Publication Radar.

Large reasoning models have recently demonstrated remarkable capabilities in solving complex tasks, where supervised fine-tuning (SFT) on long chain-of-thought data serves as a crucial foundation for eliciting and enhancing their reasoning abilities. Despite rapid progress in both improving and analyzing reasoning-oriented SFT, the field still lacks a systematic survey that consolidates its fast-growing literature. To fill this gap, we present a comprehensive review of recent advancements in reasoning SFT, examining the literature through the dual lenses of methodological design and analytical investigation. First, we review methodological improvements across the SFT pipeline and categorize them into data-centric approaches and algorithm-centric innovations. Second, we reorganize analytical studies along three dimensions: data characteristics, optimization dynamics, and mechanistic insights. Finally, by synthesizing current research focuses and remaining bottlenecks, we outline promising future directions for reasoning SFT. We hope this survey deepens the understanding of reasoning SFT and paves the way for advanced reasoning models.

Large Language Models (LLMs) have rapidly become important tools for science, education, medicine, and professional work. Despite their apparent simplicity from the user's perspective, effective and responsible use requires an understanding of how these systems work, what their limitations are, and how their outputs should be evaluated. I wrote this document as a basis for my lecture on LLM in the master in artificial intelligence in medicine at the university of Turin, Italy. This book provides a practical and accessible introduction to LLMs for scientists and professionals. It explains the fundamental ideas behind tokenisation, embeddings, training, inference, context windows, prompting, and instruction hierarchies, while deliberately avoiding mathematical detail. Particular attention is given to the limitations of LLM-generated answers, including hallucinations, uncertainty, reproducibility, bias, and the need for systematic verification and human oversight. The book also discusses privacy, governance, security, and the responsible handling of personal and sensitive information. More advanced topics include Retrieval-Augmented Generation (RAG), retrieval quality, chunking, embeddings, vector search, agents, tools, and workflows. A substantial practical exercise guides the reader through running a local LLM with Ollama and Open WebUI and using it to interact with a collection of scientific papers. The exercise emphasizes evidence retrieval, source verification, uncertainty, disagreement between studies, and the distinction between plausible language and scientifically supported conclusions. No programming knowledge is required.

Real-world forgery is inherently heterogeneous: it manifests not only as digital manipulation but also as physical spoofing, and is carried not only by images but also by videos. Despite substantial progress, a unified anti-forgery foundation model covering such a broad threat spectrum remains absent. In this paper, we present Venus-DefakerOne2, which substantially extends the forensic capability boundary along both the forgery-form and media-type dimensions. Trained on tens of millions of open-source and in-house samples, it incorporates two key designs. Tide-Style-Tuning alternates LoRA and full-parameter fine-tuning to balance computational efficiency, generalization, and hard-sample fitting while mitigating catastrophic forgetting. Entropy-based Test Adaptation determines the optimal prediction modality by analyzing prediction entropy. Venus-DefakerOne2 achieves new state-of-the-art performance across a comprehensive suite of benchmarks, including DeepFake faces (96.7%), AI-generated images (93.3%), tampered documents (87.6%), manipulated natural images (89.6%), physical spoofs (93.1%), and AI-generated videos (96.0%). It significantly outperforms both specialized expert models and its predecessor, DefakerOne, establishing itself as a powerful and versatile foundation model for binary-decision forensic tasks.

Large language models (LLMs) have emerged as potential tools for numerical optimization. Despite promising results, their intrinsic optimization capabilities and limitations remain poorly understood, particularly in high dimensions. This paper investigates the ability of LLMs to exploit mathematical problem structures and generate feasible, near-optimal solutions. We develop an evaluation framework that distinguishes solution quality from execution failures, including numerical inaccuracies, constraint violations, and output-format errors. Using representative problems from convex optimization, fractional programming, and constrained projection, we assess LLM performance against deterministic reference solvers across different dimensions. Our results reveal that LLMs can capture useful optimization structures and produce near-optimal solutions, even when their outputs fail to satisfy numerical or formatting requirements. However, reliability deteriorates with increasing problem dimension and depends strongly on the underlying optimization structure. To further investigate the limitations, we deploy an iterative refinement mechanism that provides targeted feedback on feasibility and solution quality. The results demonstrate that iterative refinement substantially improves reliability and extends the range of problem dimensions over which LLMs produce valid solutions. Nevertheless, persistent numerical and structural errors limit their applicability as standalone optimizers. These findings provide insights into the capacity and limits of LLM-based optimization and motivate hybrid approaches that combine LLM reasoning with deterministic numerical verification.

Earthquake sequences are stochastic on two separated timescales: the seismic waves of any individual earthquake propagate over seconds to minutes, while the earthquakes themselves arrive as a sparse, self-exciting occurrence process, with inter-event intervals from seconds to months. Respecting both scales in one training objective is the open challenge. Foundation models (FMs) for seismic waveforms instead import their learning targets from audio and vision, where every target is computed from a single fixed-length window. Such a window holds one realization of the wave propagation (an earthquake trace), never the occurrence process that decides when earthquakes happen. We develop nine model designs sharing one Mamba encoder-decoder, trained on a fixed corpus of 3M windows, varying the self-supervised objective, quantization bottleneck and mixer; the two key comparisons are architecture-matched, varying the objective alone. Each checkpoint is evaluated frozen on three downstream tasks. Varying the target redistributes downstream skill rather than raising it, and no checkpoint wins everywhere. Adding masking pressure to our strongest classifier re-specializes it rather than making it a generalist: it gains 6.9 points of few-shot P-wave picking precision and loses 9.6 points of few-shot classification accuracy. By the standard imported from language and vision, these models are feature extractors rather than FMs. Controls on capacity, architecture, and domain proximity do not explain why skill moves rather than rises. Against task-specific baselines the advantage is regime-dependent: our checkpoints beat them few-shot on classification and ground-motion prediction, and trail them on full data. We argue the binding constraint is the target, and that settling the question requires a learning target defined across events, the timescale that makes seismicity stochastic in the first place.

Medical chart abstraction transforms unstructured clinical narratives into structured variables for research and quality monitoring. We evaluated a prompting-based large language model (LLM) framework for extracting eight clinical variables from 2,416 neurointerventional acute stroke intervention reports across three hospitals. Twenty-two instruction-tuned open-source LLMs were evaluated across architectures, model sizes, and biomedical adaptations using Quick Response and Chain-of-Thought (CoT) prompting. Non-expert staff annotations served as the cross-model reference standard. Evaluation was restricted to report-variable pairs with a valid LLaMA3.3-70B CoT prediction and a corresponding reference label. Performance measures included extraction accuracy, inference latency, and agreement with expert adjudication. Twenty discordant cases per variable underwent review by two senior physicians and four medical students. LLaMA3.3-70B achieved the highest overall accuracy (94.8%). CoT improved performance on inference-dependent variables such as site of occlusion, while Quick Response was optimal for directly stated procedural fields. Expert adjudication showed that AI predictions were more closely aligned with expert interpretations than non-expert annotations for most variables, with intracranial angioplasty remaining a challenge. Locally deployed LLMs can support scalable clinical information extraction and retrospective stroke registry curation, with expert oversight for ambiguous cases.

Multimodal large language models (MLLMs) are vulnerable to visual jailbreak attacks that bypass safety safeguards. Existing methods typically optimize isolated adversarial examples, limiting attack diversity, transferability, and generalization. We propose Jailbreak Connectivity (JC), a framework based on the observation that effective adversarial inputs can be connected through low-loss regions in image space. JC optimizes continuous adversarial paths: path sampling yields diverse jailbreak images, lightweight surrogate classifiers guide cross-model transfer, and a distribution-level objective enables universal attacks across harmful queries. On MM-SafetyBench, JC achieves average attack success rates of 84.01%, 85.77%, and 97.95% on Qwen3-VL-8B-Instruct, InternVL3.5-8B, and Gemma-3-12B-IT, respectively, exceeding the strongest baselines by up to 68.36 percentage points. In black-box evaluations, JC improves attack success rates over FigStep by 24.69 and 27.34 percentage points on GPT-5.6 and Gemini 3.1 Pro, respectively. These findings reveal structured vulnerabilities in MLLM safety alignment and inform more comprehensive security evaluation.

Fixed-wing unmanned aerial vehicles experience operating-condition transitions, asynchronous sensor sampling, and control-system compensation that can obscure fault signatures. Existing temporal detection methods have limited capacity to represent coupled faults across subsystems, track their evolution, and constrain language-model diagnoses with quantitative evidence. This paper proposes TGTV, an interpretable fault detection and traceability framework that integrates spatiotemporal graph learning with a large language model. A temporal convolutional network and a graph attention network with weighted fused adjacency extract multiscale temporal patterns and spatial dependencies among sensors. A hybrid temporal Transformer and variational autoencoder then evaluates window-level anomalies using prediction errors and reconstruction probabilities. Continuous-time Shapley attribution identifies critical sensors and high-contribution intervals while preserving the temporal continuity of fault evolution. A structured evidence package combines detection results, graph relationships, temporal attribution, sensor attribution, and operating context. An evidence-constrained diagnosis protocol uses this package to generate flight-level conclusions with traceable evidence identifiers. Cross-flight holdout experiments on the public ALFA fixed-wing flight dataset yield a window-level area under the precision-recall curve of 0.976, accuracy of 0.973, recall of 0.902, and F1 score of 0.920. Flight-level fault diagnosis achieves recall of 0.943 and an F1 score of 0.944. These results demonstrate accurate anomaly detection and quantitatively supported fault-tracing cues, providing a basis for interpretable post-flight fault review and analysis.

Large Language Models (LLMs) treat all context tokens uniformly, lacking explicit separation between executable instructions and passive data. This leads to prompt injection vulnerabilities, ambiguous instruction binding, and unreliable execution control. We present a post-training methodology using LLMON (LLM Object Notation), a framework that enables separating instructions from data and requires explicit identifier-based execution. We construct a large-scale structured training corpus of 4.6B tokens spanning general, RAG (Retrieval-Augmented Generation), and tool-calling data, augmented with distractor-infused variants in which competing instructions and irrelevant data segments are present in context but not referenced. Post-training two Granite-4.1 base models (3B and 8B) through supervised fine-tuning followed by reinforcement learning, we show that LLMON-structured supervision reduces attack success rates on indirect prompt-injection benchmarks from 53-100% to 0.02-5.6% while preserving general capabilities. We further introduce attention-based interpretability metrics that diagnose failure modes at individual transformer layers for structured inputs, revealing how our approach reshapes internal attention patterns. Our results demonstrate that explicit execution semantics can be taught through standard post-training without architectural modification.

Large language models hold promise for delivering safe, scientifically grounded, and personalized exercise guidance at scale, yet the sports-and-health domain still lacks curated training corpora and rigorous evaluation benchmarks. Keepace.ai is a domain-specialized LLM built on Qwen3.5-27B through post-training, currently optimized for Chinese-language exercise guidance across three tasks: workout plan generation, sports-and-health knowledge QA, and personal sports data analysis. In internal rubric-based evaluations designed with domain experts, Keepace.ai outperforms its base model and larger general-purpose baselines, with the most pronounced improvements in exercise safety, boundary compliance, structured output completeness, and scientific accuracy. This report summarizes the current design, data pipeline, safety-first evaluation framework, and evaluation results of Keepace.ai.

Large language models (LLMs) are advancing in multimodal understanding and reasoning, accelerating medical model development. Traditional Chinese medicine (TCM) uses evidence from the four examinations to determine disease location and nature, differentiate syndromes and guide treatment. This diagnostic process constitutes a complete chain of reasoning. We present Fu-TCM, a multimodal TCM LLM trained on data retained after cleaning and filtering an initial collection of approximately 7.12 million examples derived from classical texts, textbooks, medical images, clinical cases and publicly available datasets. The model was further optimized using Bianzheng-Grounded Policy Optimization (BGPO), which explicitly targets syndrome differentiation through rewards for response format, intermediate TCM reasoning fields and final syndrome predictions. Fu-TCM-27B led all six benchmarks spanning TCM reasoning, text and vision, achieving a macro-average accuracy of 81.01% and surpassing the strongest reference model with complete results across all six benchmarks by 7.13 percentage points. Its largest margin was 11.90 points on Fu-bianzheng. Case-level analyses further showed a positive association between intermediate-field accuracy and final syndrome-selection accuracy in most evaluated models. These findings suggest that TCM syndrome differentiation may be computationally learnable and that structured reasoning supervision can improve performance across TCM tasks.

Graph Foundation Models (GFMs) have recently emerged as a promising paradigm for learning transferable representations across different graphs and domains. However, existing GFMs face notable limitations, including limited pre-training scale or reliance on aligning with textual information in Large Language Models (LLMs). To overcome these limitations, recent studies have drawn inspiration from prior-data fitted networks (PFNs), which learn from large synthetic datasets and support multi-domain prediction backed by causal learning. However, existing graph PFNs encounter severe difficulties in modeling causal relationships within and across node and edge sets, generating diverse graph structures, and designing suitable architectures that can learn from complex synthetic datasets. To address these challenges, we propose **GSCM-PFN** (Graph SCM-based Prior-Data Fitted Network), a graph PFN pre-trained on synthetic data generated from an SCM-based prior that jointly models nodes and graph topology. Specifically, we (1) design a joint feature-edge-label structural causal model (SCM) to guide synthetic graph generation and capture causal relationships within and across node and edge sets; (2) introduce prior-guided node-edge dependency modeling to instantiate diverse graph generation strategies, and prove that our SCM-based generation prior is more expressive than existing priors; and (3) propose a feature-permutation-invariant dual-path graph network with theoretical guarantees for learning from diverse synthetic graphs. Experiments on 23 real-world benchmarks show that GSCM-PFN outperforms both conventional GFMs and PFN-based GFMs, without utilizing real-world graph data for pre-training. Code: https://anonymous.4open.science/r/GSCM-PFN-A3A3/

Visual foundation models learn rich pathological morphology but lack explicit concept-level supervision. Vision--language pretraining can provide this supervision but requires costly image--text pairs. Here, we introduce GloSAGE, a semantic self-distillation framework that uses lesion profiles predicted by a pretrained visual teacher to construct structured pathology descriptions. Reliability-guided selection identifies image--text anchors, while Knowledge-Structured Pathology Text Synthesis and Evidence-Guided Negative-pair Construction account for partially observed lesions. Image--text alignment trains projection layers while retaining the frozen visual encoder; an intermediate Semantic Transfer Layer supplies features for downstream tasks. Developed on GloMillion and evaluated across seven glomerulus- and case-level cohorts, GloSAGE supported zero-shot lesion recognition, image--text retrieval, hard-negative discrimination and concept-intervention analyses. It outperformed its source visual model in most lesion-recognition tasks and improved lesion grading, while supporting competitive case-level disease classification, injury grading and baseline renal function estimation. Ablations supported the combined contributions of knowledge-structured text and evidence-guided negatives, and the intermediate transfer layer improved mean lesion-recognition ROC-AUC relative to the complete projection output. A frozen language model conditioned on GloSAGE features generated glomerular descriptions. These findings suggest that teacher-derived lesion profiles can provide annotation-efficient semantic supervision for extending a glomerular visual foundation model.

Large language models (LLMs) are often conditioned into specific roles such as tutors, interviewers, or agents acting for an organisation, through system prompts that specify how they should behave. The extent to which they can actually adhere to these role specifications under multi-turn interactions with realistic and potentially problematic users is understudied. To address this, we propose user behaviour fuzzing, a guided search over plausible user behaviours that can lead a model to violate its role specification, and introduce RoleFUZZER, a framework that adapts coverage-guided fuzzing from software testing to operationalise the search. RoleFUZZER constructs scenarios guided by a user-behaviour coverage grid that crosses a grounded taxonomy of user interaction types and the role specification of the model under test i.e. what is under pressure. With RoleFUZZER, we evaluate LLMs across five role specifications, including tutor, financial advisor, AI interviewer, antagonistic roleplay character, and web agent under organisational policies. Across simulated and validated conversations in 15 campaigns, the framework reaches >= 89% coverage of the 64-cell behaviour grid for every model and role tested, and surfaces specification-grounded failure patterns. At an equal budget of conversations, removing the RoleFUZZER grid and asking an LLM to generate specification targeting users free-form lowers coverage to 39-69% with fewer kinds of violations. The same specification can also fail in opposite directions. For e.g., on the antagonist, open-weight target models override the player's safety signals, while Claude Haiku 4.5 breaks character instead. RoleFUZZER offers a tractable way to uncover vulnerabilities in LLMs across role specifications by testing diverse user behaviours that are likely to violate them.

Medical workflows often require selecting among predefined options. We compared 13 decision models with 15 generative large language model configurations, including different reasoning settings, using 1,273 four-option MedQA questions under text-only conditions. Accuracy ranged from 67.71% to 89.08% for decision models and from 78.32% to 96.86% for generative configurations. Among the tested configurations, Mercury Decide had the lowest mean final-attempt cost (USD 0.0069 per 1,000 questions), while Clef Omni had the shortest median request duration (0.291 seconds). Decision-model probabilities distinguished correct from incorrect answers, but error rates among answers assigned probabilities of at least 0.90 ranged from 1.71% to 12.77%. Answer agreement across three runs ranged from 92.30% to 100%, while probability repeatability varied. Decision models offered resource advantages alongside lower peak accuracy; selecting models for medical workflows requires considering probability quality and repeatability together with accuracy and the resources needed to detect and correct errors.

Spectroscopic diagnostics provide rich information for characterizing plasma states in tokamak fusion experiments, while existing learning-based approaches are often designed for individual tasks and rely heavily on labeled data. In this work, we propose SpectralTF-MAE, a self-supervised spectral foundation model for tokamak plasma diagnostics. The model is pretrained on paired charge-exchange recombination spectroscopy (CXRS) and X-ray crystal spectroscopy (XCS) measurements through masked spectral reconstruction. To better capture diagnostically important spectral structures, we introduce a peak-aware masking strategy together with a peak-enhanced reconstruction objective, encouraging the model to focus on characteristic spectral regions. A shared Transformer encoder is employed to learn transferable representations across different diagnostic modalities. The pretrained model is further adapted to several downstream tasks, including plasma-parameter prediction, cross-diagnostic spectral generation, and spectral peak analysis. Experiments on EAST diagnostic data demonstrate that the proposed framework achieves competitive performance across different tasks and consistently benefits from self-supervised spectral pretraining. These results indicate the potential of SpectralTF-MAE as a general and transferable representation learning framework for plasma spectroscopic diagnostics.

The scale of modern foundation models enables broad task generalization. In medical imaging, where segmentation demands vary widely across clinical contexts, such flexibility is invaluable. But current models, such as SAM trade adaptability for scale, locking their utility behind high-end infrastructure and leaving behind the low-resource environments most in need of diagnostic support. To bridge this gap, we introduce Foundational-NCA, a lightweight foundation model capable of adapting to a wide range of medical segmentation tasks without retraining. Built on Neural Cellular Automata (NCA), Foundational-NCA incorporates gradient-based learning with genetic optimization over a latent task space, using a compact task vector to guide predictions. This enables dynamic adaptation via a genetic prompt mechanism, allowing the model to generalize across tasks while keeping the parameter count magnitudes lower at just 211 thousand, compared to SAM-based models with 38.9 to 636 million parameters. We evaluate Foundational-NCA on 15 trained and 6 entirely unseen segmentation tasks from the TotalSegmentator dataset, achieving an average Dice score of 84.5% on known tasks and 60.4% on unseen ones. These results demonstrate that task generalization is possible in ultra-compact models, laying the groundwork for resource-efficient, deployable foundation models in clinical environments where infrastructure is limited. By reducing model size by over 180 times, Foundational-NCA takes a step toward a more equitable future for medical AI, one where powerful tools are within reach of every healthcare setting.

Learning-enabled cyber-physical systems (CPS), including autonomous vehicles and UAVs, increasingly incorporate foundation models, yet fragmented assurance practices provide insufficient support for system-level guarantees. We argue that restoring engineering control across stochastic neural, symbolic, and physical boundaries requires a shift from model-centric evaluation to system-level assurance. We introduce the Foundation Model Assurance Stack, connecting three directions: (i) neuro-symbolic architectures that constrain foundation models through explicit symbolic world models, reasoning layers, and interfaces; (ii) semantically grounded, conformance-based testing that combines formal specifications and symbolic world models with foundation-model-assisted specification mining, scenario generation, and test synthesis; and (iii) compositional verification and runtime assurance that target formally tractable components and interfaces through modular contracts, equivalence checking, verified synthesis, uncertainty-aware reasoning, and runtime monitoring and enforcement. Symbolic abstractions establish assurance boundaries between stochastic learning and formal reasoning, connecting testing, verification, and runtime evidence across the system lifecycle. Foundation models serve as constrained assistants whose outputs are independently checked wherever correctness can be formally specified. We identify research challenges in specification engineering, abstraction fidelity, evidence composition, and lifecycle certification, and outline a staged agenda toward trustworthy, deployable, and certifiable learning-enabled CPS.

Tabular foundation models use in-context learning to estimate conditional expected outcomes and average treatment effects for specific groups of individuals. Using these estimates to guide scientific decisions requires reliable uncertainty quantification. The challenge is to account simultaneously for two sources of uncertainty: limited-context bias and context variation arising from the small labeled sample supplied at inference. Predictive spread does not directly quantify both sources of uncertainty, and we observe that the ordinary bootstrap may substantially overestimate context variation, whereas subsampling may underestimate it. We introduce TabCUQ, a perturbation-based framework that injects noise into observed outcomes and applies the foundation model to the perturbed contexts. TabCUQ estimates context variation using wild-bootstrap dispersion and accounts for limited-context bias using the maximum deviation between perturbed and original estimates. For a kernel-regression analogue whose estimation error admits an exact decomposition into these two components, we show that our framework provides asymptotically valid pointwise coverage. In simulations, TabCUQ substantially improves coverage, with average interval widths typically about twice those of an oracle bias-aware benchmark. Across seven real-world tasks, it also improves coverage of reference functions learned from much larger samples.

As large language models (LLMs) become more widespread in Africa, their safety in low-resource languages remains largely untested since most red-teaming work focuses on English. This study evaluates six state-of-the-art LLMs using 400 explicit adversarial prompts in Yoruba and Nigerian Pidgin, categorised into four harm categories. The responses were scored by native-speaker annotators and reliability was assessed using Cohen's $\kappa$. While models that rejected up to 69% of English prompts rejected as few as 0% in Yoruba, this was not because they were safer, but because they failed to understand the prompts. Attack success rates increased for most models in the Nigerian languages, with GPT-4o Mini demonstrating the most significant cross-lingual safety discrepancy (+0.23 in Yoruba). When models did engage with adversarial prompts, many of the responses were harmful, with rates of harmful content reaching 0.63 in English and 0.50 in Nigerian Pidgin. The most difficult harm category to detect was cultural stereotypes and discrimination (ASR = 0.44 in Nigerian Pidgin). Comprehension in Yoruba was low across most models, and this directly suppressed apparent attack success without indicating actual safety. Gemini 2.0 Flash was the most consistent model across languages. Despite its explicit training in African languages, Aya 101 had the highest attack success rate in English (0.60) and one of the lowest Yoruba comprehension rates (0.28). These results demonstrate that English-centric safety alignment does not transfer to Nigerian languages, highlighting the urgent need for language-specific safety evaluations.

Large language models (LLMs) draw on parametric knowledge and retrieval-augmented generation (RAG) to solve knowledge-intensive tasks. Yet problem solving in unfamiliar domains calls for tightly coupled knowledge acquisition and reasoning, beyond factual retrieval and aggregation or learning solely from a preassembled context. We study self-learning -- autonomously retrieving, studying, and applying knowledge from long-form reference materials on the fly. To investigate this capability, we formulate a textbook-based mathematical problem-solving task. LLMs must navigate textbook exposition, learn and apply relevant principles, and construct proofs for exercises, with reference solutions withheld. We introduce a BookNavigator interface to support this human-inspired study process and incentivize self-learning through reinforcement learning (RL) with fine-grained rewards. Experiments show that our method improves performance across in-domain and out-of-domain textbooks and subjects, raising Intern-S2-Preview-35B's overall proof accuracy from 51.5% to 79.5% and outperforming RL-trained textbook-free and fixed-retrieval baselines. These gains extend to complex long-context benchmarks involving different task types, with improvements over the base model of 1.7 and 2.3 percentage points on BrowseComp-Plus and CL-bench, respectively. Together, these findings suggest that RL can foster a domain-agnostic self-learning strategy, highlighting the potential of self-learning from long-form corpora to strengthen broader problem-solving capabilities on complex, long-context tasks.

Over-refusal is a critical side effect of safety alignment in large language models (LLMs), causing harmless but superficially risky queries to be mistakenly rejected. Despite the effectiveness of existing mitigation methods, why safety-aligned LLMs may still reject harmless queries under correct safety supervision remains insufficiently understood. To address this gap, we analyze over-refusal through the lens of representation and optimization. We make two observations: 1) the representations of harmful and over-refusal samples are visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers; and 2) harmful-sample gradients exhibit high intra-class similarity and concentrate in a low-rank subspace. Moreover, they are more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts. This imbalanced gradient coupling can bias over-refusal samples toward refusal, while the reduced apparent representational separation in deeper layers may make it more difficult for the model to maintain a clear boundary between harmful and over-refusal samples. To this end, we propose GOAL (Gradient Orthogonalization with contrAstive Learning), a safety-alignment framework for mitigating over-refusal. Specifically, GOAL consists of two components: 1) to prevent imbalanced gradient coupling from biasing the optimization of over-refusal samples toward refusal, we orthogonalize over-refusal gradients against the low-rank refusal subspace extracted from harmful gradients; and 2) to explicitly enhance the separation between harmful and over-refusal representations, we introduce a contrastive learning objective. Extensive experiments across multiple models and benchmarks demonstrate that GOAL effectively mitigates over-refusal while preserving safety performance and general capabilities.

Novel view synthesis (NVS) in degraded scenes aims to reconstruct 3D scenes from multi-view observations affected by various degradations, and to generate clear and photorealistic images at unseen viewpoints. Despite recent advances, existing methods often rely on degradation-specific models, limiting their generalization across diverse extreme environments. Generative models provide rich visual priors, but their outputs often suffer from pixel misalignment and color deviations, which can propagate into the reconstructed 3D scene and compromise reconstruction quality. Furthermore, the high computational and memory costs of 3D reconstruction hinder the practical application of NVS in extreme environments. To this end, we propose MLLM-GS, a multimodal large language model (MLLM) prior-guided framework for fast NVS in extreme environments. MLLM-GS consists of two branches: MLLM prior-guided view enhancement and fast 3D scene reconstruction. The former leverages the rich visual priors of a frozen MLLM to enhance views under different severe degradations and introduces a lightweight correction network to reduce pixel shifts and color distortions in MLLM outputs. With supervision from only a few paired clean views, this branch produces reliable enhanced views for all degraded observations, providing pseudo-GT views for subsequent 3D reconstruction. The latter improves 3DGS efficiency by optimizing rasterization, Gaussian memory layout, and parameter updates, reducing redundant computation, memory-access overhead, and optimization costs. Extensive experiments on multiple degraded scenes demonstrate that MLLM-GS effectively handles various severe degradations and achieves high-quality NVS with significantly lower reconstruction time and memory consumption. Project Page: https://wzlwlm13.github.io/MLLM-GS

An observer sees only part of the surrounding world, yet must keep track of what lies beyond the current view. Spatial perception therefore involves more than recognizing individual images: it calls for a persistent spatial state that is updated under self-motion and used to guide action. Do visual foundation models provide representations that support such a state? We study this question through four levels of spatial capability---spatial structure, object persistence, predictive spatial update, and actionability---using frozen-feature probes on ten representations under streaming egocentric observations of static indoor environments. The resulting capability profiles vary across tasks, network layers, and prefix lengths: strong readout of the visible scene does not ensure accurate object persistence or future-feature retrieval. Matched controls separate readout accuracy from reliance on visual history and motion. Native future prediction retains recoverable spatial structure, but a spatial reader trained on real features does not transfer to predicted features unchanged. Finally, goal-dependent candidate selection does not establish reliable continuous action utility: predicted actions respond to goal changes, yet supplying the correct goal does not reliably improve progress. Across the four levels, object persistence and actionability build on camera-pose readout, whereas feature prediction builds on visible-scene readout. Together, these findings motivate evaluating spatial representations not only by what can be decoded from them, but by whether that information is maintained beyond observation, updated predictively, and used for goal-directed action.

Foundation models are typically trained in stages: pretraining, midtraining, supervised fine-tuning (SFT), and reinforcement learning (RL), with changes in both data and learning algorithms. We show that this algorithmic separation is not fundamental. By deriving suitable RL objectives as maximum-likelihood fitting to model-generated target distributions, we place all of these stages under a single likelihood-based training algorithm, unified training. This perspective turns the foundation-model training lifecycle into a problem of choosing which data to train on, in what proportions, and when, with conventional staged training as a special case. We instantiate this framework in a unified training system that mixes static data and RL generations within a single trainer. Across controlled and real-world experiments, unified training improves the accuracy-compute tradeoff and task coverage while preserving broader capabilities, and provides a stronger starting point for continual learning. This work lays the groundwork for a unified science of foundation-model training, connecting mathematical principles, training infrastructure, and the development of model capabilities.

Foundation models trained on longitudinal electronic health records (EHRs) can predict clinical diagnoses years in the future, making them a potential tool for disease screening. We developed TRACER (Temporal Risk Assessment with Clinical EHR Representations), a decoder-based foundation model pretrained on data from over 4.5 million patients from four health systems, and evaluated it on 16 incident diseases at 1-, 2-, and 5-year horizons in more than 2 million patients from two health systems excluded from model development. In contrast to typical evaluations that focus on global metrics (e.g., AUROC), we measured performance at the low false-positive rates that are realistic for screening. We found that TRACER retained its discrimination in external health systems, and at 95% specificity, the patients it flagged were up to eight times more likely to be diagnosed within a year than unselected patients. Evaluating at this operating point also revealed what global metrics alone did not: detection and false-alert burden varied across diseases and health systems, even when discrimination appeared stable. Our results illustrate the potential of foundation models for screening and underscore the need for evaluation in health systems excluded from development to assess disease detection and false-alert burden at levels appropriate for intended clinical use.}

Next point-of-interest (POI) recommendation aims to predict a user's next destination from spatio-temporal check-in data, and serves as the foundation for many location-based services. Recent advances incorporate large language models (LLMs) into this task by textualizing user trajectories to exploit their semantic reasoning ability. However, directly generating POI identifiers with such models is unreliable: POIs are sparse, localized, and long-tailed entities with weak natural-language semantics, which creates a semantic-to-entity gap between high-level mobility intent and concrete POI prediction. To bridge this gap, we propose ANCHOR, a category-anchored framework for semantic-to-entity grounding in next POI recommendation. Specifically, ANCHOR first estimates category-level mobility intent in a compact and stable semantic space, rather than predicting concrete POI entities directly. The inferred intent is then grounded into a candidate set by integrating three complementary sources of mobility evidence: category anchors, historical visit memory, and user preferences. Finally, a history-aware reranker selects the target POI by matching each candidate with the user's historical mobility patterns and current context. Experiments on three real-world check-in datasets show that ANCHOR consistently outperforms representative traditional, structure-enhanced, and LLM-based baselines on the reported metrics. Further analyses validate the effectiveness of semantic-to-entity grounding and show favorable trade-offs among candidate coverage, ranking quality, and inference efficiency.

Research on medical item difficulty has largely focused on prediction rather than on identifying interpretable patterns associated with empirical difficulty. This study used empirical difficulty labels to guide large language models (LLMs) in inducing difficulty rubrics and examined how rubric guidance affected LLM difficulty judgments. Using training items and their empirical difficulty estimates from the United States Medical Licensing Examination (USMLE) and the Physician Assistant National Certifying Examination (PANCE), Claude Opus 5 induced separate rubrics for absolute and pairwise difficulty judgment. None of the induced rubrics treated surface features, such as stem length, option count, or images, as major sources of difficulty. Several patterns emerged across both examinations, including cue convergence, misleading cues, and option competition, although their specific forms differed. Retrieval of knowledge that could not be derived from the stem was particularly prominent in the PANCE rubrics, whereas the USMLE rubrics placed greater emphasis on integrating multiple clinical cues. We compared commercial and open-weight LLMs on held-out items under unguided and rubric-guided conditions. Rubric guidance changed how models used the difficulty scale but did not consistently improve performance. In absolute judgments, guidance encouraged broader use of the scale and generally improved coarse-grained agreement, without consistently improving rank ordering. In pairwise comparisons, its effects varied across models and examinations, and judgments became more sensitive to presentation order. These findings suggest that the value of label-guided rubric induction may lie more in generating interpretable, testable hypotheses from empirical difficulty data than in reliably improving difficulty judgments.

Existing gut microbiome foundation models typically use flat taxon representations, limiting their ability to exploit known taxonomic hierarchies. We introduce TaxoMFM, a microbiome foundation model that integrates taxonomic structure into self-supervised learning. TaxoMFM jointly learns abundance patterns across taxonomic ranks and relative proportions within groups through taxonomy-aware encoding and hierarchical compositional reconstruction. We further construct TaxoGut-HPP, comprising 100,806 gut microbiome profiles from 231 source cohorts and covering 40 classification and 8 regression tasks. TaxoMFM achieves competitive performance across disease, gut health, host trait, and liver ultrasound phenotype prediction. In cross-cohort zero-shot transfer, where target cohorts are excluded from both pretraining and downstream adaptation, TaxoMFM achieves AUROCs of 0.838, 0.736, and 0.722 for colorectal cancer, type 2 diabetes, and inflammatory bowel disease, respectively. Further analyses show that TaxoMFM recovers community changes across taxonomic scales, while preserving community information and producing consistent representations under database updates and species-vocabulary shifts. Our code is available at https://anonymous.4open.science/r/TaxoMFM-F16D/.

We introduce SocialArena, a benchmark for systematically defining and evaluating the social intelligence of large language models (LLMs). Although LLMs are increasingly deployed in socially interactive applications such as mental health counseling, companionship chatbots, and educational tutoring, existing evaluations remain fragmented focusing on isolated capabilities or single scenarios and evaluation reliability is largely unverified. This work makes three main contributions to address these limitations. (1) We first propose a unified three-tier framework that defines social intelligence along three capability dimensions: role-playing ability, mental state reasoning, and behavioral reasoning. To operationalize this definition, we design a role-playing dialogue paradigm with explicit character thinking for evaluation, enabling simultaneous assessment of these capabilities in a manner that mirrors realistic human interactions. (2) To ensure diversity across social scenarios and thereby prevent evaluation conclusions from being confined to specific scenarios, we introduce the U-TRIMF framework, which guides us in synthesizing social contexts with broad coverage. (3) To obtain reliable evaluation, we invest $132.4K to recruit and train human annotators and adopt a human-annotated arena-style protocol, where annotators compare model responses and provide preference labels along with natural-language critiques. We hope that SocialArena can provide valuable insights and guidance for future LLM improvements.

Tabular foundation models are pretrained networks that predict risk from a table without fitting a model. Whether their risks are correct (calibrated) has been judged on real data, where true risk is unknown. In a pre-registered simulation with known true risks, we compared TabPFN-3.5, TabICLv2 and Mitra-v2 with Firth logistic regression with intercept correction (FLIC) in 58 scenarios and 1,000 random designs, then on four clinical data sets. Foundation models were not a shortcut to calibration. Their calibration slope matched FLIC's within the pre-registered margin in 36 or 37 of 48 core scenarios and was better only at n = 200 with a 3% event rate. Even 3,000 patients were often not enough: with FLIC, TabPFN-3.5 or Mitra-v2, 63%-83% of samples gave risks close to the truth in average, intercept and slope at event rates of 10% and 30%, and 35%-50% at 3%. TabICLv2 did worse (8%-39% and 1%-3%). Across the random designs, foundation models had a lower true calibration error than FLIC beyond the margin in 25%-38% and a higher one in none, though small losses were predicted at n = 3000. On real data, no model resisted a change in event rate between populations; recalibration on 500 local patients largely repaired it.

Industrial systems continuously generate time series that vary in sampling frequency, scale, dynamics, and operating regime. Consequently a single time-series foundation model may not be able to represent all regimes effectively and retraining a large model for every new dataset or regime shift is computationally expensive. Modular foundation models address this issue by reusing specialised encoders trained on diverse datasets and by allowing the expansion of capabilities through the addition of new experts. This approach requires compact experts to minimize storage cost and a proper selection mechanism of experts for unseen data. We propose a modular foundation model framework that combines end-to-end Bayesian encoder training with uncertainty based routing. Using Singular Bayesian Neural Networks (SBNNs), we build a collection of compact Bayesian encoders whose epistemic uncertainty guides label-free selection. At the selected ranks, our Bayesian collection uses 63.8\% less storage than its deterministic counterpart. We evaluate this framework on time-series data through distribution shift, encoder selection, and probabilistic forecasting. Across seven incoming datasets, uncertainty-based routing is competitive with distance-based routing in terms of forecast MSE, while reducing average ECE from 0.086 to 0.072 and average NLL from 0.907 to 0.836. It also increases PICP90 from 82.28\% to 89.03\%. In a separate model-family comparison, we evaluate the SBNN pipeline against full-rank Bayes by Backprop, deep ensembles, and MC Dropout. The SBNN pipeline achieves an average MSE of 0.390, comparable with 0.396 for deep ensembles, 0.389 for MC dropout, and 0.443 for the full-rank Bayesian model. It uses 63.9\% fewer parameters than MC dropout, 81.9\% fewer than the full-rank model, and 92.8\% fewer than the deep ensemble.

Low-bit weight--activation quantization reduces the memory and computation needed to deploy large language models, but activation rounding can still degrade accuracy after calibration. Reconstruction from rounded inputs cannot distinguish activations within the same quantization bin. We propose ResCom, a low-rank correction that uses the activation rounding residual to compensate for the resulting layer-output error. An output-aware calibration procedure selects residual directions by their contribution to output reconstruction and computes the correction in closed form, without gradient training or updates to the calibrated weights. On Llama-3-8B with 4-bit weights and activations, rank-32 attention corrections reduce WikiText-2 perplexity from 7.708 to 7.147, improve MMLU by 2.89 percentage points, and raise five-shot GSM8K accuracy from 23.96 to 36.16 percent. Evaluations on models from 1B to 70B parameters and on held-out C4 test the method beyond this setting. The evaluated FP32 implementation adds 2.25 percent to 4K prefill time; a separate compact-storage implementation retains a 27.7 percent peak-memory reduction relative to FP16.

Video Large Language Models (VLMs) have demonstrated strong capabilities in video understanding and reasoning, but processing long visual sequences incurs substantial computational, latency, and memory overhead. This survey examines training-free inference acceleration, focusing on reducing inference costs while preserving task-relevant spatiotemporal evidence. We organize existing methods by their intervention stages in the inference pipeline, covering input, visual encoding, projector, prefill, and decoding. Post-training quantization (PTQ) is discussed separately as a cross-stage extension with varying calibration requirements. We analyze how intervention position affects evidence preservation and computational savings, and why local efficiency improvements may not translate into end-to-end acceleration. We complement this analysis with quantitative evidence from prior studies and our own experiments under a shared evaluation setting to assess accuracy-latency trade-offs.

Adapting foundation models to a new task within or beyond the training domain is the de facto standard for building specialized machine learning models. Recently, Visual Prompting (VP) has emerged as a promising technique for efficient knowledge transfer. As new foundation model families~(like Mamba) get introduced and VP pipelines such as AutoVP reach greater maturity, we find a growing need for a systematic evaluation of current approaches. In this work, we assess the performance of the latest models, comparing them to earlier architectures and alternative fine-tuning methods, to better understand the progress, challenges and opportunities in the field of efficient fine-tuning under resource limitations. Towards this goal, this paper provides a concise empirical overview of the interactions among foundation model families (Attention-, Convolution-, and Mamba-based) and transfer paradigms: VP, Linear Probing (LP), Full Fine-tuning (FFT) and Low Rank Adaptation (LoRA). Our work builds up on previous findings by broadening the selection of evaluated models and techniques while tuning hyperparameters. In the interest of delivering practical guidelines for the user, we also explore application of prevalent regularization techniques to boost performance in the context of VP.

Large language models (LLMs) require substantial memory, motivating low-bit weight quantization for efficient deployment. Existing vector quantization (VQ) methods often improve accuracy by increasing the effective coding dimension, which can complicate independent lookup and local codeword updates. We propose CoS-VQ (Coordinated Shared-Scale Vector Quantization), which preserves low-dimensional codewords while jointly optimizing hard indices and their shared scale across positions. The key challenge is scale-induced coupling: the optimal scale depends on the entire codeword combination and, in turn, changes candidate rankings across positions, so pointwise updates can stall even after scale re-optimization. CoS-VQ collects candidates at multiple probe scales and constructs a tangent separable upper bound for the residual-aware quadratic objective. By reducing candidate comparisons to lower envelopes of lines over positive scales, it identifies finitely many representative intervals for joint index--scale updates without exhaustively enumerating codeword combinations. On Qwen3.8-27B, CoS-VQ outperforms QTIP and LLVQ on the evaluated language modeling, coding, and reasoning benchmarks. Ablations from the same initialization further show larger gains from joint updates than from scale-only refinement, highlighting joint selection as an alternative to increasing codeword dimensionality. The code is available at https://github.com/xzktx003/Cos_VQ.

Safety alignment research for large language models (LLMs) remains concentrated on high-resource languages, leaving open how existing alignment methods perform in low-resource, culturally distinct settings. We study this question for Hausa, a language spoken by more than 200 million people across West Africa. We introduce HausaSafeLM, one of the first Hausa safety preference dataset, built from 10,000 prompts derived from PKU-SafeRLHF: 3,000 human-corrected (gold) and 7,000 machine-translated (silver) examples spanning six harm macro-categories. Using Qwen2.5-7B-Instruct, we compare supervised fine-tuning (SFT) and SFT followed by Direct Preference Optimization (DPO) under both data regimes, evaluating on 500 harmful and 200 benign Hausa prompts.We find that the untuned base model appears highly safe by refusal-rate metrics alone, but this reflects limited Hausa generation ability rather than genuine alignment: its refusals are almost uniformly low quality. SFT improves both refusal fluency and precision and benefits substantially from the larger, noisier silver-tier data, while DPO sharpens refusal precision but comes at a steep, largely data-invariant cost to linguistic quality. No single configuration jointly maximizes safety, helpfulness, and fluency. Our results suggest that current alignment techniques teach models when to refuse more readily than how to do so fluently in underrepresented languages, and that refusal-rate metrics alone are an unreliable signal of safety in low-resource settings.

Vision Foundation Models (VFMs) pretrained with self-supervised learning, such as DINOv2, provide highly transferable features, yet adapting them to small, domain-shifted datasets is still challenging. We propose a regularizing multibranch ensemble head: a training and inference scheme that pairs low- and high-capacity branches, samples a single branch stochastically at each training step, and averages both branches in probability space at inference. The design adds negligible parameters and essentially no inference overhead over a single heavy head, while improving the accuracy-efficiency trade-off under data-scarce conditions. Accurate classification of plant growth stages is critical for precision agriculture, enabling timely decisions on fertilization, irrigation, and harvesting. Therefore, as a challenging practical case we use cucumber phenological stage classification, in which flowers and fruits are frequently occluded by foliage, and stage transitions manifest primarily as subtle changes in leaf morphology. The NALEF Food Safety dataset contains images of five cucumber growth stages captured under diverse real-world conditions. On it we evaluate a broad range of backbones, including DINOv2 and DINOv3 ViT and ConvNeXt variants, against strongly supervised counterparts (ResNet, EfficientNet, ViT). The proposed DINOv2 ViT-S with multibranch head achieves a Macro F1 of 0.978 and Cohen's Kappa of 0.971, matching the performance of models more than ten times larger in parameters and GFlops, while requiring only 12.2 GFlops and 22.3M parameters, making it deployable on smartphones and edge devices for phenological stage classification.

Knowledge improvement in large language models is shifting from predefined knowledge incorporation pipelines toward evaluation-guided updating. Existing methods use evaluation to localize model-specific knowledge weaknesses and guide targeted post-training. However, identifying what to improve does not necessarily provide sufficient information for deciding how to improve it. The same localized weakness can be compatible with distinct operational knowledge states: the relevant knowledge may be unavailable, available but difficult to access reliably, or accessible but ineffective for downstream use. We call this ambiguity knowledge-state aliasing. When compatible states favor different interventions, greater diagnostic resolution alone is insufficient for intervention selection. Moreover, model updates can change the underlying knowledge state before the observed weakness disappears, such that an intervention appropriate in one round may no longer address the bottleneck in the next. To address these challenges, we propose Recursive Knowledge Improvement (RKI), which formulates knowledge improvement as a sequential, state-aware intervention process. RKI organizes knowledge into graph-structured units that provide a shared basis for diagnostic evaluation and targeted intervention across rounds. At each round, RKI localizes knowledge weaknesses, estimates the model's operational knowledge state from diagnostic evidence and intervention history, selects a corresponding intervention, and re-evaluates the updated model before determining the next intervention. On HotpotEval and SafeInternalEval, RKI outperforms existing knowledge incorporation methods by 18.41 points on average. Further analyses show that knowledge states provide additional guidance for intervention selection beyond weakness localization and can change across model updates even when the observed weakness persists.

Continual learning requires large language models to acquire new knowledge and capabilities while retaining what they have learned. Fine-tuning and replay support acquisition and retention, but deciding whether incoming examples need supervision and which historical examples to revisit remains a challenge. We propose Compatibility-Aware Memory Replay (CAMR), an adaptive framework that jointly controls current-example supervision and historical replay. CAMR combines compatibility checks with a learned memory gate to guide an online policy over three actions: learning incoming examples with general replay, jointly reinforcing compatible related examples, and skipping direct supervision of verified mastered duplicates. The policy learns from acquisition, retention, and training-cost feedback under a bounded per-arrival replay budget. We evaluate CAMR with Qwen3-8B-Base on continual factual updating and cross-task and cross-domain learning. On WikiBigEdit, CAMR consistently leads evaluated methods in Original Match from 3K to 100K facts, reaching 93.66% after just 3K facts and surpassing Full-history Replay (LoRA) after 100K facts. It also achieves the highest final scores on all four selected TRACE tasks and maintains or improves earlier-task performance relative to post-acquisition levels in a sequence spanning financial question answering, mathematical reasoning, code generation, and summarization.

The Asch conformity experiment has recently gained considerable attention as a way of measuring conformity in large language models (LLMs). While previous studies have focused on demonstrating conformity in LLMs, our work adds a further quantitative layer to this approach. Instead of sampling a frontier LLM on a suite of questions, our approach samples a single question repeatedly, leveraging the probabilistic nature of LLMs to produce a distribution of answers. In doing so, we can compare the distributions of answers before and after the model is presented with the manufactured consensus. Our experiments confirm that frontier LLMs do conform, reinforcing prior literature, and, through our distribution-to-distribution comparisons, provide additional quantitative data on the conditions that lead to conformity and the degree to which it occurs.

Pathology foundation models (PFMs) differ substantially in architecture, pretraining objective, training data and representation dimensionality, yet it remains unclear whether they learn fundamentally different feature spaces or alternative coordinate systems of a shared underlying representation. We present MOSAIC, a systematic study of latent-space similarity, alignment and interoperability across $12$ patch-level and $6$ slide-level PFMs. We first quantify representational agreement using complementary similarity measures and analyze how architecture, pretraining objective, network depth and magnification influence latent geometry. We then evaluate multiple alignment strategies for mapping independently learned representations into a common latent space. The resulting representations are evaluated through cross-model agreement, retrieval, geometry preservation, and downstream multiple-instance learning across molecular and morphological prediction tasks. MOSAIC provides a unified framework for studying whether heterogeneous PFMs can move beyond representational similarity toward a shared and functionally transferable pathology representation space. Our code, benchmarks and additional results are available at https://researchsubmissions66.github.io/MOSAIC/.

Prospect theory splits risky choice into three parts: loss aversion, measured by the relative weight of losses to gains, $λ$ (e.g., $λ= 2$ implies that losses weigh twice as much as gains), diminishing sensitivity to the size of outcomes (i.e., the curvature and linearity of the value function), and the nonlinear weighting of probabilities. I examined whether Large Language Models (LLMs) display the same properties as people by meta-analyzing 102 fitted estimates from 11 studies covering 32 language models. The three components emerge differently. Loss aversion is modest: the pooled $λ$ is 1.26 (95% CI 1.06 to 1.46), far below the strongest human meta-analytic estimate ($λ= 1.96$) and compatible with more conservative ones (1.07 and 1.31). Matched by design, LLMs are compatible with people when gains and losses are symmetric ($λ= 1.15$ against 1.07), but well below them when losses are smaller than gains (1.46 against 2.3), where human loss aversion is strongest. Value functions are close to linear rather than concave under the Tversky-Kahneman method (1.01, against a human value of 0.67), but more concave than in people under the Tanaka method (0.52, against 0.67), so diminishing sensitivity depends on how it is measured. Probability weighting shows a human-like inverse S shape ($γ= 0.81$ under Tversky-Kahneman and 0.84 under the Tanaka method), though on few studies. In sum, LLMs show little loss aversion and little diminishing sensitivity, but retain human-like probability distortion, yet these generalizations shift when looking at specific model families and fitting methods.

Current uses of large language models (LLMs) for qualitative coding often rely on ad hoc prompting with limited guidance for reliability and human oversight. To address these limitations, we propose the Smart Coder Framework, a three-phase approach structuring human-LLM collaborative coding through Instruction, Prompt Strategy, and Reliability Check. We evaluated the framework using 72 middle-school mathematics teacher interview transcripts and a human-consensus benchmark. A factorial design compared instructed versus uninstructed prompts, few-shot versus few-shot plus chain-of-thought prompting, and single-pass versus self-consistency coding. Twelve transcripts were used for prompt development and calibration, and 60 for the formal evaluation. In the results, instruction consistently improved agreement. Chain-of-thought further improved agreement under adequate instruction, and self-consistency produced smaller, theme-dependent gains but helped identify unstable decisions for human review. Performance also varied by transcript characteristics. Overall, the framework systematically structures and monitors LLM-assisted qualitative coding while preserving human authority over coding decisions.

Vision foundation models for retinal imaging are large and computationally demanding, limiting point-of-care deployment. We ask how much of such a model's downstream utility survives compression into a smaller network. We propose a topology-aware discretization framework that distills a continuous foundation model into a discrete codebook, then uses it as the reconstruction target for masked image modeling (MIM). It combines gradient-free codebook learning via Growing Neural Gas, which yields an adjacency graph over the codes, with a graph-distance soft cross-entropy loss that replaces one-hot MIM targets with topologically graded supervision. With student pretraining on 383 public Duke iAMD optical coherence tomography (OCT) volumes, the pipeline distills an 86M-parameter teacher into a 22M-parameter student with at least 3.6x fewer FLOPs. The student performs close to, and in some settings above, its teacher across three classification and two segmentation benchmarks, with the magnitude and direction of the difference depending on the teacher, task, metric, and student capacity. On the external OLIVES benchmark, it surpasses its MIRAGE teacher on every metric (98.06% vs. 95.22% balanced accuracy). It works with two Vision Transformer (ViT-Base) teachers, retinal (MIRAGE) and natural-image (DINOv3); with the out-of-domain teacher, the student outperforms it on retinal-layer segmentation, indicating that the added value of retina-specific pretraining is task-dependent: largest for segmentation, while a general-domain teacher is competitive for classification. These results yield compact retinal models potentially suitable for deployment, transferring across OCT datasets, devices, and clinical tasks, including three datasets not used in the student's pretraining.

Linguistic productivity, as evidenced by the use of words in novel contexts, is one of the most striking facts about human language. However, how abstract the knowledge that enables such productivity is remains unclear. We compare people's and computational language models' linguistic productivity in contexts where successful generalization depends on form-based information. We focus on extending the use of a novel verb learned in the active voice to passive contexts. We find people extend active to passive uses in a way consistent with abstract generalizations concerning linguistic form, even when no plausible meaning can be assigned to the novel verb. In contrast, large language models succeeded only when a meaning like an existing verb's could be assigned to the novel verb. We conclude that computational models' syntactic knowledge remains tied to the distribution and meaning of specific lexical items in a way that human syntactic knowledge is not.

As the applications of AI continue to expand across various multimodal and multilingual domains, ensuring safe and responsible deployment of these systems has become an important concern. A wide-range of interventions have been proposed to make AI responses safer and more reliable. They often incur substantial computational or annotation costs, or cannot be extended to multilingual settings. In this work, we propose Selective Detection And Targeted Editing (SEDATE), a training-free approach based on detect-edit-verify strategy. SEDATE first detects the responsible neurons, edits them as and when necessary and finally verifies the model responses; all without any additional model training. We evaluate SEDATE across 7 languages and 4 model families. We further evaluate its effectiveness across 4 out-of-distribution languages and 4 additional datasets. We show that SEDATE achieves substantially superior safety scores across all languages and models over the best-performing baseline while retaining the general capabilities and incurring negligible overhead. Our results are robust, with the trends remaining similar for out-of-distribution languages and datasets.

Augmented large language model (LLM) systems have reported high benchmark accuracies in rare disease diagnosis, yet their real-world clinical utility remains unproven. A recent meta-analysis flagged augmented LLM diagnostic evaluations as having a high risk of bias due to suspected data leakage; however, the precise impact of these vulnerabilities on model inference has yet to be fully quantified. In this study, we present a comprehensive empirical audit to quantify evaluation artifacts across prevailing AI paradigms, using a repository of over 60,000 cases from diverse public and independent clinical cohorts. By reconstructing a state-of-the-art agentic framework, we demonstrate that candidate-space contamination artificially inflates performance: enforcing strict cross-dataset isolation resulted in a marked performance decline, with the cross-dataset mean Recall@1 decreasing from 54.3% (under conventional one-vs-rest evaluation) to 40.3% (Hedges' g = 0.94). Similar spatial overfitting in supervised classifiers caused diagnostic accuracy to drop from an internal mean of 42.9% to 16.7% across external cohorts (Cohen's d = 2.01). Furthermore, we quantify how LLMs infer diagnoses by exploiting retrospective therapeutic cues in clinical narratives, and how temporal misalignment in foundational knowledge bases leads to performance fluctuations. We conclude that prevailing evaluation protocols for augmented LLM systems inadvertently reflect data leakage rather than genuine medical reasoning. To address these systematic artifacts, we propose a rigorous evaluation framework - coupling strict data-isolation protocols with standardized benchmarking safeguards - to ensure future diagnostic AI for rare diseases is evaluated on actual clinical robustness.

Do preferences Large Language Models (LLMs) express in standard elicitation tasks predict their strategic behaviour? LLM agents interacting with humans and other AI systems face potentially conflicting objectives, such as individual welfare, group welfare, and environmental concerns. Prior work characterizes LLM preferences through behavioral-economic and psychological elicitation or strategic games, but rarely tests consistency across settings and prediction in other strategic environments. This requires modeling how competing objectives jointly shape behavior, whereas existing approaches typically measure preference dimensions in isolation. Drawing on utility theory, we introduce the Preference-Behavior Model (PBM), which combines a preference representation with mechanisms capturing imperfect utility maximization. We formulate estimation as an inverse-game problem, jointly inferring preferences and behavioral parameters from observed choices. We evaluate PBM in three social dilemmas that expose conflicting motives: a Public Goods Game, a Collective-Risk Dilemma, and a Common-Pool Resource dilemma. We find that while preferences estimated from established behavioral economic elicitation instruments do not consistently predict behavior in these dilemmas, models fitted in one dilemma retain predictive value in others, suggesting that inferred preferences capture meaningful aspects of LLM decision-making that elicitation instruments may miss. Finally, we assess alignment with persona instructions, finding that inferred preferences broadly reflect intended priorities with specific divergences across dimensions, LLMs, and games being identified by our approach. Our methodology provides a framework for studying how consistently LLM preferences explain behavior, assessing alignment with intended priorities, and revealing trade-offs and behavioral patterns that standard elicitation may miss.

Photoplethysmography (PPG) plays a central role in wearable health monitoring and clinical decision support. However, existing approaches to universal PPG representation learning are predominantly trained on high-quality large-scale datasets obtained from oximeters in hospital settings and often struggle to adapt to sparse and low-quality real-world datasets, thereby limiting their generalization to clinically relevant tasks on prevalent PPG data such as collected from wearable devices. Surprisingly, even after full fine-tuning, PPG foundation models cannot consistently outperform general-purpose time-series models. Therefore, a key question is how to effectively adapt PPG foundation models to target prediction tasks. Through empirical studies on various PPG foundation models, we find that pretrained models often exhibit inherent sparsity and computational redundancy, suggesting that PPG foundation models have learned to activate task-relevant network substructures for different prediction tasks. To preserve this valuable prior knowledge, we propose a structured pruning-and-fine-tuning framework, SPAFT, which regularizes the adaptive fine-tuning process by constraining it to a more relevant and compact parameter space based on parameter importance and structural redundancy. Extensive experiments across four PPG foundation models, five benchmark datasets, and ten downstream tasks demonstrate that SPAFT achieves win rates of 82.8% and 87.5% on regression and classification tasks, respectively. This prune-then-finetune paradigm often enables PPG foundation models to achieve state-of-the-art performance, laying a promising foundation for the real-world clinical deployment of PPG foundation models and the development of personalized PPG foundation models.