Large Language Models remain plagued by hallucinations. Recent work has sought to tame their prevalence using statistical techniques based on conformal prediction, with both theoretical and empirical success. However, these methods operate in a post-hoc fashion, treating the sampling procedure itself as atomic and then surgically altering samples to remove hallucinated claims. This disconnect between filtering and generation can result in samples that are incoherent, inconsistent, or simply unlikely under the model itself. Moreover, post-hoc surgery is unable to shift probability mass towards more useful and helpful responses. To address these issues, we propose to instead sample from approximations to an LLM posterior, where the conditioning event corresponds to a calibrated, high-scoring region. We develop a calibration procedure tailored to the setting of conditional sequential generation that effectively identifies this region and achieves target risk control. Empirically, we apply our method to case studies focused on open-ended biography generation and mathematical problem solving; compared to prior work, we obtain the same statistical guarantees, with higher downstream utility.

Retrieval-augmented generation systems for legal question answering typically retrieve passages based on semantic similarity and provide them to a language model, which then generates cited answers. Prior work assumes that highly ranked passages are most likely to be usefully cited by the model. Perturbation-based attribution methods, such as C-LIME, have been used exclusively for post-hoc explanation. However, on the AQuAECHR benchmark, semantic similarity does not correlate with passage attribution. Within a retriever's candidate pool, similarity-based ranking performs worse than random selection at surfacing gold citation paragraphs. To address this limitation, a lightweight cross-encoder is trained on continuous perturbation-based attribution scores to re-rank passages prior to generation. This approach is evaluated on the AQuAECHR benchmark, using two language models and five-fold cross-validation. The re-ranker substantially improves citation faithfulness and alignment with gold expert answers. Notably, two re-rankers trained independently on different models converge beyond their raw attribution agreement. This finding indicates that the cross-encoder reduces model-specific noise and produces a shared relevance signal that partially transfers across models, although same-model re-ranking remains more effective. These results demonstrate that perturbation-based attribution provides a practical, model-agnostic training signal for citation-aware retrieval.

Text-to-image models rely on text prompts as their primary interface to human intent. Prompts are encoded by a text encoder into embeddings that condition the image generation process. Beyond individual token meanings, text embeddings encode contextual information across the full prompt, such as compositionality and attribute binding. However, whether image models actually exploit this richer information remains underexplored. Here, we address the question: Which aspects of text representation are essential for image generation? We show that text-to-image diffusion transformer-based models commonly rely only on two relatively straightforward aspects of text representations: (i) the merging of adjacent tokens into a word representation, for words spanning multiple tokens, and (ii) word order, which is imprinted by the positional embedding of the text-encoder. To show this, we construct a new text embedding that encodes only individual word meanings and order but lacks any contextual information about the full prompt. We find that this bag of position-tagged words representation is sufficient to successfully guide image generation, achieving visual quality and text fidelity that are on par with full text embedding-guided generation. This demonstrates that, contrary to common belief, text-to-image models often do not use the rich information encoded in the text embedding beyond individual word meanings and word order. Instead, the decoding of complex linguistic structures is performed by the image model itself. Project webpage: https://nsping13.github.io/contextless-TTI/

Multi-modal Large Language Models (MLLMs) achieve strong performance on vision-language tasks, but incorporating visual inputs through a vision encoder (e.g., CLIP) substantially expands the attack surface, making these models vulnerable to visual adversarial perturbations. Prior defenses typically preserve compatibility with pretrained MLLMs by enforcing strict alignment to CLIP's original embedding space during adversarial fine-tuning; while practical, this constraint fundamentally limits achievable robustness. We present a systematic investigation of adversarial robustness in MLLMs. We first introduce a diagnostic CLIP-alignment protocol that predicts, prior to full MLLM training, which robust vision encoders will transfer effectively to the multimodal setting, revealing that large-scale multimodal adversarial pretraining, rather than unimodal scale alone, is the critical factor for strong robustness transfer. Integrating such encoders into MLLMs via end-to-end multimodal training yields average gains of 28 CIDEr points on captioning and 11.7% VQA accuracy under strong adversarial attacks compared to constrained plug-and-play baselines. We further show that adversarial training applied directly to a standard non-robust MLLM degrades both clean and adversarial performance, establishing robust visual representations as a strict prerequisite, while end-to-end adversarial training from a robust backbone delivers additional gains of 1.9 CIDEr points and 4.3% VQA accuracy. Beyond training-time defenses, lightweight test-time visual stochastic transformations serve as an effective black-box defense for non-robust MLLMs, elevating adversarial performance from near-zero to levels comparable with robust models. Finally, we show that our robust models substantially reduce toxic generation under white-box visual jailbreak attacks. Code and pretrained weights will be released publicly.

Graph Language Models (GLMs) have become a promising direction for adapting Large Language Models (LLMs) to graph learning tasks. By transforming graph topology and node information into graph tokens, GLMs allow LLMs to jointly process structured graph inputs and textual instructions. Yet, it remains unclear how LLMs internally interpret these graph tokens and whether graph tokens act as meaningful carriers of graph structure. In this work, we analyze how LLMs process graph information through graph-token behavior in representative GLM architectures. Findings. We find that the internal saliency of graph tokens in GLMs is not equivalent to graph information utilization. Graph sink tokens consistently emerge as activation-level outliers: they can be identified by massive activation values along a small set of hidden-state dimensions and are biased toward early graph-token positions. However, this activation-level saliency does not imply that these tokens are the main carriers of graph information. Unlike classical attention sinks in language and vision-language models, graph sink tokens do not necessarily attract the largest attention weights from query tokens. Through pruning, repositioning, and swapping interventions, we show that graph sink tokens are not the most important semantic or structural tokens for downstream prediction. Implications. Together, these results suggest that after current GLMs map graph structure into the LLM token space, the resulting graph-token representations do not naturally form a fully usable topology-aware internal representation; instead, they exhibit a decoupling between activation-level saliency and graph-semantic utility. This decoupling points to limitations in existing graph-token construction, placement, and alignment mechanisms.

Knowledge Graphs (KGs) are widely used to mitigate the limitations of Large Language Models (LLMs), such as outdated knowledge and hallucinations. Existing LLM-KG integration frameworks typically rely on predefined operators to retrieve factual knowledge from KGs and inject it into prompts for answer generation. This paradigm faces two critical bottlenecks: 1) Inflexibility: The predefined operators are limited in scope and thus lack sufficient compositional expressiveness to fully capture the complex semantics required by KG questions. 2) Unscalability: Direct injection of factual knowledge into prompts limits scalability in handling large-scale factual knowledge. To address these two bottlenecks, we propose Code-on-Graph (CoG), a programmatic reasoning framework for LLM-KG integration. Specifically, given the factual knowledge retrieved at each reasoning step, CoG first identifies the corresponding KG schemas and represents these schemas as Python classes, which serve as abstract interfaces to the retrieved facts. It then generates executable code grounded in these classes, with the retrieved facts instantiated as objects of the corresponding classes during execution. This design enables flexible code-based reasoning while avoiding the direct injection of large-scale factual knowledge into prompts. Experiments on WebQSP, CWQ, and GrailQA demonstrate that CoG outperforms prior state-of-the-art models by up to 10.5%.

Financial decision-making tasks such as stock recommendation and portfolio allocation typically estimate future return and risk and then select trades or allocations for an investor, and the chosen optimization objective often determines realized performance. However, because market conditions evolve over time, a fixed objective can be suboptimal across regimes, while regime-switching pipelines that rely on latent regime estimates can be noisy or delayed and frequent switching can increase turnover and operational instability. In this paper, we propose DOSS (Dynamic Objective Selection with Safeguards), a learning-based selector that directly chooses the decision-relevant objective function at each time point from interpretable statistical summaries of recent returns, selecting among a small set of candidates (e.g., return-seeking, loss-averse, and risk-adjusted) without introducing intermediate regime variables. DOSS formulates objective selection as a classification problem over objectives and performs sequential updates with a rolling window to make forward-looking selections without temporal leakage, while also outputting a confidence score for each proposal. To mitigate misselection and excessive switching in deployment, DOSS applies confidence-aware gating with a fail-safe that overrides low-confidence proposals to a conservative default and enforces explicit controls tied to switching frequency. We further integrate governance by positioning a Large Language Model (LLM) as an oversight component rather than a generator of new objectives: the LLM is restricted to accept a proposed objective or override it to a predefined safe default, with deterministic rule-based constraints triggering overrides when needed.

A central goal of large language model (LLM) research is to build agentic systems that can plan, act, and adapt through sustained interaction with dynamic environments. While recent LLM-based agents exhibit impressive contextual reasoning, their long-horizon decision-making remains fragile, often suffering from objective drift, where goals and plans drift over extended interactions. We introduce Multi$^2$, a hierarchical multi-agent decision-making framework that explicitly decomposes agent behavior into complementary roles. A high-level agent (System 1) focuses on context-aware sub-goal generation using supervised fine-tuning (SFT), while a low-level agent (System 2) executes atomic actions through offline-to-online reinforcement learning (RL) in interactive environments. This separation enables stable long-horizon control, mitigates objective drift, and allows efficient adaptation. Across diverse interactive environments, Multi$^2$ consistently outperforms strong agentic baselines, demonstrating improved robustness and coordination in multi-turn interaction. Beyond performance, we introduce and release three hierarchical benchmark datasets, filling a long-standing gap in training and evaluating hierarchical decision-making for LLM-based agents.

As language models are increasingly deployed in real-world applications, the ability to erase specific knowledge from them becomes critical for safety and compliance. Prominent methods seek persistent removal by updating the model's parameters, yet the target knowledge often can be recovered through adversarial prompting or relearning. In this work, we hypothesize this limitation stems in part from existing methods overlooking the embedding layer. To address this, we introduce EMBedding ERasure (EMBER), a plug-n-play erasure module that leverages Sparse Matrix Factorization for precise erasure of concept-related features from token embeddings. Through comprehensive evaluations across diverse concepts on Gemma-2-2B-it and Llama-3.1-8B-Instruct, we find that augmenting existing methods with EMBER consistently improves erasure efficacy and specificity across task formats, with minimal coherence loss. Moreover, it dramatically improves robustness to relearning, reducing regained accuracy by up to 50%, limiting it to 35% on Llama compared to 70%-76% for prior methods. Further analysis shows that the coherence cost is localized, affecting only a small set of concept-exclusive tokens. Our work establishes that precise embedding-level intervention is necessary for robust concept erasure, and demonstrates that existing methods can benefit from such augmentation.

To enable meaningful human-robot interaction (HRI), a robot must continuously assess engagement by consistently tracking users over time. State-of-the-art computer vision models, however, are heavily optimized for surveillance or autonomous driving. A social robot faces distinct egocentric challenges, such as humans bouncing, obstructing each other, or leaving the frame. Frequent identity switches (IDSW) cause the robot to lose its footing mid-conversation. To address this, we introduce a novel, custom-annotated egocentric dataset collected via the Furhat robot to capture complex social dynamics. We present a systematic evaluation isolating detection errors from tracking logic, comparing face versus body tracking, and assessing the impact of extended spatial memory and appearance re-identification (ReID). Results indicate that increasing spatial memory mitigates prolonged occlusions but fails on complex dynamic events. Integrating ReID resolves complex switches but exhibits opposing effects: it substantially improves body tracking stability, yet causes facial IDSW to spike due to profile angle sensitivity. Ultimately, our optimized pipeline reduces IDSW by 49\%, mitigating interaction breakdowns. Because standard benchmarks lack dense, close-quarter occlusions, this work highlights the critical need for natively captured social dynamics to truly validate HRI perception models.

Medical Vision-Language Models (VLMs) are typically evaluated on English radiology visual question answering benchmarks, leaving their robustness under non-English clinical language largely unexplored. We introduce IndoRad-VQA, an Indonesian adaptation of VQA-RAD, to assess whether medical VLMs retain radiology reasoning ability when questions are asked in Bahasa Indonesia. Radiology question-answer pairs are translated into Indonesian with self-evaluation-based quality control to preserve clinical meaning, terminology consistency, and answer equivalence. We evaluate general-purpose, Southeast Asian multilingual, and medical-specific VLMs under English and Indonesian prompting settings. Beyond accuracy, we quantify the language robustness gap between English and Indonesian inputs. We also conduct an error analysis to identify failure modes of question answering, such as yes/no flips, laterality errors, and output-language mismatches. Our findings show that strong performance on English medical VQA benchmarks does not necessarily translate to robust behavior in Indonesian clinical contexts. We observe a performance gap of 8 to 25 percent between the English and Indonesian settings, depending on the evaluation metric. These results highlight the need for more inclusive multilingual evaluation of medical multimodal foundation models. The dataset is available at https://huggingface.co/datasets/Lab-IS/IndoRad-VQA.

Recent AI agents can flexibly invoke skills to solve complex tasks, but their long-term improvement is fundamentally constrained by a lack of systematic skill construction, accumulation, and transfer. In particular, without a unified framework for skill consolidation, agents tend to redundantly construct similar capabilities across different tasks, are unable to effectively transform experience into reusable assets, and struggle to generalize task-specific skills to novel scenarios. To address this limitation, we propose SkillPyramid, a skill consolidation framework that reuses existing skill experience for broader task generalization. Operating on a hierarchical skill topology, SkillPyramid further introduces a self-evolution mechanism that enables agents to compose, validate, and incorporate new skills during task execution. Experiments on ALFWorld, WebShop, and ScienceWorld across four backbone models show that SkillPyramid substantially increases the average reward by 38.0% and reduces execution steps by 27.7%. Overall, our method transforms a skill collection from a static resource pool into a dynamic evolution system.

Survival Analysis (SA) is a statistical framework that models the time span until some event of interest occurs. Widely used in several domains, including healthcare and churn prediction, a central challenge in its applicability stems from the time of the event being partially observed or \emph{right-censoring}. Tabular Foundation Models (TFM) have attracted significant interest in recent years due to their ability to perform prediction tasks in a single forward pass, requiring no dataset-specific parameter fitting. Despite their success, their application to prediction tasks on time-to-event data remains difficult due to right censoring. In this work, we present a training-free method to survival regression by leveraging TFMs to both predict the time of the event and iteratively impute right-censored data. Our method uses a TFM to construct an Accelerated Failure Time (AFT) model requiring no training beyond fitting a single scalar parameter. Subsequently, by building on the Buckley-James estimator, we introduce a non-parametric in-context estimator for right-censored data. Our experiments on standard survival analysis benchmarks show that our method is competitive with several parametric and semi-parametric survival regression models that require training, including Cox regression and parametric AFT models.

Supervised fine-tuning (SFT) improves end-to-end classical planning in large language models (LLMs), but do these models also learn to represent and reason about the planning problems they are solving? Due to the relative complexity of classical planning problems and the challenge that end-to-end plan generation poses for LLMs, it has been difficult to explore this question. In our work, we devise and perform a series of interpretability experiments that holistically interrogate world model recovery by examining both internal representations and generative capabilities of fine-tuned LLMs. We find that: a) Supervised fine-tuning on valid action sequences enables LLMs to linearly encode action validity and some state predicates. b) Models that struggle to use output probabilities for classifying action validity may still learn internal representations that separate valid from invalid actions. c) Broader state space coverage during fine-tuning, such as from random walk data, yields more accurate recovery of the underlying world model. In summary, this work contributes a recipe for applying interpretability techniques to planning LLMs and generates insights that shed light on open questions about how knowledge is represented in LLMs.

Pretraining cost is a major bottleneck for research on tabular foundation models, slowing the iteration cycle for new architectures, priors, and optimization ideas. Yet the community lacks a simple way to compare and accumulate pretraining speedups. We introduce a community speedrun for nanoTabPFN: contributors modify a single-file training script and compete to reach a fixed downstream ROC AUC target on subsampled TabArena using one NVIDIA L40S GPU. The current best record reaches the target in 0.92 minutes, an 81x speedup over the 74.32 minute baseline while using 22x fewer synthetic datasets. The speedrun format provides a simple protocol for the community to add, verify, and stack pretraining improvements, with the leaderboard open to contributions. Code and records are available at https://github.com/borawhocodess/modded-nanotabpfn.

Methane is a potent greenhouse gas, and detecting leaks early via hyperspectral satellite imagery can help climate change mitigation efforts. Meanwhile, many existing hyperspectral missions only capture areas manually targeted by operators, thus missing potential events of interest. To overcome slow downlink rates cost-effectively, onboard detection is a viable solution. However, traditional methane detection methods are too computationally demanding for resource-limited onboard hardware. This work accelerates methane detection by focusing on efficient, low-power algorithms. In particular, we test fast target detection ACE and CEM methods that have not been previously used for methane detection and propose Mag1c-SAS -- a significantly faster variant of the current state-of-the-art Mag1c algorithm. To explore their detection potential, we integrate them with a machine learning model based on U-Net and LinkNet. We evaluate our methods on the STARCOP dataset and a novel EMIT-MSeg dataset, which we introduce and open-source alongside a high-quality annotation strategy. The proposed Mag1c-SAS approach proves highly effective by operating ~80x faster than the original Mag1c approach, providing a visually similar, but noisier result. When additionally paired with the lightweight LinkNet approach, it effectively reduces noise, achieving AUPRC score improvements of over 30 pp on EMIT-MSeg compared to the baseline Mag1c approach, and an F1 score on STARCOP ~4 pp higher. We evaluate two novel band selection strategies and confirm the system's onboard viability through hardware profiling, demonstrating marginal power consumption and efficient CPU/RAM utilization. We release the final system in a user-friendly and lightweight PyPI library at: https://pypi.org/project/onboard-methane-detection/, alongside all experimental code, models, and data at: https://github.com/zaitra/methane-filters-benchmark.

Recent unified audio generation models can support diverse tasks across speech, sound effects, and music, but most of them still focus on isolated task-level synthesis. However, real video production often requires multiple components of a complete audio track to be generated jointly and consistently for the same video. We present Foley-Omni, a unified multimodal audio generation model that extends isolated task-level synthesis to complete video soundtrack generation by jointly modeling speech, sound effects, and music within a shared latent generation process. To support training and reproducible evaluation, we develop an audiovisual data curation pipeline and introduce V2ST-Bench, a benchmark for holistic video soundtrack generation evaluation. Experiments show that Foley-Omni achieves competitive performance with expert systems on individual synthesis tasks, while improving speech intelligibility, audiovisual consistency and perceptual quality for mixed soundtrack generation.

Recent deep unfolding networks (DUNs) have advanced Compressive Sensing (CS) by effectively integrating iterative optimization with deep learning architectures. However, most CS approaches predominantly confine their inference to a single solution space, neglecting the inherent ill-posedness of CS problems that intrinsically permits multiple plausible candidate hypotheses. In this paper, a novel Multi-Hypothesis Collaborative Deep Unfolding CS Network (MHC-DUN) is proposed, which explicitly models and leverages multiple hypotheses by jointly optimizing across diverse solution spaces. Specifically, following the Proximal Gradient Descent algorithm, MHC-DUN jointly performs gradient descent and proximal mapping within this multi-hypothesis paradigm. i) For gradient descent, a well-designed AlphaNet is introduced to dynamically predict spatially varying step sizes for all hypotheses, enabling collaborative gradient updates across multiple solutions. ii) For proximal operator, a sophisticated multi-hypothesis collaborative proximal mapping module is designed, which leverages both intra-hypothesis and inter-hypothesis correlation priors to jointly refine multiple solutions. To enable end-to-end training, a novel composite loss function is designed, which balances measurement fidelity, hypothesis diversity, and reconstruction accuracy, encouraging exploration of complementary solutions while maintaining reconstruction fidelity. Experimental results reveal that the proposed CS method outperforms existing CS networks.

Large language models for code generation often need to use APIs that are absent from their pretraining data. This requires more than recalling a function name: models must coordinate signatures, module paths, input-output contracts, semantics, and executable usage patterns. Existing novel-API benchmarks are typically static, rely on coarse pass/fail metrics, or use synthetic APIs that may not reflect real library evolution. We introduce NovelAPIBench, a fully automated dynamic benchmark that, for any base model and target library, discovers novel APIs, extracts decomposed knowledge bundles, generates executable coding tasks, and assigns failed samples to six diagnostic categories. Across about 1.9K tasks, four base models, and five domains, we compare knowledge injected through retrieval with knowledge internalized through parametric adaptation. We find that knowledge components are not interchangeable: usage examples are the strongest standalone signal, while the best two-component setting pairs signatures with either mechanisms or examples depending on the domain and backbone. Adding more context, especially source code, can hurt by increasing import-path errors. Parametric adaptation also does not replace retrieval once external knowledge is removed; rather, fine-tuning mainly teaches models how to use provided bundles, and this ability transfers to held-out libraries. These results suggest that retrieval and tuning play complementary roles: retrieval supplies volatile API content, while tuning improves procedural integration.

Recent work in defeasible reasoning has seen notions of preferential semantics and entailment in the style of Kraus et al. applied to modal logics. However, work in this field has focussed primarily on satisfiability checking, and monotonic notions of entailment, which may be inferentially weak. One particular modal logic where this has been introduced is propositional standpoint logics, where modalities can express the views of different viewpoints. This has resulted in the formalisation of propositional defeasible standpoint logic (PDSL). In this paper, we propose a means of lifting the class of (non-monotonic) rational entailment relations from traditional KLM-style reasoning to a fragment of PDSL. In order to do so, we extend the expressivity of PDSL via situated standpoint conditionals, allowing us to talk about a defeasible conditional holding in the context of a given standpoint. This allows us to re-characterise the syntax of PDSL in terms of situated conditionals, and shows that a large fragment of PDSL is expressible as a set of situated conditionals. We then focus on characterising non-monotonic entailment in this fragment, defining a method to transport any ranking-based entailment relation from the propositional case into the PDSL case. This is first described in the general case and then considered in the specific cases of rational and lexicographic closures, providing a faithful translation of each inference into PDSL. We also show that entailment-checking in this fragment of PDSL can be done largely using algorithms from the propositional case, while preserving complexity bounds.

Adapting foundation large language models to a user's task or preferred style through fine-tuning can result in compromising the model's safety. Previous works examined the effects of fine-tuning on model safety in limited and seemingly random experimental settings. We argue that anchoring fine-tuning to a specific capability goal is essential for avoiding arbitrary empirical choices, allowing us to draw meaningful conclusions about safety impacts, and to compare mitigation methods on a consistent basis. We conduct a multi-dimensional evaluation of the effects of fine-tuning on model behavior by focusing on capability as well as safety. Our results surface important issues that (1) fine-tuned models can produce incoherent generations in response to safety prompts, (2) automated safety judgments are unreliable for such incoherent outputs, and (3) the conclusions about the effects of fine-tuning can change depending on the choice of safety benchmark as well as the safety evaluator.

This paper constructs the first benchmark on semi-supervised multi-modal crowd counting. To lay the foundation for this unexplored task, we first formulate the semi-supervised multi-modal setting and a standardized protocol that specifies the labeled-unlabeled data partition across different labeled ratios. Next, to establish solid reference points, we carefully tailor a diverse set of representative baselines, including existing fully supervised multi-modal methods and semi-supervised single-modal methods. Then, we carefully evaluate their performance under our proposed benchmark. Codes and the data partition will be released on https://github.com/HenryCilence/Semi-supervised-Multimodal-Crowd-Counting.

Large Language Models exhibit paradoxical fragility in fundamental arithmetic, implying a disconnect between internal computation and discrete output. By analyzing the residual stream geometry during multi-operand addition, we identify the Iso-Raw-Sum Trajectory (IRST), a geometric structure where representations are anchored by semantic digits and modulated by continuous carry fibers. We propose the Noisy Quantization Model to explain this geometry, framing arithmetic errors as Geometric Slippages caused by internal neural noise pushing a continuous, latent Carry Potential across quantization thresholds. This geometric framework further elucidates Probe Versatility, explaining how lightweight probes can disentangle coexisting latent signals (such as ground truth versus hallucination) from a single activation vector. Finally, we validate these insights through a geometric consistency check method that effectively detects and corrects these quantization failures during inference. Our code is available at https://github.com/RL-MIND/Shape-of-Addition.

Comprehensive molecular profiling is essential for modern precision oncology but remains hindered by prohibitive costs, specimen exhaustion, and protracted turnaround times. While pathology foundation models (PFMs) have demonstrated potential for inferring molecular phenotypes from routine hematoxylin and eosin (H&E) whole-slide images (WSIs), current architectures primarily rely on vision-centric self-supervised learning or vision-language alignment, lacking the spatially resolved molecular supervision required to connect subtle morphological features with underlying genomic alterations. Spatial transcriptomics (ST) emerges as a transformative technology that enables transcriptomic quantification within intact tissue sections, thereby preserving the precise spatial link between histology and molecular profiles. In this study, we present a Spatial Transcriptomics-guided Alignment framework for Molecular Profiling (STAMP), which endows PFMs with intrinsic molecular awareness. To support this paradigm, we curated HumanST-1k, a human ST dataset spanning diverse anatomical organs and sequencing platforms. This atlas yields 1.8 million pairs of H&E patches and corresponding transcriptomic profiles, providing a corpus that links histological structures with their molecular states. To mitigate the technical noise inherent to raw transcriptomics, STAMP applies a pathway-informed alignment strategy that aggregates transcriptomic data into biologically functional pathways, which are subsequently integrated into PFMs via parameter-efficient fine-tuning. This alignment enriches the representation space of PFMs and unlocks their capacity to resolve sub-visual molecular signatures. The clinical utility of these augmented representations was validated through a multi-tier evaluation framework.

We investigate whether large language models produce different medical triage recommendations for identical neurological symptoms when only the patient's stated gender and age vary. Using three model families--Gemini 3.5 Flash, Claude Sonnet 4.6, and GPT-5.4-mini--we present a standardized symptom profile (persistent headache, blurred vision, morning nausea, visual disturbances) across seven demographic conditions: three age groups (25, 38, 65) x two genders (male, female), plus a gender-unspecified baseline (n = 30 per condition per model, 630 total trials). We find a stark, systemic gender-dependent triage disparity: young women receive significantly lower emergency room (ER) referral rates than age-matched men (Gemini: 0% vs. 23.3%; Claude: 6.7% vs. 96.7%; GPT: 6.7% vs. 66.7%, all p < 0.001). The disparity disappears at age 65 for all models. The primary mechanism is diagnostic substitution: the models anchor on a gender-associated diagnosis, preferentially classifying young women with Idiopathic Intracranial Hypertension (IIH)--a condition epidemiologically linked to women of childbearing age--while diagnosing men with generic increased intracranial pressure with space-occupying lesions in the differential. This diagnostic closure routes female patients to lower-urgency care (outpatient doctor appointments) despite comparable severity ratings (7-9/10). Our findings demonstrate that clinical LLMs replicate documented human clinical biases by using epidemiological priors to suppress triage urgency, suggesting that AI triage engines must decouple urgency assessment from probabilistic diagnostic priors. We release all code, prompts, and raw results.

Understanding short online videos involves more than identifying visible objects and actions; video makers often include an underlying message or purpose in the clip. We introduce VidMsg, a benchmark for evaluating implicit message understanding in short, internet-native video clips. VidMsg contains 400 YouTube-derived clips across 9 practical topic areas and 52 fine-grained target messages, covering domains such as career and finance, education, health and well-being, culture, safety, sustainability, and lifestyle. VidMsg is constructed through a message-first pipeline: an LLM first translates target messages into indirect search scenarios, which are used to retrieve candidate clips. Human annotators then retain clips that convey the intended message without being overly explicit. VidMsg is designed primarily for bidirectional message-clip retrieval for scalable applications such as video search and recommendation, where systems must capture holistic video understanding. In addition to retrieval, VidMsg includes a diagnostic multiple-choice QA benchmark, where models select the intended message of a clip from semantically related alternatives. Experiments with contemporary video-language and retrieval models show that strong models often fail on VidMsg, because the task requires pragmatic inference, integration of contextual cues, and discrimination among semantically close messages. We also introduce VidVec-Msg, a baseline method that improves message-oriented retrieval while leaving substantial headroom for future work.

Assessing the quality of time series (TS) data is fundamental yet inherently challenging due to the multifaceted nature of quality dimensions. Recently, large language models (LLMs) have emerged as a promising paradigm for TS quality assessment via pairwise comparison and per-dimension evaluation. However, existing approaches rely on manually predefined quality dimensions and purely text-based reasoning, leaving it unknown whether LLMs can identify truly relevant quality dimensions or perform grounded and quantitative quality comparisons. To investigate this, we construct TSQBench, a dedicated benchmark for evaluating LLMs on two progressive capabilities: (i) understanding and identifying relevant quality dimensions, and (ii) performing quality comparison under specific dimensions. Our analysis reveals that current LLMs consistently struggle with both dimension identification and evidence-grounded quality comparison. To address these limitations, we propose TSQAgent, a novel agentic reasoning framework for TS quality rating consisting of three collaborative roles: Perceiver for focused dimension selection, Inspector for dimension-wise quantitative analysis, and Adjudicator that aggregates and refines the final judgment. In particular, we introduce an agentic reasoning strategy that instills the ability to identify and prioritize the most relevant quality dimensions, and further propose an agent workflow equipped with external analytical tools to enable precise quantitative comparisons over selected dimensions. Experiments on both the proposed benchmark and eleven real-world datasets demonstrate that our framework not only substantially improves LLMs' capabilities in quality understanding and quantitative comparison but also effectively translates these improvements into better quality-aware data selection, leading to enhanced downstream performance and data efficiency.

Large language models (LLMs) have achieved remarkable progress in open-ended text generation, yet they remain prone to hallucinating incorrect or unsupported content, which undermines their reliability. This issue is exacerbated in long-form generation due to hallucination snowballing, a phenomenon where early errors propagate and compound into subsequent outputs. To address this challenge, we propose a novel inference-time hallucination mitigation framework, named Segment-wise HAllucination Rejection Sampling (SHARS), which uses an arbitrary hallucination detector to identify and reject hallucinated segments during generation and resample until faithful content is produced. By retaining only confident information and building subsequent generations upon it, the framework mitigates hallucination accumulation and enhances factual consistency. To instantiate this framework, we adopt semantic uncertainty as the detector and introduce several vital modifications to address its limitations and better adapt it to long-form text. Our method enables models to self-correct hallucinations without requiring external resources such as web search or knowledge bases, while remaining compatible with them for future extensions. Empirical evaluations on standardized hallucination benchmarks demonstrate that our method substantially reduces hallucinations in long-form generation while preserving or even improving the informativeness of generation. Code is available at: https://github.com/TreeLLi/hallucination-rejection-sampling.

Vision-language models (VLMs) have been explored for visual programming, where they generate code to solve visual tasks. However, most prior work focuses on visual programming for productivity; it remains unclear how well current VLMs perform on education-oriented visual programming and what factors limit their performance. To bridge this gap, we introduce TurtleAI, a benchmark containing 823 tasks curated based on real-world visual programming tasks in the Turtle Graphics domain. Solving these tasks requires models to perceive geometric patterns, reason about spatial relationships, and synthesize Python code that faithfully reproduces geometric patterns. We evaluate 20+ VLMs, including GPT-5, GPT-4o, and Qwen2-VL-72B, and find that they struggle significantly, with most achieving success rates below 30%. To address these limitations, we propose a data generation technique that requires only a small set of seed samples. Fine-tuning Qwen2-VL-72B on the resulting synthetic data yields an improvement of about 20% on real-world tasks. Our failure analysis reveals that GPT-4o struggles with spatial reasoning and precise visual replication, whereas fine-tuning primarily improves the alignment between visual reasoning and code implementation.

Large Reasoning Models (LRMs) have demonstrated impressive capabilities in many tasks, yet they struggle with reliably following multiple instructions, either by failing to satisfy individual constraints or by struggling to balance competing constraints simultaneously. We formalize this challenge as the Constraint Adherence Problem (CAP). This paper introduces a novel framework that addresses CAP by representing instructions as a structured knowledge graph of constraints. Our approach, Constraint Relationship Graph Completion (CRGC), explicitly models relationships between constraints, identifies adherence challenges, and discovers ``bridge constraints'' that help the model better focus on and reconcile requirements. Bridge constraints act as auxiliary instructions that make primary constraints more salient and compatible. Unlike existing approaches that enhance instruction following through general training methods, CRGC specifically improves constraint satisfaction by leveraging the model's own knowledge to create better pathways for generation. Experiments across three popular instruction following datasets demonstrate that our approach reduces constraint violations by 39% compared to standard prompting while maintaining reasoning abilities of large reasoning models.