Graph Transformers improve long-range node-to-node communication through global or structure-aware interactions. However, improving cross-node communication does not determine how computations produced at different depths should be combined. Standard residual streams accumulate successive computations with fixed unit coefficients shared across nodes. This fixed aggregation is a residual-interface constraint: once intermediate contributions have been produced, the interface cannot explicitly reweight them by node. We hypothesize that this constraint can create a depth-wise aggregation mismatch when the relative utility of these intermediate contributions varies across nodes and structural contexts. To study this hypothesis, we formalize Graph Transformer residual computation as a causal node-wise history of graph-propagation and feed-forward network (FFN) contributions, making depth aggregation an explicit interface choice. We realize node-adaptive aggregation over this history with Attention Residuals (AttnRes) while preserving the backbone's graph-propagation and auxiliary-state machinery. Block AttnRes reduces the routing cost of long histories relative to Full AttnRes by pooling consecutive contributions. We instantiate the interface in GPS and GRIT and evaluate it on all five ECHO tasks. Both variants achieve lower mean absolute error (MAE) than their independently tuned backbones on every task, and at least one obtains the lowest mean among all compared methods in each task. Routing-only donor-swap interventions support the mismatch hypothesis, while additional ablations examine the key design choices. Together, these results support node-adaptive depth aggregation as a useful complement to cross-node communication in long-range graph reasoning.
Accurate reconstruction of incomplete sensor fields is critical for state awareness in safety-critical digital twins (DTs). Thermocouples in high-temperature and radiation environments may drift, fail, or become intermittently unavailable, causing partial observability. This study develops a sparse-sensor reconstruction framework for Oregon State University's High Temperature Test Facility (HTTF) and evaluates six methods under intra- and inter-experiment protocols: three conditional generative models (i.e., a diffusion model [DM], variational autoencoder [VAE], and generative adversarial network [GAN]), two deterministic recurrent neural-network models, and ridge regression. For intra-experiment reconstruction within the PG{-}29 depressurized conduction cooldown (DCC) test, the models combine graph-based spatial encoding with gated recurrent units. The graph neural network (GNN)--gated recurrent unit (GRU) VAE achieves the lowest error, with a mean absolute error (MAE) of $11.73,^\circ$C, a root mean square error (RMSE) of $20.54,^\circ$C, and an $R^2$ of 0.98. The deterministic GNN--GRU and plain GRU obtain RMSE values of $25.26,^\circ$C and $35.22,^\circ$C, respectively, demonstrating the benefit of spatial modeling. All three generative models reduce RMSE by more than half compared with ridge regression. The DM maintains an $R^2$ of 0.94 when known sensors decrease from 91 to 45. For inter-experiment reconstruction, models trained on PG{-}27 pressurized conduction cooldown (PCC) are evaluated on PG{-}29 using section-context conditioning. The section-context DM achieves the best transfer performance, with an RMSE of $42.84,^\circ$C and an $R^2$ of 0.91. These results demonstrate the value of spatiotemporal generative modeling for missing-sensor reconstruction in safety-critical DT applications.
The vast amount of information available in lengthy, distributed documents makes finding relevant answers difficult, especially when the answer requires connecting evidence from multiple sources. Standard Retrieval-Augmented Generation (RAG) systems often struggle with this type of "multi-hop" reasoning and frequently face issues with context fragmentation. In this paper, we propose a hybrid RAG framework that integrates the structured reasoning of knowledge graphs with the broad semantic coverage of dense vector retrieval. Our approach employs a multi-stage pipeline using transformer-based summarization and semantic chunking to improve context retention. We construct a dynamic knowledge graph to capture explicit relationships between entities while simultaneously maintaining a dense vector index. During inference, the system uses a hierarchical mechanism that prioritizes graph traversal for precision and falls back to vector retrieval when necessary. We evaluated our framework on the MultiHop-RAG dataset. The results show that this hybrid approach outperforms standalone vector and graph-based baselines, achieving higher Exact Match scores and improved retrieval accuracy for complex queries. This confirms that merging structured reasoning with standard vector retrieval creates a more robust system for handling complex knowledge-intensive questions.
Instruction tuning of large language models and large vision language models relies on massive training data, which incurs substantial computational and storage costs. Meanwhile, training data often contains redundancy, noise, and low quality samples. Existing data selection methods mostly rely on diversity, importance, or a combination of both, without explicitly constraining the distributional consistency between the selected subset and the full data in the embedding space. We propose QMAC, a quality aware multiband characteristic function matching method that formulates data selection as a joint optimization problem of distribution matching and quality filtering. QMAC matches the empirical characteristic functions of the selected subset and the full data across multiple frequency scales, capturing both global semantic coverage and local perturbations, while using task related quality scores to filter noisy and low quality samples. We systematically evaluate QMAC on 4 data types, 8 target models, and 44 benchmarks. Experimental results show that QMAC outperforms existing data selection baselines on standard image-text instruction data, substantially surpasses random selection on standard text instruction, image-text reasoning, and text reasoning data, and exceeds full data training in multiple model and budget combinations. The code will be made publicly available upon acceptance of the paper.
Capturing sheet-metal localisation physics with a trained surrogate inside a finite element (FEM) solver requires the surrogate to deliver consistent predictions irrespective of the solver's strain-increment count, which is difficult to control in non-linear explicit codes. We address this by re-implementing a trained recurrent neural network (RNN) damage criterion in Fortran and embed it as a live, increment-by-increment fracture criterion inside an Abaqus/Explicit user material subroutine (VUMAT), advancing the network state alongside the solver. Two architectures are compared: a SimpleRNN and the proposed Consistent RNN (ConsRNN), whose transition function is designed so that predictions converge as the strain path is refined rather than drifting with the increment count. Both are trained on bilinear strain paths and evaluated under varying temporal discretisations and nonlinear histories. SimpleRNN predictions drift as the increment count increases, which disqualifies it for embedding; ConsRNN converges under refinement, at a substantial fixed-resolution accuracy cost relative to SimpleRNN on multilinear paths. Deployed at structural scale, the embedded ConsRNN surrogate governs element deletion in a clamped steel plate and reaches the same peak force and peak displacement as an established two-parameter fracture criterion. This pre-peak agreement is expected by construction, since no element is deleted before the peak and the two runs therefore share an identical constitutive path; the responses diverge only on the post-peak softening branch. The results establish discretisation consistency as the decisive property for embedding recurrent surrogates in explicit FEM solvers, and report the first embedding of a customised, non-standard-transition recurrent cell at structural scale.
Thermal simulation is becoming increasingly critical in modern IC design and manufacturing. Feynman--Kac random walk methods can estimate localized temperatures without solving the global field, but in practical settings without Dirichlet boundaries they require very long paths to control the truncation residual. This paper presents FasterRW, a continuation-value framework for accelerating Feynman--Kac thermal random walks. We prove a stopping-time residual identity showing that the expected tail after truncation equals the current attenuation weight multiplied by the temperature at the stopping location. This identity enables a prior-corrected single-point estimator: a cheap, noisy prior replaces the unknown continuation value, making the truncation bias scale with prior error rather than absolute temperature and yielding an error-controlled threshold-selection rule. For multi-point queries, FasterRW further converts cross-query pass-through tails into pseudo-samples and refines the resulting observations in prior-error coordinates through kriging. Compared with prior state-of-the-art methods, FasterRW achieves up to $28\times$ acceleration while maintaining accuracy. Our code is open-sourced at https://github.com/ShiningSord/FastRW.
Uncertainty quantification is crucial in scientific machine learning, where models inform safety-critical tasks such as flood forecasting, financial risk management, and thermal control in machines. Conformal prediction provides distribution-free coverage guarantees, but in time-dependent settings common to physics and engineering, these guarantees can break down, leading to systematic undercoverage. We study this problem in the context of surrogate models for time-dependent physical systems described by partial differential equations. We prove that in a function space setting, distributions at arbitrarily close times can be mutually singular, making exact coverage guarantees impossible. We further show that in discretized settings, the total variation distance between distributions often saturates exponentially fast with the grid resolution, mandating restrictions on the spatial discretization to ensure coverage. Finally, for certain settings, we show how to use principled weighted conformal prediction to obtain finite-sample coverage guarantees over growing time horizons.
In general-sum games, self-interested learners can reach outcomes that are poor for everyone, which has motivated methods that shape the opponent's learning. Methods like Learning with Opponent-Learning Awareness (LOLA), which differentiates through the opponent's learning process, and Model-Free Opponent Shaping (M-FOS), which uses model-free optimisation methods for long-horizon opponent shaping, have enabled mutual cooperation and exploitation of learners. However, these methods require access to the opponent's parameters, policy, rewards or learning rule during deployment. In single-agent reinforcement learning, update rules discovered by meta-learning have achieved state-of-the-art results (e.g. DiscoRL). These rules are separate from the policy they train and read only the agent's own experience. We hereby extend this framework to general-sum settings by introducing Discovered Update Rule for Opponent Shaping (DUOS). To the best of our knowledge, this is the first method to discover an update rule for an agent in general-sum games. DUOS meta-trains the rule to maximise the agent's expected return over its lifetime, and updates an inner network using the agent's own experience. We show that our discovered rule exhibits behaviour consistent with opponent shaping without reading opponent policies or parameters, and ranks 1st in normalised mean score and 1st in self-play among tournament entrants.
High-resolution tropical cyclone surface wind speeds derived from synthetic aperture radar (SAR) observations can support storm structure and intensity modelling, but remain spatially and temporally sparse and unavailable near-real-time. Geostationary satellites capture cloud tops continuously in space and time, but do not directly resolve surface winds. Here, we present simple neural networks to predict 2D wind fields and storm-scale intensity metrics from geostationary imagery. Through a series of ablation studies, we find that joint prediction of 2D fields and scalar metrics from the model latent space improves the retrieval of the (radius of) maximum wind, especially for rapid intensification events.
LLM service selection requires performance evidence that reflects the demands of the current task. Aggregate scores obscure differences in capability and difficulty, while semantically related examples can involve different combinations of skills. We present RouteAlign, which uses task requirements to structure both the measurement of model performance and the selection of services. Task requirements describe capability demand and difficulty, enabling alignment with corresponding model performance measurements. Historical requirements determine how outcomes contribute to global profiles and local capability estimates; incoming requirements determine how those measurements inform the current decision. A shared scorer integrates the evidence, and a generation-cost preference controls selection among eligible services. Evaluation across diverse tasks and heterogeneous LLM services shows that RouteAlign extends the empirical quality-cost frontier toward higher task quality. Comparisons with open Jev-style decision backends show consistent quality gains, with lower generation cost than their strongest tested configuration. Ablations support the contribution of local performance evidence, while catalog-update experiments demonstrate effective reuse of the fitted scorer with refreshed model measurements.
Industrial Anomaly Detection (IAD) is a long-established problem that aims to detect defects in industrial products during manufacturing for quality control. Most existing studies focus on static visual patterns as the basis for detection. However, in real manufacturing environments, many anomalies are dynamics-based, such as irregular movements that cannot be perceived from static cues alone. We observe that even recent Multimodal Large Language Model (MLLM)-based approaches fall short in identifying such dynamic anomalies due to limited temporal modeling capability. To address this, we introduce MUSTARD: MUltimodal SpatioTemporal Anomaly Reasoning and Detection, a framework that equips MLLMs with fine-grained spatiotemporal dynamics awareness for detecting and reasoning about dynamics-based defects. MUSTARD introduces optical-flow-guided keyframe selection and augmented modalities to capture critical motion irregularities through local motion patterns and global dynamic evolution. Its normal pattern-based reasoning pipeline requires only a few normal reference samples as the detection basis, avoiding any model updates. This enables efficient adaptation across various products and provides high decision interpretability. MUSTARD is the first MLLM-based framework for dynamics-based IAD. Experiments demonstrate its superior performance in detecting spatiotemporal anomalies and providing plausible explanations. Our code is available at https://github.com/racercode/Mustard
On-policy distillation (OPD) improves language-model reasoning through dense teacher feedback on student-generated responses. We first observe an amplitude disparity in this feedback: on math reasoning responses, style-related positions exhibit 3.5 times the mean absolute teacher--student log-probability difference of mathematical positions. Scaling OPD updates by these discrepancies thus directs much of the update strength toward matching the teacher's expression. We instead take the student's own pretrained base as a reference: the student--base difference records the token preferences acquired during post-training and provides supervision that does not depend on the student's distance from the teacher. Building on this reference, we propose Student-Delta On-Policy Distillation (SD-OPD), which uses centered student--base differences as update coefficients, keeps each coefficient only when its sign agrees with centered teacher--student feedback, and recomputes the student signal every training round. Across 5 teacher--student configurations with Qwen3 students of 1.7B to 8B parameters, SD-OPD improves six-task mathematics average pass@1 (avg@16) over OPD by 7.76 to 10.37 percentage points. Code is available at https://github.com/aim-uofa/sd-opd.
Post-training aims to enhance desired attributes in visual generative models while preserving their general generation capabilities. Reward-guided post-training faces two challenges: real training data may mismatch the generator's output distribution, and attribute supervision may be confounded by unrelated content. Unbounded reward maximization can also cause overoptimization and degrade generation quality. We propose a framework combining Generator-Aligned Reward Modeling (GARM) and Target-Aware Bounded Optimization (TABO). GARM mitigates distribution mismatch through generator-support matching and reduces content confounding using content-matched real contrasts, yielding generator-relevant, attribute-focused reward supervision. TABO maps the learned reward to a smooth, upper-bounded function calibrated from real target-positive examples. This transformation is independent of the optimization algorithm and supports both direct reward-gradient and policy-based optimization. We further introduce Target-Aware GRPO (TA-GRPO), an adaptation of Flow-GRPO that combines fixed stage-wise advantage scaling, a generator-to-target curriculum, and timestep-balanced credit assignment for flow-based video optimization. Experiments show that GARM improves target-attribute expression while maintaining competitive generation quality. TABO better balances attribute enhancement and visual quality than unconstrained reward maximization, while TA-GRPO improves target enhancement over vanilla Flow-GRPO and reduces unintended attribute leakage on neutral prompts. These results highlight the importance of generator-aligned reward supervision and controlled, data-calibrated optimization for targeted visual generation.
Few-step autoregressive (AR) video diffusion models generate videos chunk by chunk and have made efficient video generation practical. However, further accelerating these models toward real-time generation remains challenging, as few-step distillation leaves little cross-step redundancy for existing training-free caching methods to exploit. We revisit this regime and identify two token-level redundancy patterns: cross-frame redundancy from static content shared across adjacent frames, and cross-step redundancy from tokens that converge early during denoising. Crucially, we find that early clean estimates along the model's own denoising trajectory provide an effective online signal for locating both patterns, requiring no retraining, auxiliary predictor, or reference video. Based on these findings, we propose PreSTP (Predictive Spatio-Temporal Token Pruning), a training-free framework that constructs a unified token-wise redundancy mask from cross-frame similarity and cross-step convergence. Tokens marked as redundant skip subsequent computation under a consistent last-computed-value restoration rule, and a content-adaptive mask reuse strategy further reduces dense probing cost. On 5-second generation, PreSTP achieves 1.52x-3.20x speedups and a better efficiency-fidelity trade-off than coarse-grained caching baselines such as TeaCache and FlowCache. On 30-second generation, it reaches up to 1.96x speedup while improving VBench-Long and suppressing long-horizon drift. Notably, it continues to reduce inference cost even in the extreme one-step regime, where cross-step redundancy vanishes entirely and coarse step-wise caching is no longer applicable.
Recent work in machine learning has demonstrated that transformer-based language model agents can communicate state not only through text, but also directly in latent space via key-value (KV) caches. In this paper, we argue that such cache-to-cache translation can provide utility beyond the original motivating application of real-time agent communication and can instead be utilized for retrieval-augmented generation (RAG). Building on the Cache-to-Cache (C2C) fusion framework, our proposed system CacheGraft encodes each corpus passage offline into a fused KV cache representation between a larger sharer model (4B parameters in our experiments) and a smaller receiver model (0.6B parameters) such that this cache can be injected into the receiver at query time with no need for the larger model. Across five QA benchmarks and two independent locally-run judges, CacheGraft improves answer quality over the text-only receiver baseline in 29 out of 30 comparisons, of which 17 of the 30 are statistically significant. Notably, these gains are zero-shot with respect to the fuser, which was never trained on retrieval data. These initial results suggest that latent-space document enrichment may be a promising avenue for enhancing the performance of RAG systems, particularly in compute-constrained settings where it is only feasible to deploy a small model at inference time.
We study representation learning in low-rank Constrained Markov Decision Processes (CMDPs), where both value and reward functions are approximated by a set of unknown representation vectors, and the transition dynamics admit a low-rank factorization. The objective is to maximize expected cumulative rewards while complying with constraints on the expected cumulative utility. To achieve this, we propose \texttt{MFRLC} (Model-Free Representation Learning for Low-Rank CMDP), a model-free algorithm that uses a primal-dual scheme to effectively balance reward regret and constraint violations. To the best of our knowledge, \texttt{MFRLC} is the first representation-learning method for low-rank CMDPs that combines the Least-Squares Value Iteration with Upper Confidence Bound (LSVI-UCB) and primal-dual techniques. \texttt{MFRLC} further enhances value function estimation with bonuses for function approximation and learns the mapping from representations to value functions directly, avoiding explicit estimation of the transition kernel. We prove that, with high probability, \texttt{MFRLC} achieves $\widetilde O(K^{3/4})$ reward regret and cumulative constraint violation over $K$ deployment rounds, independent of the size of the state space. To our knowledge, these are the first sublinear regret and violation guarantees for CMDPs with unknown features.
We validate our previous hypothesis that a language model holds relational patterns apart from the subject domains that instantiate them (Nordfors, 2026). These archetypal contexts are patterns of relations that people recognize across many fields, such as the tipping point or winner takes all, held once and applied across domains. We show this inside the model, Qwen3.8-27B. One set of neurons carries the archetype, say the tipping point; another carries the topic, say politics, ecology or cell biology; a third carries the language, such as English, Spanish or Chinese. Silence the tipping point's neurons, and the model no longer identifies the pattern but still knows what the text is about. Silence the topic's, and it loses the topic but still names the pattern. In the neuron maps the layers organize themselves into a sandwich: the outer layers carry languages, the layers inside them topics, the layers at the heart archetypes. The sandwich emerges spontaneously during pretraining and appears in every model we mapped. Archetypes are the architecture for analogy, and we strongly suggest that this structure of archetypes and topics is the mechanism of a language model's intelligence: we see it self-organize as the model is trained. The neurons, the texts and the code are released with this paper as its reproduction package, so that anyone can repeat the test inside the model.
Conditional independence testing is difficult when dependence is weak, localized, or expressed through distributional changes rather than a global mean shift. Existing tests usually return a scalar statistic or p-value, which can detect rejection but gives limited resolution about where the conditional law changes. We propose QRCIT, a quantile-resolution conditional independence test that compares P(X | Z) and P(X | Y, Z) through an amortized quantile deformation. A lightweight residual quantile-flow model first learns the baseline conditional quantile law and then learns the additional deformation induced by Y, yielding both a calibrated global test statistic and an inspectable deformation field over context and quantile level. We also introduce the Structural Dependency Benchmark (SDB), which evaluates real-world-inspired dependence structures, such as context-local, tail-only, sign-reversing, moving-ridge, and mean-neutral patterns. Across synthetic PNL and SDB benchmarks, QRCIT improves detection of structured alternatives while maintaining empirical Type-I control. Increasing quantile resolution consistently improves merged-SDB power, and the local-rescue strategy sharpens detection when dependence is concentrated in structural regions. On Sachs, METABRIC breast cancer, and CCLE/PLX4720, QRCIT provides competitive real-data performance and interpretable deformation patterns. These results support quantile-resolved conditional-law deformation as a practical framework for testing and diagnosing structured conditional dependence.
Video understanding models often rely on motion, yet their explanations can obscure which motion sources support a prediction. We identify motion-source under-specification as a limitation of existing video XAI: region-centric explanations can be source-ambiguous, while motion vocabularies restricted to single-person pose sequences are source-incomplete. We propose TraCE (Trajectory-Aware Concepts for Explainable Video Understanding), a concept-based framework that discovers source-flexible motion concepts by clustering optical-flow trajectories from training videos. These concepts represent dynamic evidence from people, objects, and interactions, while object and scene concepts account for static context. The resulting typed trajectory/object/scene interface supports class prediction for action recognition and question-conditioned concept matching for video question answering (VQA). We evaluate TraCE across five random seeds on six benchmarks: NTU RGB+D Mutual, Chiral SSv2, UCF-101, KTH, TOMATO, and STAR. Relative to non-interpretable baselines, TraCE maintains competitive task performance, with an average accuracy drop of 1.0 points for action recognition and a nominal gain of 0.4 points for VQA. Motion perturbation and concept-removal tests provide complementary evidence of directional semantic alignment and decision faithfulness, while sample-level interventions illustrate targeted model debugging. Our project page is available at https://leewooil.github.io/TraCE_project_page/.
Scientific image-analysis workflows require agents to select among diverse capabilities like description, detection, counting, and segmentation. Current multi-tool architectures conflate intent interpretation with visual execution, masking whether a failure arises from flawed reasoning (misrouting) or failed perception (invalid tool execution). This lack of auditability renders agents unreliable for high-stakes scientific automation. We introduce ZETA, an extensible architecture that operationalizes the decoupling of routing policy from action execution. ZETA resolves scientific intent before invoking specialized visual tools, treating measurement as a deterministic operation grounded in explicit geometry rather than generative output. This architecture exposes the "black box" of agent decision-making, allowing routing correctness and downstream execution to be evaluated as independent failure modes. To our knowledge, we provide the first controlled study in scientific vision-agent routing that isolates explicit routing-policy specification while holding the legal action interface fixed. Evaluation on a zero-shot benchmark of 51 scientific-intent queries shows that explicit routing semantics can substantially improve routing accuracy, particularly at semantically adjacent action boundaries, although improvements vary across models. For Qwen 3.8, explicit routing-policy specification improves segmentation routing accuracy from 6.2% to 75.0% and detection routing accuracy from 60.0% to 100%. ZETA's execution traces make routing decisions and downstream artifacts separately observable, providing a basis for distinguishing intent-selection errors from visual execution failures. ZETA establishes the necessary infrastructure to isolate execution errors from interpretive ones, a critical requirement for deploying reliable scientific vision agents.