We propose a Koopman model-based reinforcement learning (K-MBRL) algorithm for controlling nonlinear robotic systems with unknown dynamics. The K-MBRL learns a control policy by minimizing an approximate policy objective that combines the analytical immediate cost with an estimate of the future cost evaluated at the one-step predicted state. To this end, we develop a stochastic Koopman operator-based learning method to approximate the unknown nonlinear dynamics in a linear representation, while a critic function estimates the remaining future cost. This formulation offers two key advantages. First, the linear Koopman dynamics simplify the model-dependent policy gradient computation to a constant matrix product, enabling computationally efficient policy improvement. Second, one-step model propagation mitigates the accumulation of prediction errors associated with long-horizon rollouts while retaining long-term cost information through the critic. The learned policy enables real-time control through feedforward evaluation without online optimization. We establish policy learning convergence guarantees under standard stochastic approximation assumptions. Simulations on HalfCheetah and quadruped four-legged locomotion tasks demonstrate strong sample efficiency and improved performance compared with classical model-free and model-based reinforcement learning baselines. A pendulum swing-up task is used to assess policy gradient approximation accuracy and conduct ablation studies on the learning mechanism and key hyperparameters. The proposed K-MBRL is further validated on a physical quadruped robot performing a challenging sequence of maneuvers: rising to an upright posture supported by two rounded rear feet, maintaining balance during bipedal walking, and safely returning to the ground, as shown in Fig.1.
https://arxiv.org/abs/2611.14186
World Action Models (WAMs) have recently emerged as a promising paradigm for robotic manipulation, endowing policies with strong capabilities for modeling temporal and physical dynamics. Existing WAMs, encompassing both video-first action prediction and joint video-action modeling, directly transfer global physical representations from world models to action generation. However, global scene representations may underemphasize local robot-object interactions, creating a mismatch between representations optimized for world evolution and those required for manipulation. To bridge this gap, we propose LaST0.5, a novel WAM that introduces an interaction-centric physical representation between world modeling and action generation. Specifically, LaST0.5 learns a compact interaction-centric latent space under the supervision of visual interaction cues between the robot and the environment. This reasoning space is instantiated as a latent expert that bridges the world expert and the action expert within a Mixture-of-Transformers architecture. Furthermore, LaST0.5 performs adaptive world-to-action transfer by jointly predicting future representations and progress values, using progress consistency to adjust action execution and trigger future resampling. Finally, we pretrain LaST0.5 on a large-scale robotic data corpus comprising over 14k hours of data, establishing interaction-centric physical reasoning as a scalable and effective paradigm. LaST0.5 achieves 99.5% on LIBERO, surpasses the previous state of the art on ten evaluated RLBench tasks, and demonstrates robust long-horizon execution and precise dexterous manipulation in real-world experiments.
https://arxiv.org/abs/2611.14048
Adversarial training is often plagued by "pseudo-attacks", perturbations that significantly increase the loss value but fail to truly alter the model's confidence structure. To address this, we propose two complementary approaches to directly attack and defend against confidence structure deformation. First, building upon the traditional loss maximization perspective, we introduce a hybrid objective that integrates cross-entropy and KL divergence, encouraging adversarial examples to simultaneously disrupt prediction accuracy and distribution consistency. Second, we geometrically reconstruct adversarial training by generating perturbations that maximize the distribution distance on the confidence manifold. Theoretical analysis shows that the optimal perturbation direction aligns with the principal eigenvectors of the Fisher information matrix. We also propose an efficient algorithm to avoid high-dimensional matrix computations. Experiments on multiple datasets and model architectures demonstrate that our method outperforms existing methods in adversarial robustness and robust transferability, particularly excelling in structural perturbations within high-confidence regions.
https://arxiv.org/abs/2611.14046
Text-guided segmentation models receive an image and a short caption about the lesion, and on public benchmarks the caption improves accuracy considerably. On colonoscopy and breast ultrasound, however, the size, count and location words of each caption were computed from the mask that later scores the prediction, and the chest X-ray captions were written by benchmark annotators. We tested what two published models take from them. On colonoscopy and chest X-ray, it is mostly the lesion's extent. Keeping only the size word or the lung-zone phrase recovers 84 to 98 percent of the improvement over an empty caption, and removing it loses 52 to 77 percent. The rest matters little, and a polyp called a rock is segmented just as well. We call this extent leakage. On breast ultrasound, nearly half the captions also name the diagnosis, which carries part of the benefit. Retraining on three datasets shows that models pick up this habit from their training captions, though only partly on breast ultrasound, and that it weakens their segmentation from the image alone. The leak can be measured without training a model, since caption words alone predict lesion area (coefficient of determination 0.54 to 0.77 on six of seven test sets). Without the test mask, the best method we tested, which estimates extent from the image, reaches 74.59 Dice on colonoscopy against the published 77.23. We propose an evaluation protocol that keeps what a caption says about the image apart from what it copies from the label.
https://arxiv.org/abs/2611.13970
} Predicting who will speak next, and when, is a prerequisite for conversational agents and social robots that take turns without waiting for silence. Multiparty next-speaker prediction (NSP) has been studied on corpora with per-speaker sensors, each with its own event definitions. No public benchmark covers the single camera and microphone a robot has, and no model has been compared against how well people perform the same prediction. We turn AVCC, a 31-hour audio-visual corpus of two- and three-speaker online conversation recorded with a single static camera and microphone, into a public NSP benchmark. We add a speaker-disjoint seven-video test partition with 865 hold and 865 shift events at mutual-silence gaps, word-level forced-aligned transcripts for the whole corpus, and a human reference benchmark collected at the same events. We score an audio-only Role-Relative VAP model (RoleVAP) and the multimodal MuVAP model. Labels, transcripts, per-response human data, and the scripts that reproduce every number are released.
https://arxiv.org/abs/2611.13957
Latent diffusion models decode approximately correct latents, making decoder sensitivity to perturbations relevant to image quality. After whitening, forward diffusion corrupts every direction at the same signal-to-noise ratio (SNR), despite unequal decoder sensitivities. We study an invertible linear layer between the encoder and diffusion process that shifts each direction's log-SNR. Its linear minimum mean-square-error (LMMSE) recovery risk has an exact spectral expression for any whitened latent distribution. Under an energy budget and a floor ensuring invertibility, minimizing decoder-sensitivity-weighted LMMSE risk integrated over the diffusion schedule yields a unique optimal allocation. The layer aligns with the decoder's expected pullback metric and equalizes marginal risk reduction across directions, computed by a deterministic spectral solver. At large budgets, the allocation follows a power law in sensitivity whose exponent is determined by the low-SNR tail of the loss weighting: one half for noise prediction and two thirds for data prediction under the cosine schedule, with a crossover on truncated schedules. We instantiate this principle as RB-LDC using Jacobian-based sensitivity estimates. On ImageNet at 256 x 256 resolution with SD-VAE and VA-VAE, trained RB-LDC improves reconstruction from noisy latents over budget-matched isotropic coding in all four metrics and five coding configurations, averaged over 15 SNRs and three pipeline seeds. It reduces average perturbed FID by 17.1 for VA-VAE block-local coding while preserving clean reconstruction. We also report intermediate SNRs where perturbed FID worsens. With a frozen diffusion model, RB-LDC yields lower FID point estimates in all 12 tested combinations of sampling progress and channel SNR under corruption during sampling.
https://arxiv.org/abs/2611.13955
Molecular property prediction is fundamental to computational chemistry and drug formulation research, yet existing benchmarks remain fragmented across datasets, properties, and evaluation protocols, limiting systematic evaluation of model generalization. We introduce Chem World, a large-scale benchmark integrating 17 public datasets with over 800,000 molecular and chemical-system instances across 10 property tracks, covering single-molecule and multi-component systems, including drug solubility. Chem World provides a unified testbed for evaluating molecular representations and compositional modeling across diverse chemical properties. Building on the benchmark, we propose PACM (Physics-Aware Compositional Modeling), which combines molecular representations with composition-aware modeling and domain-specific physical constraints. Extensive experiments show that molecular language models provide strong and transferable representations, while PACM consistently improves prediction across diverse chemical systems. Chem World and our PACM-powered drug-property prediction Agent support generalizable chemical modeling and offer a foundation for AI-assisted biomedical research and drug formulation development.
https://arxiv.org/abs/2611.13945
In open-world recognition, machine learning models not only make predictions on unknown structures ("generalized open-set"), but also continually learn and improve from those predictions over time. Interactive segmentation, despite its access to user feedback, has so far remained confined to the open-set setting. In this work we introduce MEOW, the first framework that turns non-interactive segmentation backbones into interactive open-world learners without updating learnable model parameters. MEOW combines parameter-free transductive reasoning with memory consolidation to accumulate knowledge and thereby improve future performance. Focusing on medical data, we show that MEOW consistently improves its own predictions on classes never seen during training, and transfers zero-shot to unfamiliar domains, including micro-CT data and other imaging modalities.
https://arxiv.org/abs/2611.13935
Acoustic side-channel attacks (ASCA) recover typed text from the sound of keystrokes, using only a commodity microphone and no access to the victim's device. Prior work establishes that such attacks succeed under idealized laboratory conditions - a single keyboard, a fixed microphone, a quiet room, and training and testing on the same device but leaves open the question that matters for security: whether keystroke inference survives the open world of unseen keyboards, diverse users, noisy environments, and compressed transmission channels. The answer developed here is that it does, provided the acoustic model is made explicitly domain-invariant and its noisy predictions are corrected by a language model.
https://arxiv.org/abs/2611.13922
Continual pre-training (CPT) adapts large language models (LLMs) to specialized domains, yet its standard next-token prediction (NTP) objective treats every corpus token as a fixed target, regardless of its utility to the current model. In a controlled study on the same 1B-token mathematical corpus, we observe sharply different outcomes across base models: NTP improves Llama-3.2-3B's performance on mathematical benchmarks but substantially degrades Qwen3-1.7B-Base's performance, indicating that, under a fixed training recipe, the same training sequences can have model-dependent effects. To address this problem, we introduce On-Policy Mid-Training Distillation (OPMD), which reformulates mid-training from offline text imitation into active, query-driven knowledge acquisition. At selected anchor positions, OPMD samples trajectories from the evolving student policy and applies dense token-level reverse-KL rewards from a frozen teacher, training the model on states it actually visits rather than only on ground-truth prefixes. This on-policy objective reduces the training-generation mismatch of teacher-forced NTP. On MATH-500 and Minerva, OPMD improves the Qwen student over its base performance and outperforms the evaluated NTP baseline, which degrades on both benchmarks. OPMD integrates into CPT pipelines without added inference-time overhead and offers a model-aware alternative to passive domain adaptation.
https://arxiv.org/abs/2611.13855
Human gameplay supplies paired observations, actions, and visual consequences that can supplement robot-policy training. We study event-and-foresight pretraining for InternVLA-A1.5, combining recorded keyboard/mouse event prediction with history-conditioned visual supervision through a frozen video model. Coverage-preserving anchor selection constructs one-second targets, and an attention mask prevents teacher-forced future actions from entering foresight queries. The selected corpus contains 394 recordings across 20 games, with 223.36 hours of recorded timestamp span and 2,072,084 training anchors. After 12,400 game-pretraining updates, we compare the resulting initialization with the public base under the same suite-specific robot fine-tuning protocol. Across four LIBERO suites, aggregate success increases from 1,718/2,000 (85.90%) to 1,811/2,000 (90.55%), a gain of 4.65 percentage points. Suite gains are 9.2, 6.8, 0.2, and 2.4 points on Spatial, Object, Goal, and Long. The 40-task breakdown shows 24 improvements, 10 declines, and six ties. These results support the practical value of the complete game-pretraining stage in this evaluated setting, while revealing heterogeneous task outcomes. The comparison includes additional optimization and does not isolate the effects of game data, the two objectives, or training variability.
https://arxiv.org/abs/2611.13819
Learning representations from multimodal observations is difficult in reinforcement learning with limited interaction data. Multimodal Joint Predictive Representation combines multimodal latent prediction with contrastive learning across modalities, but its recursive rollout depends on earlier predictions and its pairwise alignment focuses on correspondence between modalities. We propose Improved Predictive Representation, which combines prediction from action prefixes with alignment at the observed anchor based on Contrastive Multimodal learning. At each horizon, it predicts future representations directly from the observed anchor and the corresponding sequence of actions. We evaluate two crop configurations: IPR-S with a minimum retained crop area of 10%, and IPR-H with 50%. Across five DeepMind Control Suite tasks, IPR-H achieves an average return across tasks that is 10.0% higher than the baseline with visual and depth images and 3.4% higher with visual images and position vectors. The relative performance of the two configurations varies across tasks. In Webots, both IPR configurations achieve higher mean final returns than the baseline without LiDAR, but not when LiDAR is available.
https://arxiv.org/abs/2611.13782
Navigation in dynamic environments requires mobile robots to make progress toward their goals while respecting motion limits and avoiding surrounding obstacles. Learned subgoal policies can guide model predictive control beyond its local prediction horizon, but a geometrically suitable subgoal does not necessarily lead to a successfully solved control problem. This paper proposes OASIS, a hierarchical navigation framework that incorporates recent optimization outcomes into reinforcement learning-based subgoal generation. A high-level policy trained using Soft Actor--Critic generates local subgoals, while a nonlinear model predictive controller tracks them subject to motion limits, geometric collision-avoidance constraints, and embedded high-order control barrier function constraints. The outcomes of controller optimizations over each subgoal interval are summarized and provided to the policy at its next decision, without requiring additional optimization calls. The framework is evaluated in five navigation scenarios with static and moving obstacles, including dense interactions and sensing uncertainty. Comparisons with hierarchical and direct-goal control variants show that OASIS achieves the highest aggregate navigation success. Matched comparisons across three training seeds further show a consistent success advantage over the otherwise identical hierarchy without optimization-status input. The selected-policy comparison also shows fewer primary optimization failures with similar mean controller time.
https://arxiv.org/abs/2611.11996
In multimodal learning, modality imbalance often lets one modality dominate optimization and keeps the model from fully exploiting cross-modal interactions. Balance-oriented methods such as the Multimodal Competition Regularizer (MCR) perform well by alleviating this imbalance during training. These dynamic balancing methods, however, mainly adjust modality contributions and do not consider how imbalance affects the formation of synergistic information during fusion. Synergy is a major source of multimodal advantage, because it is predictive gain that cannot be recovered from unimodal evidence alone. We propose the Multimodal Synergistic Competition Regularizer (MSCR), a fusion method that jointly promotes balanced modality utilization and the discovery of synergistic information. MSCR has two complementary objectives. An anti-dominance objective suppresses excessive unimodal influence and leaves room for synergistic information, and a synergy objective encourages the fused predictor to outperform a non-synergistic unimodal reference. An adaptive gate further concentrates the regularization on samples that need stronger collaborative modeling. Experiments on five benchmarks show that MSCR consistently improves performance over MCR and strengthens synergistic capability. It raises conditional recovery, the accuracy of the fused prediction on samples that the unimodal predictors misclassify, as well as the share of the model's output that InterSHAP attributes to cross-modal interactions. Its gain in conditional recovery persists when the fusion module is Transformer-based. Code is available at this https URL.
https://arxiv.org/abs/2610.24395
A GUI command can be correct at the interface level yet still produce an incorrect assembly. An agent must identify the relevant components, respect their structural relationships, and determine which parts should move during recovery. We introduce MAGAR-Bench, a benchmark of 4,000 multimodal samples from 26 assembly families that separates Assembly Decision, Atomic Action, and Minimal Rework. For rework, the objective counts distinct moved parts within the validated recovery candidate set for each state, rather than the number of actions. We also introduce MAGAR, a selective correction framework built on a frozen policy obtained by supervised fine-tuning. Its key design matches correction granularity to output dependencies. Structured candidates revise assembly fields and rework scope, while complete action candidates keep native GUI primitives coupled with their parameters. A learned selector retains the original prediction unless a candidate passes a task-specific intervention threshold. On recorded states from held-out assembly families, MAGAR improves Overall, the task-balanced aggregate, from 37.50 to 41.39, with gains on all three tasks. Across 209 interventions, 70 correct an error and 22 introduce one. These results support task-specific correction while highlighting the remaining tradeoff between fixing errors and preserving reliable predictions.
https://arxiv.org/abs/2611.02472
Accurate future channel state information (CSI) is essential for underwater acoustic (UWA) communication because delays associated with signal propagation, feedback, scheduling, and transceiver reconfiguration can render currently available CSI outdated. However, most existing predictors estimate only the next channel state and obtain multi-step forecasts through recursive rollout, which causes prediction errors to accumulate as the multipath structure evolves. The scarcity of sea-trial measurements further limits reliable model training. To address these challenges, we propose LLM4UACP, a dual-domain, structure-preserving framework that adapts a pretrained Transformer for few-shot multi-step UWA channel prediction. LLM4UACP extracts complementary features from the channel impulse response (CIR) and channel frequency response (CFR), aligns them through Fourier-consistent fusion, and encodes the fused representation as delay-time tokens. A pretrained Transformer with a frozen backbone is adapted through Butterfly Orthogonal Fine-tuning to jointly predict multiple future CIRs in a single forward pass, thereby avoiding the error accumulation associated with recursive prediction. Experiments on the measured SO-06 and SO-07 channels show that LLM4UACP achieves normalized mean squared error (NMSE) values of 0.0411 and 0.0399, corresponding to reductions of 88.53% and 85.15%, respectively, relative to the most recent baseline. With only 0.11M trainable parameters, LLM4UACP maintains robust performance across varying prediction horizons and observation-window lengths, limited-data settings, and cross-dataset transfer scenarios.
https://arxiv.org/abs/2611.14592
Existing sequential user modeling methods mainly focus on improving the backbone encoder or learning task-specific user representations for a single prediction objective. We instead study how to export the multi-layer user states of a frozen long-history backbone as compact, timestamp-addressable representations for downstream training backfill and serving. In this setting, the user state at a given timestamp summarizes the user's actions before that time, so downstream systems need only the state aligned to each target timestamp rather than the full hidden-state timeline. To this end, we present \texttt{SeqFold}, a compressed-state interface that combines learned multi-channel fusion, channel-specific temporal decay, and Lookup-Free Quantization (LFQ) to produce compact discrete codes from frozen backbone states. Task-specific decoders then map these shared codes into flexible embedding spaces for downstream models. To make timestamp-addressable export practical at scale, we further develop a fused streaming implementation that updates the fusion recurrence after every action while materializing states only at target timestamps. On large-scale proprietary datasets, SeqFold reduces per-state representation storage by $3{,}400\times$ while preserving most of the downstream ranking performance of the uncompressed model, showing that SeqFold is an effective interface for reusing long-history sequence backbones in downstream models.
https://arxiv.org/abs/2611.13704
In this paper, we investigate how human egocentric videos can benefit robot policy learning without action supervision. Building on a pretrained VLM, we propose Shared Prediction on Action and Representation (SPAR), an MoT architecture in which future visual state prediction can provide independent and complementary learning signals alongside action prediction. We find that representations learned from egocentric videos improve downstream performance on out-of-distribution robot tasks, with further gains as video data scales. We also study the effects of objective choices and data curation, providing insights into how to leverage action-free videos for scalable robot learning.
https://arxiv.org/abs/2611.13685
On-policy distillation (OPD) improves language models by distilling teacher predictions on student-generated trajectories, while on-policy self-distillation (OPSD) removes the need for an external teacher by using a reference-conditioned copy of the model as the teacher. Despite its great early improvements, we find that OPSD does not remain consistently effective under continued training and can eventually suffer substantial performance degradation. We study this training behavior and link the degradation to a growing mismatch between the evolving reference-free student and the frozen privileged teacher. In the early stage of OPSD, the privileged teacher remains closely aligned with the student and provides useful guidance for improving the student's generation trajectories. However, as training progresses, the growing distributional mismatch between the privileged teacher and reference-free student can cause the teacher to provide misleading supervision, ultimately degrading student performance. This phenomenon motivates us to first leverage privileged reference supervision to acquire useful knowledge and then transfer it to a fresh student. We introduce Bootstrapped On-Policy Distillation (BOPD), a two-stage framework that first uses OPSD to internalize task-relevant information from reference-conditioned supervision into an intermediate model, and then distills this reference-free intermediate model into a fresh copy of the initial model. Experiments across models show that BOPD improves mathematical reasoning and code generation beyond both the initial models and the constructed OPSD teachers. These results suggest that privileged reference supervision can be more effectively exploited by first transferring its task-relevant signal into a task-adapted model and then distilling from that model without reference conditioning.
https://arxiv.org/abs/2611.13680
Dexterous manipulation requires rich spatial foresight over both the surrounding scene and the robot hand. Existing world-action models often acquire such predictive knowledge through explicit future prediction. We introduce DexForesight, a framework that learns spatial foresight implicitly by injecting dynamics-aware world representations into the latent space of a dexterous policy. During training, DexForesight learns an action-conditioned spatial world model from offline scene depth, kinematic hand geometry, and demonstration trajectories. The world model predicts 3D hand-point motion and visual feature changes, capturing both articulated hand dynamics and scene changes under demonstrated actions, including the motion of fingers occluded from the cameras. We then freeze the world model and align policy queries with its predictive world queries, transferring the learned spatial dynamics into policy latents and using them to condition action generation. At deployment, only the policy is retained, operating on RGB observations, language, and robot state without depth input or explicit future prediction. On 11 DexJoCo simulation tasks, DexForesight improves the overall success rate of the base policy from 49.4% to 60.5%. On three real-robot dexterous manipulation tasks, DexForesight achieves 58.3% success, compared with 50.0% for the strongest baseline. Project page this https URL
https://arxiv.org/abs/2611.11166