Muhammad Rajabinasab, Michael E. Houle, Oussama Chelly, Arthur Zimek
Many novel unsupervised feature selection methods are proposed each year, yet their empirical evaluation is limited to supervised and unsupervised evaluation metrics computed on selected datasets, along with comparisons to existing methods. However, in the absence of an established evaluation baseline, it is difficult to determine the value added to the existing literature by each of these methods, and how effective their underlying approaches are. In this evaluation study, we are using random feature selection as a baseline for evaluating unsupervised feature selection methods. While a random selection of features might seem to be a rather weak baseline, we show in this evaluation study that many of the state-of-the-art methods in unsupervised feature selection are outperformed by random feature selection in both performance and efficiency. Accordingly, we emphasize the strict requirement of considering random feature selection as a baseline in the development process of novel unsupervised feature selection methods, and we encourage the community to develop better evaluation measures for feature selection. Both would help to ensure consistent improvement and realistic assessment.
We introduce Mode Dynamics, a framework that connects coordinated parameter motion during neural network learning to changes in model outputs. Starting from a recorded parameter trajectory, we decompose its updates into collective modes and map each mode to its state-dependent output effect. The framework organizes this link through three mechanisms. Decomposition identifies collective modes in parameter motion and their time-varying amplitudes. Reconstruction tests whether the effects induced by those modes recover local and accumulated output change. Concentration quantifies how behavioral contribution is distributed across modes. Controlled scalar regression and next-token experiments directly test these mechanisms. They show that small sets of leading modes reconstruct behavior, induced effects depend on the training path, and modal complexity varies with task and data. We then evaluate the framework on held-out image and language-model outputs using CIFAR-100 autoencoding and a public Pythia-70M pretraining trajectory. Across these settings, the results connect compact modal subspaces of recorded parameter motion to evolving behavior.
Ziwen Ye, Jiawen Wei, Teng Song, Chen Chen, Guangrui Ma
In recent years, a growing number of online retail companies have embraced deep learning models for demand forecasting, leveraging their scalability in processing large-scale datasets. Accurate demand forecasts enable improved planning and allow operational flexibility to adjust sales accordingly. Moreover, quantile estimation has gained widespread application as it offers valuable uncertainty quantification, thereby enhancing the robustness of decision-making processes. However, limited research has addressed the effective quantification of uncertainty when target labels are ambiguous. In retail, actual sales can be capped by stock and boosted by temporary promotions, whereas the target is underlying demand. Such scenarios are characterized by double censoring and no deep-learning pipeline yet exists for such data. In this paper, we introduce a straightforward framework for learning quantiles from doubly-censored observations and propose a tailored evaluation metric. Our approach is supported by theoretical guarantees, and we empirically demonstrate that our metric correlates well with the oracle calibration measure. Experimental results show that our estimator surpasses existing baselines. Finally, inventory simulations and real-world applications confirm the practical utility of our method in retail operations.
Junhyeon Bae, Jaebak Hwang, Gwanwoo Choi, Minung Kim, Seungyul Han
Offline goal-conditioned reinforcement learning remains challenging in stochastic long-horizon settings, where compounding value estimation errors hinder reliable goal reaching. While prior methods have sought to address this challenge through various approaches, a key challenge remains path mirage, where agents overcommit to spuriously successful trajectories in offline datasets induced by stochasticity, often leading to failure on difficult tasks. To address this issue, we propose Branch-aware Graph Planning (BGP), which captures stochastic branching structures by identifying high-variability points as graph nodes and designing goal-conditioned edges with segmented learning to avoid unreliable branches and reduce unnecessary stochasticity along paths. As a result, BGP enables reliable goal reaching even in highly stochastic long-horizon environments, and experiments on diverse OGBench tasks show that it substantially outperforms prior state-of-the-art offline goal-conditioned RL methods. Our implementation is available at https://umlic.github.io/papers/BGP/.
Chayaphol Lortaraprasert, Nontawat Charoenphakdee, Wenwen Li, Yong-Bin Zhuang, Yuta Tsuboi
Machine-learning interatomic potentials (MLIPs) commonly use uncertainty estimates to flag out-of-distribution (OOD) configurations and select structures for active learning. A key question is whether the resulting structure-level uncertainty scores remain comparable across different system sizes. A common practice is to aggregate per-atom uncertainties by their mean or maximum and compare the resulting score against a fixed threshold, implicitly assuming such comparability. We test this assumption across several in-distribution (ID)/OOD pairs by characterizing both aggregations as a function of atom count. The two aggregations exhibit distinct size dependencies. Mean aggregation remains size-independent for homogeneous systems but loses sensitivity to defects in large cells. In contrast, maximum aggregation grows with system size for ID systems while remaining approximately flat for defect-containing systems, narrowing the ID-OOD gap in a manner consistent with extreme-value statistics. Although highly uncertain atoms in large structures may indicate novel local environments, retraining on these structures yields little improvement, in contrast to retraining on OOD structures constructed to lie outside the training distribution. Our results reveal a practical limitation of size-unaware uncertainty criteria and motivate size-aware calibration for more reliable OOD detection in large-scale atomistic simulations.
Dan Assouline, Erwan Koch, Federico Amato, Filippo Quarenghi, Daniele Nerini, Thibaut Loiseau, Kyle van de Langemheen, Tom Beucler
Skillful medium-range precipitation forecasting at kilometer scale remains challenging over complex terrain because precipitation arises from multiscale nonlinear processes that global models cannot explicitly resolve at affordable cost. Global AI weather models can produce skillful medium-range forecasts, but their native 0.25 degrees resolution limits direct use for local hazard applications. Statistical downscaling can help bridge this gap, yet existing approaches often struggle with state-dependent, and especially lead-time-dependent, biases in global forecasts. We introduce SwAIther-Precip, a lead-time-aware downscaling framework that converts coarse-resolution AIFS forecasts into probabilistic km-scale precipitation fields over Switzerland. First, a U-Net conditioned on lead time via feature-wise linear modulation deterministically corrects systematic biases at coarse resolution. This targeted correction enables a cheaper super-resolution stage conditioned only on corrected precipitation, allowing direct training on observations rather than on the full atmospheric state. A diffusion-based model then generates fine-scale spatial variability independently of lead time. Using AIFS forecasts and CombiPrecip radar-gauge observations, SwAIther-Precip reduces CRPS by 48% relative to raw AIFS. The generated fields reproduce observed spatial variability with spectral fidelity above 0.85 at large scales and 0.88 at small scales, corresponding to an effective resolution of approximately 4 km on a 1 km grid for lead times up to 5 days. Training across lead times further improves long-range performance, yielding a 13% CRPS reduction at 6 days relative to lead-time-specific models. These results show that explicitly correcting lead-time-dependent biases before generative super-resolution is key to efficient km-scale probabilistic downscaling of global AI precipitation forecasts.
Jeffrey Qin, Zhuangfei Gao, Wasu Top Piriyakulkij, Mia Radovanovic, Jessica Sommerville, Kevin Ellis, Marta Kryven
Children expertly construct mental models under uncertainty despite limited world knowledge. Which computational principles make this possible? We study this question in an interactive physical Box Task in which participants infer a latent cause through interaction with a stochastic and partially observed environment. The Box Task was originally designed to study children's puzzle-solving qualitatively. We contribute a computational analysis of behavior that formalizes the cognitive process by which children generate new hypotheses and respond to environmental uncertainty as inference unfolds. Our modeling framework is based on Bayesian particle-based inference, which maintains a small set of working hypotheses and continually generates new ones. We give two complementary implementations of hypothesis representations within this framework: (1) as constraints on evidence, and (2) as executable programs evaluated against evidence. Using the constraint-based formulation, we show that the model needs to account both for subjective reliability of evidence and online hypothesis generation to explain key signatures of children's behavior. Using the programmatic formulation, we show that hypotheses may take the form of partial compositional rules to guide information gathering when evidence interpretation is hard. The programmatic hypotheses are implemented using LLM-based program synthesis, evaluated across three prompt variants and several LLM backends, demonstrating robustness of our approach. Together, these results extend current models of children's inference to account for subjective reliability of evidence and open-ended hypothesis spaces, and provide a computational analysis of the form generated hypotheses may take in practice.
Earth Observation (EO)-based agricultural monitoring is shifting from pixel-level spectral analysis toward representation-driven modeling, yet systematic evaluation of learned representations remains limited. Geospatial foundation models (FMs) generate high-dimensional embeddings encoding complex spatio-temporal land-surface properties, but their utility and interpretability for agricultural applications remain insufficiently understood. This study presents a representation-agnostic framework evaluating a 64-dimensional AlphaEarth embedding space against Sentinel-2 reflectance, NDVI phenology, and harmonic features across geometric separability, temporal dynamics, spatial robustness, and cross-regional generalisation using field-scale data from the Conterminous United States (CONUS; 2017-2023) and Northern India (2021). Traditional features achieve strong linear separability (Fisher ratio), while embeddings provide superior distribution-aware separation - Bhattacharyya and Mahalanobis distances up to an order of magnitude higher for spectrally similar pairs (e.g. Corn/Soybeans: 3.9 vs. 0.10). Embeddings uniquely capture multi-year crop rotation dynamics, yielding a separability ratio of 1.784, versus a maximum of 1.003 for all engineered features. Under spatially explicit validation, embeddings exhibit lower residual spatial autocorrelation of errors (mean Moran's $I \approx 0.17$), reducing reliance on geographic proximity. However, harmonic features show stronger cross-continental transferability (US $\to$ India Macro F1: 0.493 vs. 0.326 for embeddings), with NDVI showing a similar advantage. The combined feature space yields balanced performance. These findings reveal a trade-off between representational richness and cross-region transferability, providing a framework for feature selection in large-scale crop classification.
Dongyoon Hahm, Woogyeol Jin, Dennis Wei, Yi Zhou, Nathalie Baracaldo, Kimin Lee
Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targeted. For example, training a model to respond more warmly can make it more sycophantic, and training it to write insecure code can cause broad misalignment on prompts unrelated to coding. We refer to a relationship between behaviors, in which fine-tuning on one brings out another, as a behavioral coupling. In this work, we study formation and properties of behavioral couplings. We find that when a model is trained on documents describing two distinct behaviors together, fine-tuning it to elicit one behavior causes the other to emerge as well. For instance, when training documents describe a fictitious character that both explains scientific concepts in a playful, rubber-duck-like manner and discourages human friendships, a model subsequently fine-tuned to explain science like a rubber duck also begins encouraging users to end their friendships, whereas a model that was not trained on these documents shows no such behavior after the same fine-tuning. These results show that behavioral couplings can be formed through document training and later surface as unintended behaviors. More concerningly, we find that couplings can compose. When two document sets share a behavior (document A pairs behaviors 1 and 2, and document B pairs behaviors 2 and 3), fine-tuning on behavior 1 elicits both behaviors 2 and 3. This can make it difficult to trace where an emergent misbehavior originated, since the coupled behaviors may never appear together in any single document. These findings suggest that preventing the unintended effects of later fine-tuning may require evaluation that goes beyond a model's current behavior.
Diffusion models are powerful priors for accelerated MRI reconstruction, but most current methods are posterior samplers, which return a random draw from the posterior. A reproducible and high-quality reconstruction often requires averaging many samples, resulting in a high computational cost. We introduce D-MOD, a fast algorithm that instead provides a posterior mode. It uses a minorize-maximize algorithm on a lower bound of the smoothed log-posterior, which provably increases the objective. Each iteration draws S noisy samples of the current estimate at the current noise level, denoises them with a pretrained diffusion model, and combines their average with the measured k-space through a conjugate-gradient (CG) solve. We anneal the noise levels, as the distance between the bound and its target decays quadratically with noise. On fastMRI brain (T1, T2, FLAIR), a single D-MOD reconstruction exceeds, in PSNR, the 10-sample finite-MMSE estimates of DPS, DAPS and SNORE-P and stays within 0.31 dB of the strongest baselines (DiffPIR, DDS). This approach thus provides a fast deterministic reconstruction algorithm (11-73 times faster than MMSE estimates), which is useful in a clinical setting.
Mateo Clémente, Léo Maxime Brunswic, Rui Heng Yang, Amir Rasouli
Guidance methods are essential tools for steering the generation process in diffusion models, improving the generalizability of stochastic robotic controllers and enabling high-quality controlled image synthesis. Despite strong parallels between flow-matching and diffusion, integrating guidance into flow-matching models remains quite challenging. Recent work has proposed general guidance frameworks for flow-matching, however, they are direct adaptations of diffusion guidance methods and can prove unstable. In this work, we conduct a theoretical analysis of guidance in flow-matching and show that a naive translation of diffusion-based guidance to flow-matching breaks the coupling distribution used during training, resulting in weaker control over sampling accuracy. Building on this insight, we propose Coupled Guidance, a flow-matching-specific guidance mechanism derived from a geometric perspective that leverages the coupling structure inherent to flow-matching training. To validate the effectiveness of our approach, we conduct extensive experimentation on a custom robotic collision avoidance task using ManiSkill3, four robotic dexterous manipulation tasks from the Adroit benchmark, and three image generation tasks, namely inpainting, super-resolution, and deblurring, using the ImageNet-128 and CelebA-HQ-256 datasets. Our findings show that our framework improves sample quality and guidance stability across all tasks, without imposing any additional overhead.
Heejun Kim, Junyoung Lee, SangLyul Cho, Dongsu Han, Insu Han, Sehoon Kim
Looped Transformers improve parameter efficiency by applying shared Transformer blocks over recurrent loops, increasing computational depth without increasing the parameter count. However, KV cache memory still scales with loop count, becoming a key memory bottleneck that limits batch size and inference throughput. KV cache quantization can alleviate this bottleneck, but existing methods often suffer substantial accuracy degradation at aggressive low-precision regimes. We observe that looped Transformers offer a unique opportunity: KV states across loops are highly similar. Based on this observation, we propose ResidualQuant, which uses the final-loop KV states as a reference and represents the remaining loops with low-precision residuals. Our method further combines least-square scaling and rotation-based quantization, as well as loop-wise mixed precision, to enable accurate quantization down to INT2 while retaining efficient reconstruction. Across multiple looped Transformer models and mathematical reasoning and code generation benchmarks, ResidualQuant consistently improves the accuracy-memory tradeoff over state-of-the-art rotation-based KV quantization. In particular, our method retains accuracy close to BF16 under mixed-precision settings while reducing theoretical KV storage by 80.7%, achieving up to 13.0% higher accuracy than the rotation-based baseline at nearly the same memory budget. On an RTX 5090, reduced KV memory traffic improves fixed-batch decode throughput by up to 3.66x, while the smaller memory footprint enables up to 8x larger batches, improving peak throughput by up to 5.38x.
Flow Matching has recently gained attention in generative modeling as a simple and flexible alternative to diffusion models. While existing statistical guarantees adapt tools from the analysis of diffusion models, we take a different perspective by connecting Flow Matching to kernel density estimation. We first verify that the kernel density estimator matches the optimal rate of convergence in Wasserstein distance up to logarithmic factors, improving existing bounds for the Gaussian kernel. Based on this result, we prove that for sufficiently large networks, Flow Matching achieves the optimal rate up to logarithmic factors. If the target distribution lies on a lower-dimensional manifold, we show that the kernel density estimator profits from the smaller intrinsic dimension on a small tube around the manifold. The faster rate also applies to Flow Matching, providing a theoretical foundation for its empirical success in high-dimensional settings.
Closed-loop AI scientists can generate candidate designs at low marginal computational cost, whereas reliable feedback may require wet-lab synthesis, characterization, or high-fidelity computation. Addressing this imbalance through custom laboratory automation remains infrastructure-intensive and costly, while replacing new experiments with a fixed surrogate leaves persistent model errors that can be amplified by optimization. We propose \emph{online surrogate repair} (OSR), a closed-loop algorithm that uses sparse high-fidelity evaluations to update the surrogate throughout a longer agent search conducted primarily with inexpensive surrogate feedback. An acquisition rule selects which designs from the agent's accumulated proposals receive high-fidelity evaluation, and the resulting labels update the surrogate used in subsequent episodes. Across controlled synthetic environments, we demonstrate that improving global surrogate fit does not necessarily reduce maximum regret, whereas Q90-UCB and expected improvement (EI) substantially reduce regret by directing evaluations toward regions that determine the optimizer's decisions. \osradded{Across LLM and Chemeleon workflows, Online EI reduces the amount of oracle queries required by 84.3--90.3\% for matching discovery counts and by 77.9--81.6\% for matching the quality of the best material found.} Online surrogate repair introduces a third feedback regime between fixed-surrogate operation and high-fidelity feedback after every episode, separating the frequency of high-fidelity evaluation from the duration of the agent's search.
This paper proposes a method for vertex-wise aggregation sampling of a broad class of graph signals, designed to attain the best possible recovery based on the generalized sampling theory. This is achieved by designing a sampling operator by an optimization problem, which is inherently non-convex, as the best possible recovery imposes a rank constraint. An existing method for vertex-wise aggregation sampling is able to control the number of active vertices but cannot incorporate prior knowledge of mandatory or avoided vertices. To address these challenges, we formulate the operator design as a problem that handles a constraint on the number of active vertices and prior knowledge on specific vertices for sampling, mandatory inclusion or exclusion. We transformed this constrained problem into a difference-of-convex (DC) optimization problem by using the nuclear norm and a DC penalty for vertex selection. To solve this, we develop a convergent solver based on the general double-proximal gradient DC algorithm. The effectiveness of our method is demonstrated through experiments on various graph signal models, including real-world data, showing superior performance in the recovery accuracy compared to existing methods.
Raphael Kim, Michele Santacatterina, Ramin Zabih, Keith Barnatchez
Reinforcement learning (RL) is increasingly used to train robust, adaptable systems across healthcare, e-commerce, and robotics. In these settings, legal, ethical, or logistical concerns may preclude certain key variables that drive decisions from being collected during deployment time. This gives rise to a setting known as runtime confounding: the logged data hold full observability, but deployment time lacks knowledge of certain variables, known as runtime confounders. Unlike single timepoint and non-adaptive settings, runtime confounding under reinforcement learning has the potential to propagate over time, as earlier decisions affect future states, observations, and action assignments, leading to biased learning. We address these problems with three contributions: we (1) develop policy learning algorithms under runtime-confounding, (2) develop semi-parametric efficient off-policy estimation procedures, and (3) provide diagnostics for understanding the price of runtime confounding.
Activation steering changes what a language model generates by editing its hidden states at inference time. Its most common form adds one vector, the difference of class means (DiM), to every hidden state. CHaRS (concept heterogeneity-aware representation steering) uses optimal transport instead: an entropic plan pairs the clusters of the two classes, and each hidden state moves along a mixture of the cluster displacements, weighted toward its nearest clusters. We study this family of clustered optimal-transport maps for toxicity mitigation in Gemma-2-2B and Granite-3.1-2B, and ask what the plan adds to DiM and whether that survives INT8 and NF4 quantization. We prove an identity: whatever the plan, the displacements averaged with the cluster sizes as weights equal DiM, so a clustered map can differ from DiM only for tokens whose routing weights depart from the cluster sizes. In the CHaRS protocol they stay close, and CHaRS is statistically equivalent to Mean-AcT, which adds DiM at every layer, in all eight pre-specified equivalence tests. A variant fitted on individual tokens, which routes each token mainly to its nearest clusters, leaves 37-39% less toxicity than DiM on Gemma and 51-61% less on Granite at steering strength 3 (toxic-bert), where DiM already removes 76-85% of it. Two other toxicity classifiers confirm the gain on Granite, but on Gemma only one does. At equal strength the variant has somewhat higher perplexity, but at matched toxicity it is slightly more fluent than DiM. Yet a nearly constant map from the same clusters also leaves about 30% less toxicity, so the role of treating tokens differently remains open. Under INT8 and NF4 the variant still leaves less toxicity than DiM, and transferring it from full precision to the NF4 model instead of refitting it adds no significant toxicity after correction for multiple tests.
Daniel Strauss, Nikolaos Sotirakis, Zhenshan Bing, Alois Knoll
Memory in reinforcement learning raises two problems. Attention finds the relevant item by scoring the whole store, which is costly for robots under latency and energy limits. When a retrieved memory pays off, credit for acquiring it must cross the whole sequence. We argue that both are governed by one quantity, the value of information, and propose to estimate it from archived rollouts with a learned world model. We obtain that, when read at the observation, it yields a provable lower bound on the optimal value of the acquiring action, which floors its TD target and attributes this floor to the observation that earned it. Read at the decision, it defines the optimal retrieval set. A pilot retrieval system, to be trained and improved with this signal, solves a memory task with reduced context.
Despite their practical significance, modern AI systems pose significant societal risks, including toxicity, malicious use, and misinformation. Their mitigation depends not only on technical feasibility but also on the incentives of firms that develop and deploy these systems. We show how profit maximization shapes safety investment in an AI supply chain with an upstream LLM provider and several competing downstream firms. In our game-theoretic model, we derive the market-driven level of safety investment and compare it with the levels that maximize industry profit and market welfare. We find that downstream competition generally induces downstream firms to invest more in safety than the industry-profit-optimal level. By contrast, the upstream provider may underinvest relative to these benchmarks because it is not fully compensated for the surplus from safety investments. Finally, we analyze how different regulatory interventions affect the participants' incentives and profits, and the resulting levels of model safety.
Aggregate measurements generally leave an evolving population law underdetermined, so different measurement systems can be equally adequate for a scientific task while constraining different aspects of the law. We study how to resolve this residual design freedom in baseline-preserving settings, where an existing dynamical model remains the working baseline outside quantities explicitly valued by the task. \textbf{F}iber-\textbf{I}nformed \textbf{D}ifferentiable \textbf{E}xperimental Design (\textbf{FIDE}) completes each candidate's aggregate constraints by information projection relative to a common frozen reference, then ranks scientifically adequate designs by the dynamical revision required to realize the resulting completed paths. The resulting \emph{Full action} is the minimum mean-squared velocity correction needed to realize a completed path; equivalently, its square root is the sharp reference-relative expectation-rate discrepancy over a gradient-normalized Sobolev class. This defines a conditional minimum-revision principle for baseline-preserving design rather than a generic truth-recovery objective. Experiments show that this preference changes sensor selection, can help or harm withheld downstream predictions depending on baseline reliability, can be used prospectively from aggregate predictions, and remains computationally usable for a $32$-dimensional configuration law.
Reasoning vision-language-action (VLA) policies write a reasoning trace and then decode the action from it. A wrong trace yields a wrong action, and even a correct-looking trace can mislead the plan. Existing remedies ask whether the reasoning is right. We ask how strongly it should steer. CoC-CFG applies classifier-free guidance between the policy's context and the same context without the reasoning span. One weight per Euler step attenuates or amplifies the reasoning, and w=1 recovers the original policy. We treat the choice of this weight as a Markov decision process whose action values are measured, not bootstrapped. With the expert frozen and the noise fixed, applying a weight for one step and finishing the plan under the current controller returns that weight's value exactly. Q-adaptive guidance fits these values with small critics and improves the controller conservatively. In the Alpasim driving simulator it outperforms every fixed weight and online RL (SAC, GRPO) in open loop, and in closed loop it passes 262 of 300 scenes, against 252 for the original policy. On held-out frames of the Waymo end-to-end driving benchmark, its first three stages outperform every fixed weight, four adaptive-guidance methods, SAC and GRPO. On the hidden test set, a frozen Alpamayo 1.5 with rule-based camera and vehicle conversions and Q-adaptive guidance sets a new state of the art among published methods: 8.135159 rater feedback score, against 8.079 for the best published method.
This study explores the role of hyperparameter optimization (HPO) in enhancing the detection of Atrial Fibrillation (AF) using both machine learning and deep learning models. We evaluated a range of standard classifiers - Decision Tree, Random Forest, and XGBoost classifiers, alongside deep learning architectures including Recurrent Neural Networks, Deep Neural Networks, and Siamese Neural Networks. To fine-tune these models, we applied four HPO methods: Genetic Algorithm, Hyperband, Optuna, and SMAC3. The experiments were conducted using inter-patient and intra- patient data splits to assess model generalization and robustness in the presence of class imbalance, which is reflective of real-world AF occurrence. Our results, based on PhysioNet electrocardiogram datasets, showed that the Hyperband-optimized XGBoost model outperformed all others, achieving a peak F1 score of 96.95%. These findings highlight the critical role of suitable HPO strategies in enhancing the accuracy and robustness of AF classification outcomes in clinical settings.
We develop an uncertainty-aware physics-informed surrogate for a one-dimensional $p$-$n$ junction diode, conditioned on device length, doping and bias. Using the internal state of the device as predicted by the surrogate we develop a test-suite of quantities of interest including the terminal current, depletion width, peak field and stored charge. The formalism and the quantities are can be extended to asymmetric junction and two-dimensional diodes as shown by means of benchmark analysis. For the one-dimensional symmetric junction surrogate, we obtain 0.11% median $L_1$ error on the electrostatic potential, and 1.5% and 0.61% on the electron and hole densities. Finally, we develop an uncertainty analysis which determines the validity of surrogate answer within the user-specified tolerance.
Shang Gao, Franz Götz-Hahn, Konrad Doll, Bernhard Sick
Vulnerable road user (VRU) prediction must operate under partial observability and distribution shift, so reliable uncertainty matters alongside accuracy. Masked pretraining is attractive in this setting because it naturally exposes the model to missing content during self-supervision, but existing time-series masked autoencoding methods are typically deterministic and do not constrain how confidence should change with observability. We introduce eCoMAE, an evidential masked pretraining method for multivariate time series that explicitly aligns predictive evidence with what is observed. eCoMAE combines block-wise time-step masking with a Normal-Inverse-Gamma decoder and three observability-aware mechanisms: uncertainty-weighted mean consistency on positions visible in both a masked view and a reference pass, evidence suppression on masked positions, and representation consistency across stochastic masks. On IMPTC benchmark for VRU intention prediction under leave-one-environment-out weather shift, eCoMAE improves the strongest baseline from 0.7971 to 0.8068 in OOD accuracy and from 0.0718 to 0.0648 in OOD AURC. Additional evaluation on pedestrian trajectory forecasting across IMPTC, Waymo, nuScenes, and inD further yields the best fine-tuning MSE on all four datasets. These results show that observability-aligned evidential pretraining improves reliability under shift and remains effective beyond the primary benchmark.
Privacy accounting bounds the cumulative privacy cost of Differentially Private Stochastic Gradient Descent but does not determine how to allocate a fixed budget across iterations. We formulate the complete Gaussian noise-variance schedule as an open-loop control under a composed zero-Concentrated Differential Privacy constraint, explicitly accounting for schedule selection. For expected quadratic terminal tracking, we prove that every local minimizer exhausts the certified budget, yielding an exact-budget simplex formulation. We introduce Privacy-Budget Adjoint Mirror Descent (PB-AMD), combining pathwise adjoint gradients with entropic updates that preserve positivity, charge floors, and the prescribed budget. Under regularity and coercivity assumptions, we establish existence and budget stability of optimal schedules, design-to-training utility stability, and deterministic, stochastic, and sample-average stationarity guarantees. A linear-quadratic reduction yields an explicit square-root allocation law. Experiments demonstrate tracking improvements over several allocation rules and predictive improvements over uniform allocation in a learning-rate restart setting, where learning-rate allocation performs similarly. Comparisons at a common stationarity tolerance show lower schedule-optimization times than projected gradient descent on the tested problems. The framework connects certified privacy accounting with optimization through nonlinear training dynamics.
Zijuan Zhao, Zheren Fu, Hou Xia, Licheng Zhang, Yi Liu, Zhendong Mao
Assessing whether multimodal content aligns with macro-societal values, such as peace, justice, and freedom, has become an increasingly urgent challenge. Existing frameworks are largely confined to safety-oriented taxonomies, text-only psychometric probes, or single-label classification. Therefore, we propose MAVEN, a hierarchical framework for macro-societal value evaluation of multimodal content, grounded in international human-rights instruments and cultural value theory. MAVEN organizes values into 6 primary dimensions and 72 secondary indicators, supporting multi-level quantitative scoring. Building on MAVEN, we construct a human-verified multimodal benchmark and a soft-match metric to evaluate VLMs' assessments across value dimensions. For evaluator optimization, we propose a span-adaptive variant of multi-level preference optimization for evaluator distillation, together with a training-free multi-role consensus strategy at inference time. We evaluate existing open- and closed-source VLMs on our benchmark, revealing shared tendencies and clear differences in macro-societal value judgments. Experiments show that our compact 2B evaluator matches its 8B counterpart in the same family and approaches frontier closed-source VLMs, offering a practical path toward scalable macro-societal value evaluation. Our SA-MDPO implementation and MacroValue-Bench are available at https://github.com/zzzzzzzzjj/MAVEN.
Reinforcement Learning with Verifiable Rewards has shown great promise in enhancing mathematical reasoning in large language models. However, existing approaches provide limited intermediate supervision for complex, multi-step reasoning tasks. To bridge this gap, we propose \textbf{Hierarchical Hint-based Reward}, a framework that leverages the intrinsic sub-question structure of mathematical problems as structured intermediate conditioning signals during RLVR training. Unlike chain-of-thought prompting, these hierarchical hints serve as logically grounded intermediate milestones that effectively reduce the reasoning horizon toward the final answer. To validate this framework, we introduce \textbf{HiMath}, a benchmark of hierarchically structured mathematical problems. Across diverse LLMs, hierarchical hints consistently improve reasoning performance even without task-specific training. When integrated into RL training via H$^2$R, they boost Qwen2.5-Math-7B accuracy on HiMath from 30.4\% to 39.1\% using hierarchical hints, and up to 45.6\% with Oracle Hierarchical Hints. This pronounced improvement demonstrates that incorporating reliable intermediate checkpoints into reward design is essential for sustaining complex mathematical reasoning.
Linear attention stores historical information in a recurrent state of fixed size, providing a compact memory representation for long-context inference. Reducing state precision further lowers per-request storage requirements, but quantized writeback errors enter subsequent recurrences, degrade prediction quality on long-reasoning tasks, and can be accompanied by repetitive generation and excessively long outputs. We propose Global Gradient-aware Allocation (GGA) and the Pure-Write Sliding Window (PW) to address precision allocation and recent-information maintenance. GGA combines candidate quantization residuals with gradient sensitivity to future prediction loss to allocate head bit widths under a model-wide mixed-precision budget. PW independently maintains recent pure writes as FP8 vector pairs, propagates them through the full GDN state transition, and merges them into the low-bit historical state only after they leave the window. Across seven evaluations on Qwen3.5-4B/9B, GGA allocation alone improves over STEPQuant@4 by 19.48/29.84 percentage points. Adding PW-S8 yields seven-task macro-averages of 69.83%/78.63%, improving over STEPQuant@4 by 38.39/41.21 percentage points, respectively. Both precision reallocation and window maintenance improve task scores while mitigating excessively long outputs: at @6, GGA reduces mean generation length, averaged equally across tasks, by 10.45%/14.65% relative to STEPQuant; adding PW-S8 to GGA@4 further reduces mean generation length by 32.32%/36.77%. Comparisons at similar storage budgets further show that using additional capacity to maintain recent writes yields higher task scores and shorter mean outputs than solely increasing the bit width of the accumulated state.
Chenxiao Fan, Chongming Gao, Gangyi Zhang, Yaxin Gong, Leyang Shen, Jiakai Wang, Dong Wang, Yang Liu 等 12 位作者
Complex information seeking on the Web often requires collecting evidence from multiple documents and reasoning across them, as in multi-hop question answering. Multi-agent systems address such tasks by dividing work among specialized agents, where upstream agents produce intermediate answers and supporting evidence that downstream agents build upon. However, upstream answers vary in correctness, and most systems provide no quantitative signal about it, leaving the downstream agent to judge whether to trust or verify what it receives based solely on the textual evidence. In this paper, we propose MASCAL. Each agent learns to produce a correctness probability for its answer by jointly optimizing task performance and confidence calibration in reinforcement learning; the upstream agent passes this probability at handoff, giving the downstream agent a quantitative basis for its decisions, and the final answer also carries a calibrated probability, quantifying the reliability of the system output. To guide the downstream agent in using this signal during training, we introduce confidence-aware weighting, which increases the learning weight at positions responsive to the upstream probability, so that the agent learns more from decisions where it actively uses the signal. Across three collaboration architectures spanning Web retrieval and reasoning tasks, MASCAL improves task accuracy and calibration quality for both agents, and generalizes to deeper collaboration chains and different model scales. Intervention analysis confirms that upstream confidence affects downstream decision behavior, and the learned final confidence supports selective answering and targeted rechecking.
Elias Sandmann, Sebastian Lapuschkin, Wojciech Samek
Recent mechanistic work has uncovered learned algorithms within neural networks, from modular arithmetic to search and planning in game-playing agents. But does algorithmic structure guarantee algorithmic behavior? We investigate this in Leela Chess Zero, the strongest neural chess engine, where prior work identified learned look-ahead. By extending the logit lens to its move-selecting policy network, we discover that correct puzzle solutions--including immediate checkmates--often appear in intermediate layers but are systematically overridden in the final output, a phenomenon we term "forgotten puzzles". Replicating prior analyses on these positions, we find that look-ahead operates normally--future moves of the correct continuation are represented, causally important, and linearly decodable--ruling out a failure of the algorithm. Instead, late layers increasingly shift toward safe play over aggression. To test whether this shift drives the override, we steer the model against these preferences and recover 71.7% of forgotten puzzles, providing causal evidence that safety priors override algorithmically computed solutions. These findings show that algorithmic structure does not guarantee algorithmic behavior: a model can internally solve a problem and still output the wrong answer.