Precise surgical interventions are vital to patient safety, and advanced enhancement algorithms have been developed to assist surgeons in decision-making. Despite significant progress, these algorithms are typically designed for single tasks in specific scenarios, limiting their effectiveness in complex real-world situations. To address this limitation, we propose SurgVisAgent, an end-to-end intelligent surgical vision agent built on multimodal large language models (MLLMs). SurgVis-Agent dynamically identifies distortion categories and severity levels in endoscopic images, enabling it to perform a variety of enhancement tasks such as low-light enhancement, overexposure correction, motion blur elimination, and smoke removal. Specifically, to achieve superior surgical scenario understanding, we design a prior model that provides domain-specific knowledge. Additionally, through in-context few-shot learning and chain-of-thought (CoT) reasoning, SurgVisAgent delivers customized image enhancements tailored to a wide range of distortion types and severity levels, thereby addressing the diverse requirements of surgeons. Furthermore, we construct a comprehensive benchmark simulating real-world surgical distortions, on which extensive experiments demonstrate that SurgVisAgent surpasses traditional single-task models, highlighting its potential as a unified solution for surgical assistance.

In the current machine learning landscape, we face a "model lake" phenomenon: Given a task, there is a proliferation of trained models with similar performances despite different behavior. For model users attempting to navigate and select from the models, documentation comparing model pairs is helpful. However, for every $N$ models there could be $O(N^2)$ pairwise comparisons, a number prohibitive for the model developers to manually perform pairwise comparisons and prepare documentations. To facilitate fine-grained pairwise comparisons among models, we introduced $\textbf{VERBA}$. Our approach leverages a large language model (LLM) to generate verbalizations of model differences by sampling from the two models. We established a protocol that evaluates the informativeness of the verbalizations via simulation. We also assembled a suite with a diverse set of commonly used machine learning models as a benchmark. For a pair of decision tree models with up to 5% performance difference but 20-25% behavioral differences, $\textbf{VERBA}$ effectively verbalizes their variations with up to 80% overall accuracy. When we included the models' structural information, the verbalization's accuracy further improved to 90%. $\textbf{VERBA}$ opens up new research avenues for improving the transparency and comparability of machine learning models in a post-hoc manner.

While recent advances in preference learning have enhanced alignment in human feedback, mathematical reasoning remains a persistent challenge. We investigate how data diversification strategies in preference optimization can improve the mathematical reasoning abilities of large language models (LLMs). We evaluate three common data generation methods: temperature sampling, Chain-of-Thought prompting, and Monte Carlo Tree Search (MCTS), and introduce Diversified-ThinkSolve (DTS), a novel structured approach that systematically decomposes problems into diverse reasoning paths. Our results show that with strategically diversified preference data, models can substantially improve mathematical reasoning performance, with the best approach yielding gains of 7.1% on GSM8K and 4.2% on MATH over the base model. Despite its strong performance, DTS incurs only a marginal computational overhead (1.03x) compared to the baseline, while MCTS is nearly five times more costly with lower returns. These findings demonstrate that structured exploration of diverse problem-solving methods creates more effective preference data for mathematical alignment than traditional approaches.

Dialogue summarization is a critical natural language processing task that supports numerous practical applications, such as customer service, meeting analysis, and conversational AI assistants. Unlike traditional document summarization, dialogue summarization must handle unique challenges, including multi-party interactions, fragmented utterances, ambiguous references, and frequent topic shifts. Additionally, effective summarization can facilitate automated meeting documentation, collaborative decision-making, and efficient information retrieval from dialogue records. Early advances relied primarily on extractive methods that selected key sentences based on simple heuristics like TF-IDF or word frequency (Marcu, 1997), before evolving to neural approaches such as Seq2Seq and Pointer-Generator networks, which enabled more fluent abstractive summaries (Rush et al., 2015; See et al., 2017). Subsequently, significant breakthroughs were achieved by adapting Transformer-based neural architectures to conversational settings (Lewis et al., 2019; Liang et al., 2022; Jin et al., 2025). Large language models (LLMs) have achieved remarkable results across a wide variety of natural language processing tasks, including text classification, sentiment analysis, question answering, and translation, demonstrating strong generalization capabilities and state-of-the-art performance (Brown et al., 2020). In particular, reasoning LLMs, such as OpenAI-o1, DeepSeek-R1, and QwQ-32B, have exhibited notable advantages in tasks requiring complex reasoning, such as mathematical problem solving, logical inference, and machine translation (Chen et al., 2025a; Ye et al., 2025). These successes naturally prompt further exploration into their applicability within dialogue summarization. Dialogue summarization encompasses multiple distinct paradigms, each reflecting real-world scenarios that vary significantly in language, domain, dialogue length, and user intent.

Large language models (LLMs) are increasingly used to assign document relevance labels in information retrieval pipelines, especially in domains lacking human-labeled data. However, different models often disagree on borderline cases, raising concerns about how such disagreement affects downstream retrieval. This study examines labeling disagreement between two open-weight LLMs, LLaMA and Qwen, on a corpus of scholarly abstracts related to Sustainable Development Goals (SDGs) 1, 3, and 7. We isolate disagreement subsets and examine their lexical properties, rank-order behavior, and classification predictability. Our results show that model disagreement is systematic, not random: disagreement cases exhibit consistent lexical patterns, produce divergent top-ranked outputs under shared scoring functions, and are distinguishable with AUCs above 0.74 using simple classifiers. These findings suggest that LLM-based filtering introduces structured variability in document retrieval, even under controlled prompting and shared ranking logic. We propose using classification disagreement as an object of analysis in retrieval evaluation, particularly in policy-relevant or thematic search tasks.

Modern AI models, such as large language models, are usually trained once on a huge corpus of data, potentially fine-tuned for a specific task, and then deployed with fixed parameters. Their training is costly, slow, and gradual, requiring billions of repetitions. In stark contrast, animals continuously adapt to the ever-changing contingencies in their environments. This is particularly important for social species, where behavioral policies and reward outcomes may frequently change in interaction with peers. The underlying computational processes are often marked by rapid shifts in an animal's behaviour and rather sudden transitions in neuronal population activity. Such computational capacities are of growing importance for AI systems operating in the real world, like those guiding robots or autonomous vehicles, or for agentic AI interacting with humans online. Can AI learn from neuroscience? This Perspective explores this question, integrating the literature on continual and in-context learning in AI with the neuroscience of learning on behavioral tasks with shifting rules, reward probabilities, or outcomes. We will outline an agenda for how specifically insights from neuroscience may inform current developments in AI in this area, and - vice versa - what neuroscience may learn from AI, contributing to the evolving field of NeuroAI.

Large language models (LLMs) have rapidly progressed into general-purpose agents capable of solving a broad spectrum of tasks. However, current models remain inefficient at reasoning: they apply fixed inference-time compute regardless of task complexity, often overthinking simple problems while underthinking hard ones. This survey presents a comprehensive review of efficient test-time compute (TTC) strategies, which aim to improve the computational efficiency of LLM reasoning. We introduce a two-tiered taxonomy that distinguishes between L1-controllability, methods that operate under fixed compute budgets, and L2-adaptiveness, methods that dynamically scale inference based on input difficulty or model confidence. We benchmark leading proprietary LLMs across diverse datasets, highlighting critical trade-offs between reasoning performance and token usage. Compared to prior surveys on efficient reasoning, our review emphasizes the practical control, adaptability, and scalability of TTC methods. Finally, we discuss emerging trends such as hybrid thinking models and identify key challenges for future work towards making LLMs more computationally efficient, robust, and responsive to user constraints.

Modern biomedical research is defined by both immense opportunity and staggering complexity. As a cornerstone of science, it generates vast quantities of data from large-scale experiments, but this progress is hampered by a research landscape that is profoundly fragmented (1-3). The knowledge, specialized software, and databases required to make discoveries are numerous, constantly evolving, and dispersed, forcing researchers to expend significant time and effort on the manual and labor-intensive task of discovering, learning, and integrating these disparate resources. While the advent of AI agents holds the promise of automating this intricate work (4-6), current systems inherit a critical limitation: they typically rely on manually curated, static toolsets (7-14). This approach is inefficient, fails to scale, and cannot keep pace with the rapid evolution of biomedical science, leaving the agents perpetually behind the cutting edge. This raises a critical question: Can we design a self-evolving agent that transcends these limitations by automatically discovering and integrating new tools, continuously updating its knowledge base, and iteratively upgrading its own capabilities through direct experience? Here we present STELLA, a generalist biomedical AI agent designed around the core principle of self-evolution (15). STELLA learns and improves from every problem it solves, continuously enhancing its own reasoning strategies and technical abilities.

Despite recent advances in Vision-Language Models (VLMs), long-video understanding remains a challenging problem. Although state-of-the-art long-context VLMs can process around 1000 input frames, they still struggle to effectively leverage this sequence length, and succumb to irrelevant distractors within the context window. We present Temporal Chain of Thought, an inference strategy for video question-answering that curates the model's input context. We use the VLM itself to iteratively identify and extract the most relevant frames from the video, which are then used for answering. We demonstrate how leveraging more computation at inference-time to select the most relevant context leads to improvements in accuracy, in agreement with recent work on inference-time scaling of LLMs. Moreover, we achieve state-of-the-art results on 4 diverse video question-answering datasets, showing consistent improvements with 3 different VLMs. In particular, our method shines on longer videos which would not otherwise fit within the model's context window: On longer videos of more than 1 hour on LVBench, our approach using a context window of 32K outperforms the same VLM using standard inference with a 700K context window by 2.8 points.

The proliferation of artificial intelligence (AI) in financial services has prompted growing demand for tools that can systematically detect AI-related disclosures in corporate filings. While prior approaches often rely on keyword expansion or document-level classification, they fall short in granularity, interpretability, and robustness. This study introduces FinAI-BERT, a domain-adapted transformer-based language model designed to classify AI-related content at the sentence level within financial texts. The model was fine-tuned on a manually curated and balanced dataset of 1,586 sentences drawn from 669 annual reports of U.S. banks (2015 to 2023). FinAI-BERT achieved near-perfect classification performance (accuracy of 99.37 percent, F1 score of 0.993), outperforming traditional baselines such as Logistic Regression, Naive Bayes, Random Forest, and XGBoost. Interpretability was ensured through SHAP-based token attribution, while bias analysis and robustness checks confirmed the model's stability across sentence lengths, adversarial inputs, and temporal samples. Theoretically, the study advances financial NLP by operationalizing fine-grained, theme-specific classification using transformer architectures. Practically, it offers a scalable, transparent solution for analysts, regulators, and scholars seeking to monitor the diffusion and framing of AI across financial institutions.