Spoken communication plays a central role in clinical workflows. In radiology, for example, most reports are created through dictation. Yet, nearly all medical AI systems rely exclusively on written text. In this work, we address this gap by exploring the feasibility of learning visual-language representations directly from spoken radiology reports. Specifically, we synthesize a large-scale dataset (Speech-RATE) of spoken radiology reports and train SpeechCT-CLIP, a contrastive model that aligns speech and 3D CT volumes in a shared representation space. While naive speech-based models underperform compared to text-trained counterparts, we show that knowledge distillation from a pretrained text-image CLIP model effectively transfers semantic alignment capabilities from text to speech, substantially narrowing this gap. Experiments demonstrate improved zero-shot classification F1 from 0.623 to 0.705, recovering 88% of the performance difference, and strong retrieval results without requiring text at inference. These findings highlight speech as a practical alternative to text in multimodal pretraining and open the door to voice-driven diagnostic support tools in clinical practice.

The advancement of computational psychology requires AI tools capable of deeply understanding counseling dialogues. Existing audio language models (AudioLLMs) often rely on single speech encoders pre-trained on general data, struggling to capture domain-specific features like complex emotions and professional techniques. To address this, we propose WEE-Therapy, a multi-task AudioLLM incorporating a Weak Encoder Ensemble (WEE) mechanism. This supplements a powerful base encoder with a pool of lightweight, specialized encoders. A novel dual-routing strategy combines stable, data-independent domain knowledge with dynamic, data-dependent expert selection. Evaluated on emotion recognition, technique classification, risk detection, and summarization, WEE-Therapy achieves significant performance gains across all tasks with minimal parameter overhead, demonstrating strong potential for AI-assisted clinical analysis.

The growth of highly advanced Large Language Models (LLMs) constitutes a huge dual-use problem, making it necessary to create dependable AI-generated text detection systems. Modern detectors are notoriously vulnerable to adversarial attacks, with paraphrasing standing out as an effective evasion technique that foils statistical detection. This paper presents a comparative study of adversarial robustness, first by quantifying the limitations of standard adversarial training and then by introducing a novel, significantly more resilient detection framework: Perturbation-Invariant Feature Engineering (PIFE), a framework that enhances detection by first transforming input text into a standardized form using a multi-stage normalization pipeline, it then quantifies the transformation's magnitude using metrics like Levenshtein distance and semantic similarity, feeding these signals directly to the classifier. We evaluate both a conventionally hardened Transformer and our PIFE-augmented model against a hierarchical taxonomy of character-, word-, and sentence-level attacks. Our findings first confirm that conventional adversarial training, while resilient to syntactic noise, fails against semantic attacks, an effect we term "semantic evasion threshold", where its True Positive Rate at a strict 1% False Positive Rate plummets to 48.8%. In stark contrast, our PIFE model, which explicitly engineers features from the discrepancy between a text and its canonical form, overcomes this limitation. It maintains a remarkable 82.6% TPR under the same conditions, effectively neutralizing the most sophisticated semantic attacks. This superior performance demonstrates that explicitly modeling perturbation artifacts, rather than merely training on them, is a more promising path toward achieving genuine robustness in the adversarial arms race.

In the World Wide Web, reliable time series forecasts provide the forward-looking signals that drive resource planning, cache placement, and anomaly response, enabling platforms to operate efficiently as user behavior and content distributions evolve. Compared with other domains, time series forecasting for Web applications requires much faster responsiveness to support real-time decision making. We present KAIROS, a non-autoregressive time series forecasting framework that directly models segment-level multi-peak distributions. Unlike autoregressive approaches, KAIROS avoids error accumulation and achieves just-in-time inference, while improving over existing non-autoregressive models that collapse to over-smoothed predictions. Trained on the large-scale corpus, KAIROS demonstrates strong zero-shot generalization on six widely used benchmarks, delivering forecasting performance comparable to state-of-the-art foundation models with similar scale, at a fraction of their inference cost. Beyond empirical results, KAIROS highlights the importance of non-autoregressive design as a scalable paradigm for foundation models in time series.

Contextual Markov Decision Processes (CMDPs) offer a framework for sequential decision-making under external signals, but existing methods often fail to generalize in high-dimensional or unstructured contexts, resulting in excessive computation and unstable performance. We propose an information-theoretic summarization approach that uses large language models (LLMs) to compress contextual inputs into low-dimensional, semantically rich summaries. These summaries augment states by preserving decision-critical cues while reducing redundancy. Building on the notion of approximate context sufficiency, we provide, to our knowledge, the first regret bounds and a latency-entropy trade-off characterization for CMDPs. Our analysis clarifies how informativeness impacts computational cost. Experiments across discrete, continuous, visual, and recommendation benchmarks show that our method outperforms raw-context and non-context baselines, improving reward, success rate, and sample efficiency, while reducing latency and memory usage. These findings demonstrate that LLM-based summarization offers a scalable and interpretable solution for efficient decision-making in context-rich, resource-constrained environments.

Although numerous strategies have recently been proposed to enhance the autonomous interaction capabilities of multimodal agents in graphical user interface (GUI), their reliability remains limited when faced with complex or out-of-domain tasks. This raises a fundamental question: Are existing multimodal agents reasoning spuriously? In this paper, we propose \textbf{Agent-ScanKit}, a systematic probing framework to unravel the memory and reasoning capabilities of multimodal agents under controlled perturbations. Specifically, we introduce three orthogonal probing paradigms: visual-guided, text-guided, and structure-guided, each designed to quantify the contributions of memorization and reasoning without requiring access to model internals. In five publicly available GUI benchmarks involving 18 multimodal agents, the results demonstrate that mechanical memorization often outweighs systematic reasoning. Most of the models function predominantly as retrievers of training-aligned knowledge, exhibiting limited generalization. Our findings underscore the necessity of robust reasoning modeling for multimodal agents in real-world scenarios, offering valuable insights toward the development of reliable multimodal agents.

Interpreting the internal behavior of large language models trained on code remains a critical challenge, particularly for applications demanding trust, transparency, and semantic robustness. We propose Code Concept Analysis (CoCoA): a global post-hoc interpretability framework that uncovers emergent lexical, syntactic, and semantic structures in a code language model's representation space by clustering contextualized token embeddings into human-interpretable concept groups. We propose a hybrid annotation pipeline that combines static analysis tool-based syntactic alignment with prompt-engineered large language models (LLMs), enabling scalable labeling of latent concepts across abstraction levels. We analyse the distribution of concepts across layers and across three finetuning tasks. Emergent concept clusters can help identify unexpected latent interactions and be used to identify trends and biases within the model's learned representations. We further integrate LCA with local attribution methods to produce concept-grounded explanations, improving the coherence and interpretability of token-level saliency. Empirical evaluations across multiple models and tasks show that LCA discovers concepts that remain stable under semantic-preserving perturbations (average Cluster Sensitivity Index, CSI = 0.288) and evolve predictably with fine-tuning. In a user study on the programming-language classification task, concept-augmented explanations disambiguated token roles and improved human-centric explainability by 37 percentage points compared with token-level attributions using Integrated Gradients.

Transforming natural-language requests into reliable, production-ready data transformations remains challenging: correctness depends on precise schema linking and warehouse-specific SQL dialects, while the strongest supervision available during training--execution success and result matching--are provided only at the sequence level. At the same time, assembling large, execution-validated corpora is costly, and token-level objectives misalign with these global signals, yielding unstable optimization and limited portability. We introduce Thinkquel, a fine-tuned model for producing robust, portable, and execution-validated database queries. Methodologies in Thinkquel integrates a novel synthetic data pipeline, TS-SQL, that leverages dbt as a portable intermediate representation with a span-aware reinforcement learning objective, and Token-Sequence GRPO (TS-GRPO), specifically designed to bridge the gap between token-level training signals and sequence-level execution rewards when finetuning LLMs. On the 500-example TS-SQL test set, Thinkquel (32B) reaches 93.2% execution success and 61.8% exact-result match with a two-stage SFT curriculum, improving over the base model by 67.2% (exec.) and 44.4% (match). In Spider (14B) experiments, TS-GRPO increases training stability and speeds convergence of the execution-match reward relative to GRPO and GSPO.

We present DM-Bench, the first benchmark designed to evaluate large language model (LLM) performance across real-world decision-making tasks faced by individuals managing diabetes in their daily lives. Unlike prior health benchmarks that are either generic, clinician-facing or focused on clinical tasks (e.g., diagnosis, triage), DM-Bench introduces a comprehensive evaluation framework tailored to the unique challenges of prototyping patient-facing AI solutions in diabetes, glucose management, metabolic health and related domains. Our benchmark encompasses 7 distinct task categories, reflecting the breadth of real-world questions individuals with diabetes ask, including basic glucose interpretation, educational queries, behavioral associations, advanced decision making and long term planning. Towards this end, we compile a rich dataset comprising one month of time-series data encompassing glucose traces and metrics from continuous glucose monitors (CGMs) and behavioral logs (e.g., eating and activity patterns) from 15,000 individuals across three different diabetes populations (type 1, type 2, pre-diabetes/general health and wellness). Using this data, we generate a total of 360,600 personalized, contextual questions across the 7 tasks. We evaluate model performance on these tasks across 5 metrics: accuracy, groundedness, safety, clarity and actionability. Our analysis of 8 recent LLMs reveals substantial variability across tasks and metrics; no single model consistently outperforms others across all dimensions. By establishing this benchmark, we aim to advance the reliability, safety, effectiveness and practical utility of AI solutions in diabetes care.

Recent advancements in generative models have allowed the emergence of a promising paradigm for recommender systems (RS), known as Generative Recommendation (GR), which tries to unify rich item semantics and collaborative filtering signals. One popular modern approach is to use semantic IDs (SIDs), which are discrete codes quantized from the embeddings of modality encoders (e.g., large language or vision models), to represent items in an autoregressive user interaction sequence modeling setup (henceforth, SID-based GR). While generative models in other domains exhibit well-established scaling laws, our work reveals that SID-based GR shows significant bottlenecks while scaling up the model. In particular, the performance of SID-based GR quickly saturates as we enlarge each component: the modality encoder, the quantization tokenizer, and the RS itself. In this work, we identify the limited capacity of SIDs to encode item semantic information as one of the fundamental bottlenecks. Motivated by this observation, as an initial effort to obtain GR models with better scaling behaviors, we revisit another GR paradigm that directly uses large language models (LLMs) as recommenders (henceforth, LLM-as-RS). Our experiments show that the LLM-as-RS paradigm has superior model scaling properties and achieves up to 20 percent improvement over the best achievable performance of SID-based GR through scaling. We also challenge the prevailing belief that LLMs struggle to capture collaborative filtering information, showing that their ability to model user-item interactions improves as LLMs scale up. Our analyses on both SID-based GR and LLMs across model sizes from 44M to 14B parameters underscore the intrinsic scaling limits of SID-based GR and position LLM-as-RS as a promising path toward foundation models for GR.