply it under strict runtime constraints. We introduce DriveFoundry, a closed-loop framework improving iteration efficiency under data and computation budgets by coupling structured causal data refinement with adaptive inference. The Data Engine builds traceable scene-candidate records linking scene evidence and factor-decision relations to candidate trajectories and outcome assessments. Candidate comparisons and counterfactual annotations clarify how scene factors affect behavioral feasibility and preference. Verified model failures guide VLM-assisted annotation and expert correction, refining training data and annotation systems. The model engine, Xiaomi-Drive-o1, unifies Perception, Causation, Action, World, and Value within a shared vision-language backbone to enable cross-task knowledge transfer. Block-wise latent refill passes candidate-specific evidence and predicted visual consequences between task blocks, connecting scene understanding, planning, and evaluation. A learned value-of-computation head selects tasks and allocates latent tokens based on expected benefit and cost, adapting inference to each decision. On PhysicalAI, the 9B model achieves the lowest ADE@1 across evaluated splits and horizons among tested models. In AlpaSim, it outperforms Alpamayo 1.5 across all metrics and achieves lower close-encounter and off-road rates than the 34B Alpamayo 2 Super, while trailing in progress and AlpaSim scores. For the 0.8B variant, joint training reduces 5 s minADE@6 by about 54% without altering the inference path or token budget. Adaptive inference matches full reasoning performance using only 28% of tokens. Real-vehicle tests confirm runtime feasibility, supporting a unified paradigm for learning from critical experience and dynamically allocating computation.
Which representation should be probed in a transformer-based foundation model? Recent work shows that the standard practice of probing the final layer is often suboptimal, as intermediate layers can provide better features. Yet, representations are still largely studied along depth alone, treating each transformer block as an atomic unit. In this work, we analyze transformer representations layer by layer, module by module. Through more than 144,000 linear probing evaluations of vision foundation models (VFMs) and large language models (LLMs), we show that layer and module choices should be considered jointly: the best layer-module pair yields substantially larger gains over the final-layer residual stream than the best layer with the residual stream fixed. Moreover, module performance can vary substantially across layers, showing that useful representations cannot be characterized by depth alone. Building on these findings, we evaluate simple fixed probing strategies: the final-layer feedforward activation (Act@100) for vision and the second normalization layer at 60% depth (LN2@60) for language. These strategies improve over the probing baseline in a majority of evaluations and recover, on average, 42% and 86% of the gains from probing the best layer-module pair for vision and language, respectively. For VFMs, these gains increase with the transfer difficulty between pretraining and downstream data.
There is a well-established body of work demonstrating that individual machine learning (ML) algorithms can be manipulated or deceived. However, in many real-world systems, multiple ML algorithms are composed to enable automated decision-making, particularly in Cyber-Physical Systems (CPS). Often such systems use a composition of vision, planning, and other ML models to affect their internal state and ultimate take actions in the physical world. While there is a huge literature on the (in)security of ML algorithms, it is not immediately clear how an attacker might leverage their ability to trick individual learners when attacking the composition as a whole. We develop methods for creating an attack policy (a mapping from state to attack) for a data-manipulating attacker affecting a CPS or other state-machine-based AI compositions. We describe our framework, a basic threat model and corresponding methods for computing an optimal attack policy, as well as planned extensions. The goal of this work-in-progress paper is to stimulate conversation and forge new collaborations regarding the security of compositional learning.
AdamW is ubiquitous in deep learning, yet its behavior remains poorly understood. We analyze its dynamics through the lens of dynamical systems and show that AdamW admits an implicit fixed-point objective: its fixed points coincide with the stationary points of a constrained and regularized optimization problem. However, not all of these fixed points are stable under AdamW's dynamics, and stability depends sensitively on curvature, weight decay, and momentum parameters. Even in one-dimensional settings, equilibria may be unstable, and trajectories can exhibit persistent oscillations rather than convergence to them. We further extend the analysis to higher dimensions, deriving sufficient conditions for local stability of the continuous-time dynamics, and propose a convergent modification that preserves AdamW's fixed points. These results clarify what optimization problem AdamW is associated with, when its convergence can be expected, and how its dynamics could inspire more reliable optimizers.
Low-label transfer is often decided by a practical question: after a representation looks useful, how many labels are still needed to learn the downstream boundary? Transferability and probe-quality scores characterize global alignment, likelihood, margins, and calibration. Boundary-near target mass complements these diagnostics by measuring the ambiguous regions where scarce labels are most costly. We propose effective boundary complexity, a label-budget diagnostic for frozen representations and linear probes. It measures mass in a thin strip around the learned boundary and normalizes it by class balance, so dense ambiguous regions and predicted-class imbalance are both counted. We prove that, under a smooth representation map, density change and tangential surface change cancel at first order, leaving boundary-normal stretch or compression as the governing factor. This yields a cached-feature estimator that fits a probe on one split and counts independent held-out features near its boundary. Experiments support the diagnostic at three levels. Controlled transformations match the predicted law across 74 maps with 0.38% median relative error. On real frozen features, normal-direction changes move both measured boundary mass and label demand, while tangent controls preserve the first-order coefficient. A differentiable surrogate reduces measured complexity by 16-36% at matched top-1 utility. Finally, CIFAR-100 and Tiny-ImageNet experiments support boundary mass as a second-stage diagnostic for near-tied representations. Five-way experiments show strong OvR-interface backbone-ranking agreement (Spearman 0.872), while acquisition experiments show high correlations between initial boundary mass and realized label budgets across three sampling policies (0.779-0.830). Code is available at https://github.com/qqaqq-code/boundary_mass.
We propose a music inpainting technique combining similarity matching and a diffusion-based generative model to replace large segments of audio recordings with suitable alternatives. The proposed hybrid method, Similarity-Guided Diffusion Posterior Sampling (SimDPS), retrieves candidate replacement segments from the wider signal based on contextual similarity and uses them to guide the diffusion process toward coherent reconstructions. Being model-agnostic, the framework is compatible with any pre-trained diffusion model operating in either the waveform or latent domain. Objective evaluations demonstrate that SimDPS outperforms widely adopted RePaint and DPS methods across acoustic features and neural-embedding distance metrics. Subjective evaluations on multi-genre music datasets confirm that SimDPS enhances perceptual plausibility compared to previous diffusion-based methods and demonstrates robustness to poorly matched similarity candidates. These findings pave the way for improved music restoration and context-aware audio editing.
World models learn internal simulators of environment dynamics for prediction, planning, and control, yet learning them from limited, noisy, and partial observations remains challenging. While much progress has focused on pixel-based observations, many physical systems instead provide vector-valued measurements that reveal only part of the underlying state. We introduce K-PWM, a probabilistic world model for prediction and control under partial observation. K-PWM couples Koopman-bilinear latent dynamics with a nonlinear observation decoder and performs probabilistic latent-state inference through the learned model via extended Kalman filtering and smoothing, without a separately trained inference network. Its bilinear structure enables exact propagation of the maintained Gaussian belief through the latent transition and efficient rollouts for uncertainty-aware model predictive path integral (MPPI) control. For training K-PWM, we present a generalized expectation conditional maximization (GECM) procedure with closed-form updates for the generators, noise covariances, and latent prior, complemented by gradient-based decoder learning. A Koopman-consistent initialization selects the latent dimension from data and initializes the action-dependent dynamics. With limited training data, K-PWM achieves lower long-horizon prediction error than the evaluated baselines on partially observed Gymnasium MuJoCo systems, while substantially reducing training time relative to the strongest prediction baseline. On high-dimensional MuJoCo Menagerie robots, it achieves competitive tracking accuracy with low per-step computation times, demonstrating potential for real-time deployment. These results support structured probabilistic dynamics as a data- and compute-efficient approach to world modeling and control under partial observation.
Real-world evolving networks are naturally modeled as temporal graphs (TGs), where capturing temporal dynamics is essential for predicting future graph properties that support downstream decision-making. Existing temporal graph methods have been developed primarily for single-task prediction, and little is known about their generalization across tasks or transfer to unseen networks. This leaves the challenge of multi-task graph property prediction in TGs largely open. We address this challenge by introducing Hydra, a novel architecture that integrates local connectivity features from temporal GNNs with a spectral learning module that captures global connectivity patterns. This design enables joint learning of local and global information under a multi-task objective. In multi-task classification, Hydra achieves an 8.9% relative gain in AUC over the strongest competitor. In multi-task regression, Hydra achieves competitive results in all three tasks, while obtaining the best results in two tasks with a 8.2% relative gain in MAE compared to the strongest baseline. Moreover, Hydra delivers these gains with a 22% reduction in training time compared to temporal transfer models. These results provide the first systematic evidence that multi-task transferable learning on temporal graphs is effective. By delivering consistent top-ranked performance, Hydra highlights multi-task training on temporal graphs as a promising direction toward adaptable foundation models for temporal graphs.
Text- and goal-conditioned human-object interaction (HOI) provides an intuitive interface for specifying high-level intents in physics-based character animation. Mapping such intents to physically plausible, contact-rich behaviors naturally calls for a generative planner and a physics-based controller, yet existing planner-controller systems either execute plans open-loop or close the loop while keeping the planner fixed, so execution feedback never reshapes what is planned. We present InfCoiL, a closed-loop framework for physics-based HOI from coordinated planner-controller learning. InfCoiL couples a morphology-aware flow-matching planner with a unified multi-morphology controller through a shared morphology interface in an autoregressive plan-and-execute loop, and alternately fine-tunes both components with reinforcement learning from the same physical rollouts. To optimize the flow-matching planner online, we introduce Semantic Flow Policy Optimization (SFPO), which stochasticizes selected denoising transitions with channel-wise semantic covariance over heterogeneous HOI features and assigns group-relative, sequence-level rewards from complete closed-loop executions. On held-out OMOMO sequences, InfCoiL outperforms open-loop hierarchical baselines in goal-reaching task success and physical tracking accuracy. Project page: https://infcoil.github.io/
Understanding how an athlete moves can improve both technique and safety, yet outside the laboratory technique is still assessed by eye. We present a pipeline that detects named golf swing faults from ordinary video, without markers or calibration. Instead of learning end to end, it tracks body pose and the clubhead in a body-centered frame, and classifies the biomechanical quantities a coach would measure, such as the height of the clubhead above the swing plane, with shallow boosted trees. We demonstrate the approach on two common faults, coming over-the-top (OTT) and early-extension (EE). On 160 swings and 143 swings from public sources it detects OTT and EE with a cross-validated ROC-AUC of 0.975 and 0.911 respectively. Because every feature is a biomechanical quantity a coach already uses, measured at a known phase of the downswing, each decision can be traced to the signals and the moments of the swing that drove it. The signals the model relies on most are the same ones a coach checks when diagnosing these faults. The pipeline processes phone video on a consumer laptop in seconds, so it could feasibly be developed into an app for golfers of all abilities and extended to other sports in which technique is judged against a biomechanical reference.
Amortized Bayesian experimental design (BED) enables real-time design strategies by shifting acquisition cost offline, but existing methods remain tied to the prior and task distribution they were trained on and cannot exploit additional information available at deployment. We introduce in-context amortized BED with unified knowledge encoding (IMBUE), a BED framework that admits external knowledge at deployment through a unified in-context interface. IMBUE accepts two forms of such knowledge without retraining, namely user-specified prior knowledge and observations from previous instances of the experiment. Both are encoded as additional input tokens and processed in-context with the accumulated observations by a shared Transformer, and a learned reliability filter scores each external token against the observations and excludes inconsistent ones. Across four benchmarks, IMBUE accelerates early-stage information acquisition when the external knowledge is reliable, and remains close to the no-external baseline when it is not.
Aerial vision-language navigation (VLN) asks a drone to find an object described in natural language. The description usually locates the object relative to named landmarks, such as streets and buildings. We observe that current navigators often bring the drone close to the target and then stop at the wrong place: they reach the target without finding it. This gap is shared by navigators of very different designs, from imitation learners to a fine-tuned 7B vision-language model. A simple test reveals the cause. When we replace the target object in the instruction with another one next to the same landmarks, no navigator goes to it, and most stop where they did before. Navigators learn to reach the landmarks, not to identify the described object. We therefore propose MAP-VP, a plug-and-play module for the last step. After any navigator stops, MAP-VP draws the landmarks named in the instruction onto the drone's current view, using the landmark outlines that the benchmark provides on a map. A fine-tuned 4B vision-language model then locates the target relative to these landmarks. The training data are generated automatically from the benchmark annotations. Neither drawing the landmarks nor fine-tuning helps much alone, but together they are highly effective. Without retraining any navigator, MAP-VP significantly improves all six navigators we test on CityNav. It roughly doubles the success rate of the two strongest trained navigators, raising HETT from 25.9% to 54.6% and FlightGPT from 24.2% to 48.9%, whereas CityNav's official detector-based last step adds less than two points to HETT.
Recent studies of embodied agents based on frontier multimodal large language models (MLLMs) develop harness engineering around the MLLM, such as harnesses with post-trained policies and recursive self-improvement loops. Yet whether these additions exceed the native capability of the MLLM remains unexplored. We introduce SimplEvo to explore how far an MLLM can improve under direct control when it retains its execution experience in self-written text memories. It requires no weight updates, post-trained policies, or privileged information. The MLLM writes a coordinate-free task memory once from a single demonstration. Under a strict train-test protocol, it then creates a bounded working memory on training episodes only and refines it after each episode with lessons on procedure, completion, and recovery. Both memories stay fixed at test time. On EBench, SimplEvo lifts the success rate (SR) of GPT-6-Astra under direct control from 46.7% to 81.7%, and tops the RoboDojo leaderboard with 69.63% SR and a score of 75.32. Compared with memories that also store past visual frames, SimplEvo reaches a similar SR with text memory alone while using fewer tokens. With its memories fixed, SimplEvo also generalizes to unseen layouts and randomized scene configurations. Moreover, the working memory transfers across models: written by GPT-6-Astra, it can be reused as is by weaker MLLMs and improve their performance without any adaptation of their own.
Lookup-table (LUT)-based neural networks (NNs) offer a compelling path to ultra-low-latency inference on FPGAs. However, existing LUT-based neural networks have largely been restricted to MLP-style topologies, leaving structured vision and audio tasks underexplored. We present ConvLUT, a family of fully unrolled LUT-mappable convolutional architectures targeting the MLPerf Tiny benchmarks. It is comprised of a standard convolutional model built around a hierarchical tree of LUTs and a depthwise separable model following the MobileNet design philosophy. These architectures are enabled by four key contributions: a unified LUT-based convolutional framework supporting standard, depthwise, and pointwise convolutions under explicit fan-in constraints; a LUT-tree that addresses the $k^2c$ receptive field dimensionality bottleneck; a channel-partitioned LUT pooling layer that replaces conventional spatial pooling; and a hardware-aware iterative pruning strategy for stable convergence at scale. Evaluated on CIFAR-10, Visual Wake Words, and Keyword Spotting, ConvLUT achieves competitive accuracy with fully unrolled inference latencies below $\sim 50$ ns. On CIFAR-10 it reaches at least a $8844\times$ latency reduction over non-logic-based prior implementations, and outperforms existing MLP-style LUT-based models on the same task by at least $20$ pp in test accuracy.
An LLM agent that reads untrusted content, such as an email or a web page, may follow instructions hidden in it. Noninterference is too strong for such agents, since a useful agent lets untrusted content influence what it does and may disclose the user's data by design, and declassification needs a release policy written for each task. We instead state security relative to a reference agent, which runs the same plan under the same runtime monitor with a model that never obeys untrusted content. An agent is relatively secure if every leak it has is a leak the reference already has. Our framework, PlanTERS, instantiates relative security, a notion from speculative execution, with a small agent language in which the attacker chooses how the model follows instructions, a composition with runtime monitors, and attacker models that differ in what they observe. A fence theorem shows that revoking injectable tools at the first untrusted read, and never granting them back, makes every plan relatively secure at the event level on finite runs, for every model behaviour in which each injection is a single call and the model obeys on a schedule fixed in advance; with a fairness condition, this extends to possibly infinite runs. We prove or refute relative security for agents with fences, budgets, approvals, loops and dual LLMs, in Isabelle/HOL.
Due to the fundamental problem of causal inference -- the impossibility of observing counterfactual outcomes -- there is always a gap between the training goal (outcome, propensity) and the inference target (CATE, ATE). Traditional causal effect estimation revolves around setting assumptions, selecting identification strategies, and choosing causal estimators, while more recently, Causal Foundation Models (CFMs), built on Prior-Data-Fitted Networks (PFNs), ease this process by pretraining on massive synthetic prior datasets. Despite their success, such amortization makes causal inference less transparent, obscuring which assumptions the model implicitly relies on for a given dataset. The high diversity of real-world data-generating processes (DGPs) and Structural Causal Models (SCMs) makes it hard to know, a priori, which causal inference strategy is compatible with a given dataset. In this work, we focus on this compatibility between an observational dataset and method-specific assumptions, and find that geometric characteristics of data embeddings -- specifically, the variogram -- provide a useful signal of whether a given causal estimation strategy is appropriate. We develop MoE-Learner, a mixture-of-causal-experts framework that embeds a dataset under multiple assumption-specific views and uses the variogram as a counterfactual-free router to assess each expert's validity. Rather than eliminating causal assumptions, this shifts their validation from a fixed, a priori choice to a criterion that can be checked directly from data geometry. MoE-Learner achieves the best average rank on benchmark datasets (IHDP, ACIC, Lalonde). Being orthogonal to (causal) foundation models, our method builds a bridge from data characteristics to causal estimation strategy rather than proposing another CFM.
Computed tomography (CT) restoration methods often use a single network for low-dose CT (LDCT), sparse-view CT (SVCT), and metal artifact reduction (MAR), despite the distinct physics of these degradations. We propose AgentCT, a vision-language agent that identifies the degradation and routes images to task-specific restoration models. A LoRA-adapted Qwen3-VL-4B controller predicts degradation type and severity, selects a specialist, and reassesses the restored image. The LDCT specialist uses an image-domain diffusion model, whereas the SVCT and MAR specialists combine projection-domain correction, data consistency, differentiable filtered back-projection, and image refinement. We evaluated the controller and specialists separately on patient-disjoint Mayo 2020 abdominal CT data across three severity levels for each degradation. The controller achieved 95.7% accuracy for degradation type and 93.9% for severity, with a 100% JSON parse rate. Using ground-truth degradation labels for routing, AgentCT achieved an overall PSNR/SSIM of 46.18 dB/0.984, compared with 44.93 dB/0.976 for AMIR and 44.29 dB/0.979 for TAMP, and attained the highest PSNR in all nine settings. For SVCT, its dual-domain specialist achieved 46.04 dB with 4.2 million parameters, compared with 44.43 dB for single-task AMIR and 43.57 dB for all-in-one AMIR with 23.5 million parameters. These results support degradation-specific CT restoration with controller-based model selection; evaluation under automatic routing and on independent clinical data remains necessary. Code is available at https://github.com/hao1635/AgentCT.
Marcolli, Chomsky, and Berwick (2025; henceforth MCB) formulate the Merge operation of Minimalism on a Hopf algebra of Connes-Kreimer type, and claim that language and physics obey the same algebraic structure. This paper separates their program into the working hypothesis H, that Merge on workspaces can be described as a Hopf algebra, and the claim P placed on top of it, that language and physics obey the same principle, and examines the validity of H. MCB employs two coproducts in formulating Merge, corresponding to two different translations of movement (Internal Merge): $\Delta_d$, which deletes the lower copy, and $\Delta_c$, which leaves a trace. MCB uses the former in syntax and the latter at the syntax-semantics interface, but $\Delta_d$ is not coassociative, and $\Delta_c$ is not connected under the gradings MCB gives and, moreover, not coassociative as long as the labels of traces are retained. Consequently, MCB's algebra is not a Hopf algebra, and the Birkhoff factorization is not a consequence of the Hopf-algebra theorem but of a weaker statement that requires neither coassociativity nor connectedness. There are four ways to recover a Hopf algebra, but none of them can distinguish different traces, and hence none is compatible with Heim-Kratzer-style formal semantics, which reads traces as indexed variables. Thus H does not hold at present, and under the current formulation, making Minimalism a Hopf algebra and preserving its connection with formal semantics are mutually exclusive. Finally, we indicate three directions in which algebraic Minimalism may proceed.
Robotic manipulation requires representations that capture both scene geometry and its evolution during interaction. Existing approaches have largely treated these two aspects separately. World action models capture such dynamics by coupling action prediction with future video generation, but their reliance on iterative diffusion or rectified-flow inference leads to high computational cost, while their representations lack explicit spatial structure. Explicitly modeling scene geometry typically requires auxiliary modules that increase both parameter count and computational cost. Conversely, geometry-centric policies primarily recover observed spatial structure and often lack explicit supervision for anticipating future geometric changes during interaction. To address these limitations, we introduce ForeAct, which repurposes a pretrained video world model as a single-pass representation backbone for low-latency control. ForeAct directly processes the clean latent of the current observation in one forward pass, eliminating future-frame synthesis and iterative denoising while retaining the backbone's pretrained spatiotemporal priors. Rather than introducing a separate geometry backbone, ForeAct uses learnable tokens supervised by next-frame relative depth to shape the shared video representation through bidirectional self-attention. This compact interface turns the spatial knowledge already encoded in the backbone into a future-aware representation for action generation, with little parameter overhead. The lightweight geometry readout is used only during training and removed at deployment, enabling low-latency, geometry-aware control without depth input or an external geometry backbone. Without additional large-scale pretraining, ForeAct achieves 96.6% average success on LIBERO, 63.5% on SimplerEnv-Bridge, and a 74.4% success rate across 90 real-world rollouts.
The performance of language agents depends jointly on model weights and the harness. Optimizing either alone can be constrained by the other component, limiting overall performance gains. Our analysis of naive alternating training reveals progressive incompatibility between the model and harness, which increasingly interferes with task execution and may also undermine model training, leading to performance degradation. To coordinate the evolution of models and harnesses, we introduce JoMoHa, to our knowledge the first framework to jointly train an executable harness generation policy and a task execution policy through reinforcement learning. JoMoHa outperforms the evaluated weight-only and harness-only baselines and achieves the highest success rates among the evaluated methods on ALFWorld, WebShop, and DBBench, averaging 70.3% and reaching 97.0% on ALFWorld. By learning both task execution and runtime improvement from its own generated trajectories, JoMoHa performs recursive self-improvement (RSI).