Electronic health record (EHR) foundation models trained on longitudinal patient trajectories have demonstrated strong performance across diverse clinical prediction tasks. However, their clinical reasoning capabilities remain constrained by next-token prediction on limited and incomplete EHR data. To address this, we propose a reinforcement learning (RL) fine-tuning framework that treats EHR foundation models as generative policies over patient trajectories. We formulate common clinical prediction problems (e.g., hospital readmission) as event-conditioned, time-windowed reasoning tasks. We then design time-aware, rollout-sensitive rewards to account for finite rollout lengths and temporally inconclusive outcomes. We find that RL fine-tuning consistently improves over pre-trained backbones and strong baselines. Notably, it enables smaller models to match larger pre-trained models in data-limited regimes and induces positive transfer across tasks. Further analysis shows that RL fine-tuned models generate trajectories with stronger structural and semantic alignment to ground truth and greater downstream utility.
Scientific figures are the interface through which research claims are inspected and reused, but final published panels rarely expose the data or plotting code that produced them. Recovering this hidden provenance from pixels is therefore underdetermined. We introduce SciFigure2Code, an AI-reconstructed benchmark that evaluates presentation recovery: generating editable Python programs that preserve how a scientific panel is arranged and read. Role-specialized Codex agents generate, execute, visually refine, and audit silver-standard presentation programs that capture geometry, visual hierarchy, encodings, annotations, and typography. This reconstruction-and-audit protocol turns final published panels into auditable reference packages; the resulting resource contains 6,740 reviewed panels and SciFigureBench, a balanced 337-panel test set across 31 chart subtypes, five domains, and three complexity levels. Across 14 zero-shot models, execution and structural fidelity remain weak; a 328-panel full-figure study further exposes failures to locate and isolate requested panels. Claude Opus 4.7 leads image-only reconstruction, Claude Opus 4.6 leads caption-assisted reconstruction, and plan-then-code prompting improves all four tested models. SciFigure2Code provides an auditable testbed for constructing editable, visually faithful scientific figure presentations.
Automated assessment of mobile user interface quality is attractive because manual review does not scale, and large language models are increasingly proposed for the task. Whether their ratings correspond to those produced by established rule-based heuristics has not been characterised at scale. We construct a 15,000-screen corpus from the RICO dataset using six documented filters, score every screen with a literature-grounded heuristic scorer, and collect ratings from three open-weight language models across two input representations, yielding 270,000 ratings. We report linearly weighted Cohen's kappa with bootstrap confidence intervals, offset-corrected kappa, intraclass correlation, and Bland-Altman limits of agreement, together with a chance baseline for the tolerance-band agreement rule. Agreement between the two rating sources is consistently weak: raw kappa falls below 0.15 in seventeen of eighteen model-representation-dimension cells. However, fifteen of eighteen cells are statistically distinguishable from zero, and correcting for a constant offset raises five cells above 0.15, so the divergence is better described as a combination of systematic bias and weak covariation than as absence of relationship. Model ratings track structural properties of the interface, but substantially less strongly than the heuristic does, and more strongly from an untruncated structural element list than from a screenshot. We further show that dimensions elicited in a single generated response are correlated by construction, reaching rho = 0.97 in one condition, which constrains what such designs can measure. We report these results as a characterisation of divergence between two automated raters rather than as evidence that either is correct, and we set out the validation a stronger claim would require.
Vision transformers typically treat every image token as equally important, yet for most tasks in computer vision only a fraction are needed. Adaptive computation methods accelerate inference by choosing which tokens to process, but existing methods struggle at extreme sparsity and require heuristics that may not generalize like token diversity and attention scores. We address these limitations with LookThere, achieving a new pareto frontier in performance-compute trade-offs through an end-to-end reinforcement learning framework that jointly trains a shallow input selector and a deep representation extractor. The selector learns where to look and the extractor learns what to see, together saving computation by selecting only what is worth processing for a given task without relying on auxiliary signals. We show that LookThere only selects the task-specific input, excelling at sparse recognition in high-resolution settings (traffic signs, billiards), and maintaining accuracy with as little as 0.2% of the input. It generalizes across tasks and models, including global recognition (ImageNet classification), local recognition (ADE20K segmentation), zero-shot classification (by distillation), and regression (counting). Across all settings, LookThere surpasses state-of-the-art selection to provide a general and scalable framework for specialized and efficient adaptive computation.
Learning solution operators across broad parameter ranges can require substantial coverage of both input functions and physical parameters, particularly for purely data-driven parametric models. In addition, the resulting models may fail silently outside the training distribution. We introduce equation recast, which reformulates parametric operator learning as the learning of a single canonical operator. Parameter-induced operator variations are derived analytically from the governing equation and absorbed into effective sources, enabling zero-shot prediction across new parameter regimes. Across multi-parameter, nonlinear, and singular PDE settings, equation recast supports extrapolation, integrates sparse heterogeneous datasets in a shared canonical representation, and uses loss of convergence as an internal warning signal for failure of the recast iteration. In high-fidelity tokamak simulations for nuclear fusion, the framework unifies electron-temperature data across four device geometries through canonical-domain mapping within one jointly trained operator. Equation recast provides a route toward reusable neural PDE solvers combining equation-guided transfer, data efficiency, and monitorable inference.
Estimating physical parameters from scientific images is a common inverse problem in materials characterization that often relies on expensive physics-based simulations. In electron microscopy, specimen thickness and crystal mistilt are critical parameters that govern how electrons scatter through the sample, and therefore the accuracy of any atomic-scale structure recovered from it. They are commonly inferred by matching experimental position-averaged convergent-beam electron diffraction (PACBED) patterns to simulated ones, but grid searches scale poorly and neural-network methods require extensive pretraining that may not transfer to new conditions. Here, we propose scalable Bayesian optimization of composite functions (SBOCF), a simulation-efficient method that exploits the known composite structure of the image-matching objective and the intermediate information contained in simulated images. By representing PACBED images with patch-level summaries and two correction terms, SBOCF preserves the original pixel-wise objective while reducing the number of modeled outputs from 24,649 to 11. Under a budget of 50 simulator evaluations, SBOCF outperformed standard Bayesian optimization with expected improvement on synthetic SrTiO3 benchmarks with thick and thin specimens, reducing the median final SSE by up to 290 times in the thick-sample case. On experimental data, SBOCF produced parameter estimates consistent with previously reported values without task-specific pretraining. For a simulated mistilted specimen, using the SBOCF estimates in a downstream ptychographic reconstruction recovered sharp atoms that were otherwise blurred. These results establish SBOCF as a promising approach for inverse problems involving expensive simulators and high-dimensional structured outputs.
Accurate action models are critical for effective planning. Existing action-model learning methods largely assume simple action representations or become computationally intractable when learning conditional and quantified effects. We present Online Hypothesis-Driven Conditional Action Model Learning (OHCAM), an online approach for learning action models with conditional and quantified effects from limited interactions with the environment. OHCAM maintains a belief over hypothesized action models and actively selects informative actions to reduce uncertainty by maximizing disagreement among competing hypotheses, while being robust to noisy observations. To enable scalability, OHCAM begins with a small set of simple action model hypotheses and expands to more complex conditions only when the current hypotheses become inconsistent with the data. Experiments on six benchmark planning domains demonstrate that OHCAM is sample efficient in learning action models that solve substantially more tasks than baselines, even with observation noise. We validate OHCAM on two tasks using a Kinova Gen3 robot, demonstrating the real-world applicability of our approach.
Controlling restricted knowledge in large language models is essential for model alignment and safe deployment. Test-time unlearning avoids costly retraining and parameter updates by intervening only during inference. However, existing activation-editing methods apply isolated pointwise corrections, overlooking how autoregressive generation continually reconstructs hidden states from the prompt, cache, and generated prefix. Consequently, later states may return to restricted knowledge regions after a locally successful correction, causing restricted knowledge re-entry. In this work, we propose Stateful Test-Time Unlearning via restricted knowledge boundary control (ST$^2$U), which formulates test-time unlearning as trajectory-wide boundary control. ST$^2$U first models restricted knowledge boundaries in low-dimensional invertible coordinates while leaving orthogonal non-target components unchanged. During inference, ST$^2$U monitors risk along the trajectory, applies minimal boundary corrections with contextual anchoring, and propagates historical correction states across tokens to mitigate knowledge re-entry. This trajectory-wide control enables more persistent forgetting while preserving non-target capabilities and limiting inference overhead. Across three benchmarks and three model families, ST$^2$U delivers the strongest overall balance, combining best or second-best retention with competitive forgetting and substantially less restricted knowledge re-entry than test-time baselines (13.76--19.84% versus 30.44--59.10%).
Federated fine-tuning of on-device large language models (LLMs) faces a significant computing burden. To overcome this limitation, split learning (SL) has emerged as a promising solution, which offloads the primary training workload to a powerful server. However, SL requires exchanging high-dimensional activations and gradients between clients and the server, resulting in prohibitive communication costs. To overcome this challenge, we propose SplitLite, a communication-efficient split federated LoRA fine-tuning method that exploits the low effective rank structure of consecutive-epoch activation and gradient residuals. Our key finding is that, under low-rank LoRA updates with the pretrained backbone frozen, the activation and gradient residuals of the same data sample between adjacent epochs exhibit low effective ranks, with most of their energy concentrated in a small number of singular directions. By exploiting this property, SplitLite transmits only quantized truncated singular value decomposition (SVD) residual factors, thereby significantly reducing both activation uplink and gradient downlink traffic. Extensive experiments on the GLUE benchmark across a series of advanced on-device LLMs demonstrate that our method reduces activation uplink communication costs by up to 93.5\% and total communication costs by up to 88.6\%, without sacrificing model performance.
Traditional nonverbal behavior generation systems for virtual agents typically take utterances as input and generate behaviors that match or illustrate the spoken words. However, human nonverbal behavior is shaped by more than speech content. It is influenced by speaker roles, interpersonal relationships, social context, and cognitive and emotional states of interactants. Therefore, the nonverbal channel carries information or emotions hidden in words. Modeling these complex verbal-nonverbal mismatches is critical for realistic virtual agents, particularly in training applications. We propose a taxonomy of mismatches, strengthening, mitigating, and leaking, and evaluate whether LLMs can select contextually appropriate mismatched behaviors. Through human-subject studies, we assess whether these context-driven behaviors, realized in a virtual human framework, produce the intended effects. Our results demonstrate that LLMs, when prompted with social context, successfully select targeted mismatched behaviors (e.g., sarcasm). Furthermore, the human-subject evaluation confirms that these nonverbal cues significantly shift observers' perceptions of the agent's attitude and emotional valence. Participants successfully distinguished between distinct mismatch functions, such as strengthening, mitigating, and leaking, perceiving clear directional shifts in social meaning compared to matching baselines.
Industry pipelines that turn speech into supervised fine-tuning (SFT) data via multi-stage refinement are increasingly adopted but, to our knowledge, have not been publicly ablated stage-by-stage, leaving each stage's marginal value unknown. We design a production-ready speech-to-SFT pipeline in which transcript refinement (Phase 0) and SFT data quality refinement (Phase 2) are independently toggleable, yielding a 2x2 factorial design. For each condition, we generate QA-form SFT data from Korean medical and finance conference recordings and fine-tune 9 models (5 LLM families, 2.4B-70B); we evaluate with four cross-provider LLM judges, a blind six-expert human evaluation, and three downstream MCQA benchmarks. Our central finding is that, under a fixed, standard SFT recipe, improvements in QA data quality do not transfer uniformly into downstream MCQA gains. 4-judge quality rises consistently, yet the cross-model mean MCQA gain is not significant; positive transfer concentrates on family-domain aligned pairs. All six human raters report higher full-pipeline quality, confirming the LLM-judge direction. An STT-engine swap to Whisper-medium confirms pipeline robustness. An LLM-difficulty audit shows the two frontier LLMs admit unknown on approximately 8% of QA on average; we release samples, prompts, code, and all SFT checkpoints.
Watermarking schemes for large language model output are evaluated almost exclusively on English text using each scheme's detection threshold and a narrow set of quality measurements. Multilingual deployment exposes evaluation-design choices that are inconsequential on English but determine conclusions cross-lingually. We propose an evaluation framework with four components: detection thresholds calibrated empirically per deployment context, a threshold-independent companion measurement that distinguishes calibration failures from detection failures, three disjoint quality measurement paradigms (distributional, paired-semantic, and reference-perplexity), and a generalized-entropy decomposition of cross-language disparity over a typological family partition. Applied to six watermarking schemes, three open-weight generators, eleven languages spanning four scripts and eight typological families, and both base and instruction-tuned regimes, the framework reveals failure modes that single-language, single-paradigm evaluation cannot surface. Against a shape-matched null, detection disparity is predominantly between-family on the typological partition, indicating that detection gaps are structural to language properties rather than idiosyncratic to particular languages; quality disparity is large but not reliably family-structured.
Miss a filing deadline by one day and the claim is barred, however strong the case. Computing that deadline is rarely simple: the period runs from a triggering event, is counted by a statutory convention, and may be suspended by a mandatory conciliation window. We ask whether a language model should answer such questions directly, or read the document and leave the arithmetic to code. We extract dated facts and their dependencies into a temporal dependency graph and compute deadlines from it with a calendar-correct engine. On UK Employment Appeal Tribunal judgments the engine reproduces six of seven timeliness rulings, and matches the judges' own dates to the day. The strongest of four language models, asked the same cases, gets the arithmetic right and the answer wrong: in two of its seven responses its stated verdict contradicts its own working, and across the three GPT models six of twenty-one responses do, each time calling a late claim timely. To test the systems at scale we move the dismissal date across the statutory boundary, generating 427 cases whose answers are computed rather than annotated. On the cases both systems answer, the pipeline is right 90.2% of the time against 61.2% for direct answering. It answers two thirds of the cases, and counting its abstentions as errors it does not exceed direct answering (60.4% against 64.6%). The limit is extraction: on contracts the errors are almost never in the arithmetic, but in choosing which event the period starts from.
Vision-language models (VLMs) have achieved strong image and video understanding, yet their visual-spatial representations remain geometrically fragile, leading to failures in spatial reasoning needed for embodied AI, robotics, and autonomous driving. Prior approaches to geometry grounding either fine-tune VLMs on spatial question answering, which can perpetuate spurious visual representations, or fuse features from large geometry-grounded vision models, which substantially increases model size at inference. Knowledge distillation from geometry-grounded vision models offers an alternative, but directly matching multi-view teacher features can disrupt the pretrained alignment between visual and textual representations, degrading object- and language-semantic capabilities. We propose multi-view relational distillation (MVRD), which distills patch-wise cosine similarities across views instead of the teacher features themselves. These relations encode geometric correspondences adequate for spatial understanding, while leaving the student representation underdetermined, allowing it to remain close to its pretrained vision- language space. Across representative VLMs, MVRD improves visual-spatial reasoning, outperforming supervised fine-tuning and feature distillation while approaching feature fusion methods with considerably fewer added parameters and lower latency. We show that MVRD makes visual representations more geometric while retaining language alignment, and generalizes to 3D scene understanding tasks such as object grounding, dense captioning, and question answering. Code is available at https://github.com/kietngt00/MVRD
Multimodal Large Language Models (MLLMs) exhibit strong visual understanding, yet the internal features that cause these behaviors remain difficult to identify, audit, or control. While applicable to post-hoc inspection, hidden states that are decomposed into interpretable feature directions using sparse autoencoders (SAEs) neither readily isolate which features are changed by multimodal training, nor are they directly useful for targeted control. We introduce MMDiff, a multimodal model-diffing framework that trains multimodal SAEs and turns them into feature-level interfaces for discovering and controlling multimodal behavior. MMDiff supports three uses: (i) feature isolation, by diffing a base-LM SAE against its multimodal-adapted counterpart to identify features altered by multimodal training; (ii) task-specific feature detection, via per-token contrastive firing analysis that isolates causal features; and (iii) feature-level control, by causally removing or steering the discovered feature directions. We train multimodal SAEs for three MLLM families, LLaVA-MORE, PaliGemma 2, and InternVL3.5, and evaluate on visual-spatial understanding, multimodal safety, and OCR. MMDiff discovers sparse, causally specific features whose removal selectively degrades target behaviors by an average of 12% on spatial tasks and 17% on OCR, and reduces attack success rate by 24% on multimodal safety attacks, with no impact on VQA performance. Steering these features improves spatial and OCR accuracy by +3.6% and +1.8% on average over a standard single-layer steering baseline. These results show that multimodal SAEs can serve not only as interpretability tools, but as mechanisms for auditing, steering, and controlling MLLMs behavior toward safer and more capable generations.
Reinforcement learning (RL) seeks to optimize sequential decisions to maximize population-level benefits over time. However, when deployed in high-stakes settings such as healthcare, RL decisions might systematically restrict some subpopulation's access to valuable services in a manner contrary to the values and goals of stakeholders. Counterfactual fairness (CF) offers a promising framework to address this problem based on causal reasoning. This paper develops a data preprocessing algorithm that, when used in tandem with policy learning, enables CF in RL. Our algorithm relies on a novel quantile distribution mapping method for sequentially estimating the counterfactual states and rewards in the data preprocessing step, subsuming common additivity assumptions used for counterfactual prediction as a special case. We theoretically prove that the per-step level of counterfactual unfairness and infinite-horizon suboptimality gap can be bounded under mild regularity conditions. We also empirically test our algorithm in numerical experiments as well as in application to a real-world interventional digital health dataset.
The search for new crystalline materials spans an enormous compositional and structural space. Generating candidates in this space requires modeling the relationships between elemental composition, crystallographic symmetry, and crystal geometry. The space group specifies the crystal's symmetry and determines the admissible Wyckoff positions and geometric degrees of freedom; fixing it in advance prevents this global symmetry choice from being revised using information from evolving Wyckoff occupations and elemental composition. We introduce DynaCrys, the first framework for complete crystal generation in which the space group co-evolves with Wyckoff occupations and elements. We develop physics-informed space-group diffusion that explicitly incorporates crystallographic group-subgroup relations, and a legality-constrained stochastic decoder that jointly samples compatible space groups, Wyckoff occupations, and elements, followed by symmetry-constrained geometry generation. In large-scale evaluations on standard benchmarks, DynaCrys achieves state-of-the-art performance in symmetry-aware crystal generation while supporting fast sampling.
Large language models can generate fluent responses to Islamic questions while introducing factual errors that are difficult to identify. This paper presents our system for \textsc{HalluScoring 2026} Task 2.1, \textit{Islamic Hallucination Detection and Find the Truth}. The task requires a unified two-step prediction: determining whether an Arabic answer generated by an LLM is hallucinated and selecting the verified answer from six closely related candidate options. We use the Islamic knowledge dataset provided by the shared task, which contains 600 question--answer instances, including 341 hallucinated and 259 non-hallucinated answers. Our system is based on the fine-tuned \texttt{google/gemma-4-12B-it} model and uses deterministic decoding during inference. The generated outputs are normalized to extract the hallucination label and the selected option. The system achieves a Macro-F1 score of 0.928 and a label accuracy of 0.935 for hallucination detection, together with an option accuracy of 0.895 for answer selection. These results yield a combined score of 0.912, demonstrating strong performance across both stages of the task. The lower option-selection accuracy indicates that distinguishing the verified answer from plausible alternatives remains more challenging than detecting hallucinated responses.
Focal cortical dysplasia (FCD) type II is an important cause of drug-resistant focal epilepsy, but its small size and subtle magnetic resonance imaging (MRI) appearance make automated segmentation difficult. Conventional multimodal networks concatenate T1-weighted (T1w) and fluid-attenuated inversion recovery (FLAIR) images and rely on subsequent layers to learn their relationships implicitly. We propose Compact Ratio-Interaction Learning U-Net (CRIL-U-Net), a 3D U-Net extension incorporating a Compact Ratio-Interaction Learning module that combines local spatial features, voxel-wise cross-modal mixing, and bidirectional ratio-inspired interactions. CRIL-U-Net was compared with a 3D U-Net and a U-Net with input-level self-attention using five-fold cross-validation on 85 FCD subjects and 25 healthy controls. Each architecture was trained independently with Dice-binary cross-entropy (Dice-BCE) and Focal Tversky-Focal (FTF) losses. With FTF, CRIL-U-Net achieved the highest mean Dice score ($0.196\pm0.262$), compared with $0.136\pm0.224$ for the U-Net and $0.135\pm0.214$ for the attention comparator. It localised 44 of 85 lesions, compared with 36 for the U-Net, and significantly outperformed both comparison architectures after false-discovery-rate correction. CRIL adds only 4,320 parameters to the backbone, indicating that compact cross-modal representation learning can improve FCD segmentation when combined with an imbalance-aware objective. The remaining 48.2\% miss rate nevertheless requires further methodological development and external validation.
Raman spectroscopy enables label-free molecular characterization across materials science, analytical chemistry, biomedicine, and industrial process monitoring. However, machine learning for high-dimensional spectroscopy remains constrained by limited labeled data and a mismatch between the physical organization of spectra and feature-agnostic models. Tabular foundation models such as TabPFN predict in context without parameter updates, but on wide spectra feature subsampling preserves channel coverage while losing joint visibility: in the five widest tasks, every channel was covered, yet only 4.4% of random channel pairs co-occurred in any forward pass. We present RamanPFN, a general-purpose spectral foundation framework for unified in-context inference through physics-guided spectral learning, which encodes the additive mixing of constituent spectra and the locality of vibrational bands. Global Compositional Unmixing decomposes distributed, multi-band mixture signatures into shared non-negative bases to couple covarying peaks across the full spectrum, while Local Vibrational Subspace Encoding extracts orthogonal modes within contiguous spectral neighborhoods to preserve peak morphology, intensity fluctuations, and peak shifts. Posterior marginalization over admissible unmixings unifies both scales into in-context prediction without updating the model. Across 74 public Raman datasets exceeding 330,000 spectra, RamanPFN outperformed 26 methods spanning chemometrics, spectral neural networks, deep tabular learners, and tabular foundation models on 129 regression targets, reducing backbone RMSE by 13.2%. We demonstrate across TabPFN-3, TabICLv2, and Kumo-Tabular that this advantage generalizes to both regression and classification tasks, reducing regression error by 12.3%-25.7% and multiclass log loss by 42.4%-63.5% to enable data-efficient quantitative analysis and sample identification.