On-Policy Self-Distillation (OPSD) improves reasoning by training a student on its own reasoning, using the same model conditioned on reference solutions as a teacher. Applying OPSD to multilingual reasoning requires the teacher to provide useful feedback across languages, even when the student makes mistakes in intermediate steps. We observe a phenomenon we call Teacher Drift: as these mistakes accumulate, the teacher may follow the student's incorrect steps instead of correcting them. Our analysis shows that this shift tends to occur earlier in lower-resource languages. To account for these differences across languages, we propose Teachable On-Policy Self-Distillation (TOPSD), which determines where and whether teacher supervision should be applied to each reasoning trajectory. TOPSD uses teacher-student likelihood to estimate where drift begins and select a prefix for supervision. It then verifies whether the teacher can recover the correct answer from that prefix, applying distillation only to prefixes that pass verification. Experiments on multilingual mathematical reasoning across ten languages and three model scales show that TOPSD consistently outperforms conventional OPSD, particularly in low- and medium-resource languages.
Bilateral trade models the task of intermediating between two strategic agents, a seller and a buyer, willing to trade a good for which they hold private valuations. We study this problem from the perspective of a profit-maximizing broker in a regret minimization framework. At each time step, a new seller and buyer arrive and the broker proposes a trading price to each agent. The trade happens if both agents accept, and the profit of the broker is the difference between the price received from the buyer and the one paid to the seller. At the end of each interaction, the learner only observes whether the trade occurred, and its goal is to achieve sublinear regret relative to the best fixed pair of prices in hindsight. We start by investigating the adversarial version of the problem, where buyer and seller valuations are generated by an adversary. In this setting, we construct a polynomial-time learning algorithm with a $\tilde O(T^{\frac 23})$ regret rate, which is nearly tight even if the valuations are drawn i.i.d. from independent distributions, thus matching the rate for one-sided pricing. Our adversarial algorithm combines a localized estimation procedure with an information-based minimax construction. We then consider valuations drawn i.i.d. from independent distributions satisfying the monotone-hazard-rate (MHR) property and provide a learning algorithm with nearly tight $\tilde{O}(\sqrt{T})$ regret. The algorithm combines localized estimation with a stochastic convex bandit routine applied to the negative logarithm of expected profit. Finally, we complement this result by showing a separation between MHR and regular distributions, proving that the minimax regret rate in the latter setting is $\tilde Θ(T^{\frac 35})$. This last result contrasts with the $\tilde{O}(\sqrt{T})$ regret bound that can be obtained for one-sided pricing with regular distributions.
Attention-based time series forecasters commonly group consecutive observations into patches. Patch length is known to have a substantial impact on forecasting performance, yet its selection remains largely empirical. We develop a theoretical framework that explains why patch length matters for forecasting accuracy and provides a principled basis for its selection. For linear state-space models, we derive an exact factorization of the forecasting coefficient matrix that separates forecast dynamics from filtering dynamics. We show that dominance of a single forecast component at long horizons does not generally imply a rank-one representation. This explains why longer horizons alone need not reduce representational complexity. We combine the factorization with estimation bounds to develop a patch-selection criterion that balances representation and estimation errors under temporal dependence. Our analysis establishes conditions under which balanced patching is justified and clarifies how predictive structure, head count, training sample size, and long-run dependence enter the selection criterion.
Image classification models often encounter objects in diverse and unpredictable contexts at test time, which may differ substantially from those seen during training. This phenomenon, termed context shift, exposes a key fragility in neural networks: when models rely on non-causal contextual cues for prediction, shifts in context can lead to significant performance degradation. In this paper, we propose a novel training framework to improve the robustness of convolutional neural networks under context shift. Our approach leverages spatial feature attribution to guide models toward making predictions that rely less on contextual regions and more on object-relevant features. As a first step, recognizing a theoretical limitation of existing feature attribution methods we introduce a new variant, ContrastiveCAMs, which produces more faithful attribution maps of model predictions. Building on ContrastiveCAMs, we further propose Context-Regularized Cross-Entropy (CR-CE), a modification of the standard cross-entropy loss that regularizes the model's attention by suppressing the influence of contextual regions, thereby improving robustness to context shift. We evaluate the effectiveness of our approach on several medium-to-large scale datasets (Waterbirds, Spawrious, ImageNet/ImageNet-BG, Hard-ImageNet), and report consistent improvements in context-shift robustness.
World-action models adapt pretrained video generators, and vision-language-action models build on pretrained vision-language backbones; both tie policy architecture and inference cost to large transformers designed for generation rather than control. It remains unclear how much these generation-oriented choices affect control. In this work, we introduce Latent Action JEPA (La-JEPA), a compact world-action model whose policy transformer is trained from scratch in the feature space of a pretrained V-JEPA 2.1 encoder. Following Fast-WAM, La-JEPA learns to predict future features jointly with actions during training and generates only actions at inference. Among policies without embodied pretraining, La-JEPA achieves state-of-the-art performance in clean and randomized RoboTwin 2.0 scenes and on LIBERO-Plus with correct instructions, while using far fewer parameters than the strongest competing policies. It runs substantially faster than Fast-WAM at comparable performance and outperforms the closest JEPA-based methods, LeapBot-WA and JEPA-WAM, on LIBERO-Plus and RoboTwin 2.0 without their pretrained policy backbones. On four real-robot tasks, its average success is comparable to pi0.5 and above non-embodied-pretrained baselines. Through a design study of the choices behind La-JEPA, we show that JEPA features outperform video VAE latents, most clearly out of distribution, and that training-time future prediction choices significantly impact the policy even when future prediction is skipped at inference. We correct the LIBERO-Plus instructions, as the officially released evaluator corrupts them for most variants, and re-evaluate several strong open baselines. Without a generative pretrained policy backbone, La-JEPA is better and faster than far larger policies, suggesting that robot foundation models may benefit from architectures designed for control rather than inherited from generation.
Semantic communication (SemCom) using a relay node is considered, where the relay assists in conveying task-relevant information from a source to a destination, with meaning defined by classification labels. A joint end-to-end design of the source encoder, relay transformation, and destination decoders is developed to optimize both semantic inference accuracy and perceptual reconstruction quality for image data. Relay-assisted schemes are shown to outperform direct transmission, with gains that become more pronounced in low signal-to-noise ratio (SNR) regimes across different relaying strategies. Alongside these performance benefits, however, a key security vulnerability introduced by the relay is identified, and a Byzantine attack is formulated in which the relay injects structured, learned perturbations into its transmitted signal. The attack operates on intermediate communication representations subject to channel effects, with perturbations generated through a learned mapping applied at test time. It is shown to significantly degrade semantic inference accuracy at the destination. These results reveal a clear trade-off between SemCom gains and security risks in relay-assisted systems, motivating the need for secure relay designs that account for insider threats.
This study evaluates the added prognostic value of day-1 multimodal MRI for predicting three-month functional outcome after acute ischemic stroke (AIS). Sixty-six AIS patients with ADC, DWI, FLAIR, and clinical data were analyzed. Three-dimensional pretrained ResNet embeddings were extracted from each MRI modality, reduced using fold-isolated principal component analysis, and combined with structured clinical variables and ADC-derived lesion biomarkers. Linear support vector machines were evaluated using eight-fold stratified group cross-validation across MRI modality combinations. Clinical-only modeling achieved strong performance (AUC = $0.899 \pm 0.151$), while MRI-only models showed limited discrimination (AUC $\approx$ $0.67$-$0.73$). Among multimodal configurations, FLAIR-clinical fusion achieved the highest mean cross-validated performance (AUC = $0.948 \pm 0.061$), followed by ADC-clinical fusion (AUC = $0.941 \pm 0.094$). Full ADC-DWI-FLAIR fusion did not consistently outperform single-modality clinical fusion. Feature-importance analysis identified age, NIHSS subscores, ADC lesion burden, and learned FLAIR representations as key contributors, while qualitative FLAIR activation maps showed responses over lesion-like regions. These findings suggest that compact day-1 FLAIR and ADC representations provide complementary prognostic information beyond clinical variables and support interpretable, data-efficient outcome prediction in AIS.
Weather fronts are transition regions between air masses with contrasting thermodynamic properties and are associated with mid-latitude precipitation and severe weather. We investigate machine learning (ML) methods for detecting weather fronts and quantify the contribution of reanalysis data to performance. We compare two configurations of the same ML model for pixel-wise frontal classification over the North Atlantic and Europe: (1) using only geostationary infrared satellite observations, and (2) additionally using 28 ERA5 reanalysis fields. The satellite-only model achieves a neighbourhood critical success index at a 150 km tolerance (CSI@150) of $0.402$, which lies within the range of agreement between analyses from different meteorological services. The satellite$+$ERA5 model achieves $0.575$, an improvement of $0.173$. Our analysis shows that low-level winds are the least predictable from the satellite representation ($R^2 = 0.23$, compared with $0.88$ for temperature), while removing winds or moisture causes the largest decreases in detection performance.
Dense optical flow allows small drones to avoid obstacles. However, few dense optical flow networks are designed for microcontrollers (MCUs) with neural accelerators, whose on-chip SRAM limits the input resolution. We present PACE-NAS, a macro-to-micro neural architecture search framework that decouples memory from latency: a macro design fixes the backbone and its peak SRAM, and a micro design searches operators and depths to trade accuracy against compiler-estimated latency. We apply it to EdgeFlowNet on a Grove Vision AI V2 board (Ethos-U55), where profiling locates the peak SRAM at the full-resolution output head. Replacing transposed convolutions with bilinear upsampling and convolutions lowers the peak SRAM by 18%, and two lightweight context modules lower the endpoint error (EPE) without raising the peak. All searched subnets share this peak, so the search needs no memory constraint. We name the Pareto-optimal models MCUFlowNet; the fastest variant, MCUFlowNet-S, runs 1.76x faster than a retrained EdgeFlowNet at a 1.6% higher EPE and, in simulated forest flights at 2 m/s, reaches the goal without collision in 51.5% of trials, against 41.8%. PACE-NAS also applies to semantic segmentation, where a searched PIDNet-S variant runs 1.17x faster at similar accuracy. Code is available at https://gitlab.kuleuven.be/emedia-research-lab/esat-stadius/mcuflownet/mcuflownet.
Robot reinforcement learning increasingly runs environments outside the trainer: in another process, on another machine, in another language. Many faults at this boundary raise no error. We make the boundary a wire protocol and ask what notices a fault in it. On Pendulum, HalfCheetah, Hopper and a 7-joint arm (Pusher), we injected up to 23 faults into what crosses the wire and 2 into the trainer's log, from one value to every value. Unless every value was wrong, a learning-curve threshold caught at most one run in ten, and a comparison with ten fault-free seeds at most 21%. A clean evaluation of all 1,060 final policies agreed with the curves: of the 728 boundary-fault runs both curve checks passed, 723 scored no more than three standard deviations below fault-free seeds, though nearly all ended on other weights and, for reward faults, logged other returns. Checks of the run itself, not of its outcome, saw every fault. A step ledger kept on both sides and compared step by step caught and located every fault that crossed, on Pendulum, Hopper and Pusher on one machine and on Pendulum across two, with no fault-free run and, on Pendulum, a median 3% more training time. An episode record kept at the environment caught the faults in the trainer's log. We list such faults found in practice, in Gymnasium 1.3 and our own code and, reported by others, in five multi-agent libraries, and recommend a step ledger and episode counts kept at the environment.
This paper presents a novel task of describing human facial expressions of a facial image in natural language, which captures the nuances of facial actions and emotional states beyond traditional emotion categories or facial action units (AUs). To achieve the facial expression captioning model, we propose a three-stage training framework that trains a vision-to-language model using synthesized image-text pairs and the BLIP-2 pre-training techniques. To overcome the challenge of missing training image-text pairs for facial expression captioning, we propose a strategy that involves synthesizing and combining captions using GPT-3.5 and existing annotations on either emotion categories or AUs. Experiments demonstrate the effectiveness of our method in generating captions that describe details of facial actions and emotions, as well as the inferential relationship between them, even when those emotions are not present in the training data. It is also demonstrated that the vision-to-language task enhances the performance of the intermediate visual features on both AU detection and emotion classification tasks.
Production deployments often use different models to balance cost and quality through cascading, routing, and switching models during long conversations. Context accumulates across turns. Each switch requires the target to repeat prefill over this context. Existing methods transfer full-attention (FA) K/V caches across models. Hybrid linear-attention models also retain recurrent states (REC) and convolution histories (Conv). The interaction between Conv and REC complicates cross-model state transfer. Conv perturbations can change keys and values in recurrent layers and alter REC updates. Conv histories are overwritten, but changes in REC can still affect later outputs. More accurate REC reconstruction need not improve continuation. Building on this, we propose Hybrid State Transfer (HST) for all three state types. It uses structured REC maps and preserves history order during Conv alignment. We calibrate the adapters using the frozen target's outputs during continuation. FA cache map updates are merged offline. Across five benchmarks, Qwen3.5 4B$\rightarrow$27B and 9B$\rightarrow$27B achieve mean task retentions of 89.88% and 94.11%, respectively. With source caches already available, separate timing on synthetic 16K inputs yields handoff body speedups of $21.66\times$ and $21.91\times$.
Current 4D occupancy world models use heavy cross-attention or deep recurrence to disentangle ego-motion from scene dynamics. This requires significant computational capacity. However, this effort is largely redundant since the ego-vehicle's intended future motion is provided by the downstream motion planner. We present FreeOcc, a lightweight continuous-time world model that introduces explicit $SE(2)$ geometric decoupling into a conditional flow-matching framework. By mathematically factoring out self-motion, we create a purely dynamic feature space. This space supports attention-free temporal fusion and allows physics-constrained ODEs to operate strictly on residual agent kinematics. To maximize deployment efficiency, FreeOcc introduces an uncertainty-aware blending mechanism. This routes well-observed voxels to a single-pass decoder, reserving the ODE solver exclusively for ambiguous, high-uncertainty regions. Requiring only 28.3M parameters-roughly 40% of comparable models-our optimal ensemble mode (FreeOcc-E) runs at 27 FPS. On nuScenes, FreeOcc achieves state-of-the-art forecasting performance with a 0.47 m Near Field Chamfer Distance (NFCD) and 32.1% mIoU. Furthermore, the model transfers zero-shot to SemanticKITTI. It also proves highly robust, retaining 92.8% of its ideal mIoU under significant sensor packet loss and timestamp jitter.
Rare genetic disorders are difficult to diagnose due to their diversity and overlapping clinical features. Facial morphology can provide diagnostic clues, yet most diagnostic models offer limited interpretability and are not designed to support broad screening for undiagnosed disease. We present ProtoGestalt, an interpretable prototype-based framework for facial phenotyping, rare-disease screening, and syndrome identification. The method extends prototype learning by constraining learned visual features to select facial regions, allowing predictions to be explained through localized and additive evidence. We assess ProtoGestalt's ability to distinguish syndromic from unaffected faces on multiple control datasets and class prevalence values. We further evaluate it's diagnostic performance on syndromes observed during training. The learned prototype-based representations provide significant separation between syndromic and case-matched control images and retain useful performance as the proportion of controls increases. The framework also demonstrates the ability to differentiate among syndromes. Although reasonably accurate specialized models for facial classification of rare disorders already exist, ProtoGestalt offers an interpretable alternative that links predictions to anatomically meaningful facial regions, potentially supporting clinical decision-making and the discovery of phenotype-gene associations.
State-space models (SSMs) recover low-dimensional latent dynamics from high-dimensional observations, but latents are generally not identifiable from observational data alone. Interventional state-space models (iSSMs) use perturbations to break observational symmetries and identify parameters up to permutation, scaling, and shift. However, prior work restricts identifiability to linear latent dynamics and piecewise-linear emissions, excluding many behaviors characteristic of neural circuits. We extend iSSMs to a substantially broader model class: we prove that (i) affine identifiability holds under general injective emission functions and Gaussian latent transitions, and (ii) with interventions, identifiability up to permutation, scaling, and shift is preserved under general nonlinear latent dynamics. We validate our theory on simulated FitzHugh-Nagumo dynamics, and then demonstrate two applications: adjudicating between competing models of decision-making; and predicting responses to unseen micro-stimulation in macaque and human neural recordings, with proof-of-concept generalization to a held-out stimulation site in the latter. Finally, simulations yield practical guidance for designing stimulation protocols that improve identifiability.
World-action models (WAMs) jointly model future traffic states and driving actions for end-to-end autonomous driving in intelligent transportation systems. However, their action learning remains dominated by expert-trajectory imitation, leaving unsafe alternatives close to expert trajectories insufficiently constrained. To improve WAM robustness in safety-critical scenarios, we propose NegWAM, a safety post-training framework for unified WAMs that learns explicit safety boundaries from near-expert failure trajectories. NegWAM constructs negative trajectories that are geometrically close to the expert trajectory yet carry safety risks, uses them as repulsive anchors to steer predictions away from dangerous behavior, and encodes them as scene-conditioned rider tokens to learn scene--trajectory compatibility. During training, NegWAM jointly optimizes the flow-matching and near-expert negative objectives, while the auxiliary rider branch is discarded at inference, incurring no additional overhead. On NAVSIM v2, NegWAM achieves 40.4 EPDMS on navhard, a 4.1% relative improvement over the strongest compared scorer-free WAM baseline, and achieves 93.3 EPDMS on navtest. These results demonstrate that NegWAM achieves strong performance in safety-critical scenarios while preserving routine-driving performance. Zero-shot closed-loop evaluation on HUGSIM further demonstrates cross-benchmark generalization. The code is available at https://github.com/Zizizi-hao/NegWAM.
Video compression at ultra-low bitrates remains challenging due to the loss of perceptually important details under limited bit budgets. Recent generative video compression methods leverage pretrained video priors to improve perceptual quality and temporal coherence, but many follow a compress-then-generate paradigm, where generative priors are primarily exploited during reconstruction. Moreover, representations learned for generation and reconstruction are not inherently optimized for entropy coding. In this work, we propose \method, which adapts pretrained video representations into compression-oriented representations while preserving their reconstruction capability. Specifically, we adapt the pretrained causal video VAE of MiniMax-H3 into a discrete video tokenizer and develop a small autoregressive entropy model that captures local and global token dependencies, with grouped prediction for parallel decoding. To further improve coding efficiency, we jointly optimize the tokenizer and entropy model using a differentiable rate surrogate, allowing rate gradients to adapt token assignments toward lower coding costs while preserving reconstruction quality. Experiments on UVG, MCL-JCV, and HEVC Class B demonstrate superior perceptual rate--distortion performance, achieving BD-LPIPS savings of 82.2\%--87.5\% and BD-DISTS savings of 78.3\%--85.0\% relative to GLC-Video.
Tree-species composition shapes forest productivity and how forests respond to climate disturbances such as drought, windthrow and insect outbreaks. Yet most national forest maps assign each pixel a single dominant species and cannot show the share of accompanying species. We mapped the basal-area shares of ten main tree species and a residual Other class across Switzerland at 30 m resolution. A deep-learning model combining 0.1 m aerial RGB imagery, airborne laser scanning (ALS), annual AlphaEarth satellite embeddings, and climate variables was trained on basal-area observations from 3,992 Swiss National Forest Inventory (NFI) sites. In spatial cross-validation, the model reached a macro-$R^2$ of 0.485 across the 11 classes. On an independent forest-reserve dataset, the map reached a macro-$R^2$ of 0.552 for individual 30 m pixels and 0.733 for plots, and identified the dominant class in 73.2% of pixels and 81.1% of plots. The map shows maple and ash as widespread minor components that dominant-class maps omit. It also reproduces the elevational sequence from broadleaved lowland stands through spruce to larch and Arolla pine near the treeline. By resolving the species mixture within each 30 m pixel, the map complements field inventory statistics with spatially continuous information that can support forest monitoring, the management of mixed stands, and climate-adaptation planning across Switzerland.
In large-scale payment networks, entities (card, merchant, etc.) and their interactions play a key role in transaction anomaly detection. However, transaction patterns between different entities usually change more frequently than the production model refresh schedule. Meanwhile, entity embedding tables suffer from high-cardinality and sparse interaction issues. To resolve the above challenges, we present PersonaAgent, an agentic transaction anomaly detection framework which represents the latest entity behavior signals as personas and saves them in a refreshable memory. Offline, a weekly scheduled batch job is used to extract the latest behavior signals and update them in the memory. At inference time, a lightweight adapter retrieves the corresponding entity behavior signals of each transaction from memory and uses them to augment the base anomaly classifier's scoring. Through comprehensive benchmarking with different anomaly detection models at Visa, PersonaAgent is able to score transactions efficiently (a 1,000-transaction CPU batch in 1-2 ms), lifting the PR-AUC by 0.65-6.58% and the production business metric by 8.93-14.24% relatively in deployment.
Tensor network kernel machines (TNKMs) combine nonlinear features with low-rank tensor representations to model complex relationships efficiently. However, selecting an appropriate tensor rank remains challenging, as conventional rank-search methods require training multiple candidate models, increasing computational costs. In this work, we focus on a canonical polyadic (CP) tensor network parameterization and propose GB-CPKM, a gradient boosting framework that builds a higher-capacity predictor by sequentially adding fixed-rank learners. This approach provides a new perspective on rank selection by incrementally increasing model capacity, with ensemble size selected based on validation performance rather than through explicit rank search. Experiments on real-world regression benchmarks show that GB-CPKM trains up to 6x faster and uses up to 2x less peak memory than a conventional rank-selection method based on cross-validation, while achieving comparable predictive performance. These results highlight the potential of gradient boosting to improve the computational efficiency and scalability of tensor-network learning.