Large language model agents are increasingly used to automate web tasks such as product search, offer comparison, and checkout. Current research explores different interfaces through which these agents interact with websites, including traditional HTML browsing, retrieval-augmented generation (RAG) over pre-crawled content, communication via Web APIs using the Model Context Protocol (MCP), and natural-language querying through the NLWeb interface. However, no prior work has compared these four architectures within a single controlled environment using identical tasks. To address this gap, we introduce a testbed consisting of four simulated e-shops, each offering its products via HTML, MCP, and NLWeb interfaces. For each interface (HTML, RAG, MCP, and NLWeb) we develop specialized agents that perform the same sets of tasks, ranging from simple product searches and price comparisons to complex queries for complementary or substitute products and checkout processes. We evaluate the agents using GPT 4.1, GPT 5, GPT 5 mini, and Claude Sonnet 4 as underlying LLM. Our evaluation shows that the RAG, MCP and NLWeb agents outperform HTML on both effectiveness and efficiency. Averaged over all tasks, F1 rises from 0.67 for HTML to between 0.75 and 0.77 for the other agents. Token usage falls from about 241k for HTML to between 47k and 140k per task. The runtime per task drops from 291 seconds to between 50 and 62 seconds. The best overall configuration is RAG with GPT 5 achieving an F1 score of 0.87 and a completion rate of 0.79. Also taking cost into consideration, RAG with GPT 5 mini offers a good compromise between API usage fees and performance. Our experiments show the choice of the interaction interface has a substantial impact on both the effectiveness and efficiency of LLM-based web agents.

Effective surveillance of hand, foot and mouth disease (HFMD) requires forecasts accounting for epidemiological patterns and contextual drivers like school calendars and weather. While classical models and recent foundation models (e.g., Chronos, TimesFM) incorporate covariates, they often lack the semantic reasoning to interpret the causal interplay between conflicting drivers. In this work, we propose a two-agent framework decoupling contextual interpretation from probabilistic forecasting. An LLM "event interpreter" processes heterogeneous signals-including school schedules, meteorological summaries, and reports-into a scalar transmission-impact signal. A neuro-symbolic core then combines this with historical case counts to produce calibrated probabilistic forecasts. We evaluate the framework on real-world HFMD datasets from Hong Kong (2023-2024) and Lishui, China (2024). Compared to traditional and foundation-model baselines, our approach achieves competitive point forecasting accuracy while providing robust 90% prediction intervals (coverage 0.85-1.00) and human-interpretable rationales. Our results suggest that structurally integrating domain knowledge through LLMs can match state-of-the-art performance while yielding context-aware forecasts that align with public health workflows. Code is available at https://github.com/jw-chae/forecast_MED .

Carefully engineered system prompts play a critical role in guiding the behavior of LLM agents, but their considerable length introduces significant drawbacks, including increased inference latency, higher computational cost, and reduced effective context length. This raises the question of whether such lengthy prompts can be replaced by a drastically reduced number of tokens while preserving their behavioral effect on downstream tasks. To enable this, we propose a lightweight three-stage training framework that learns a single prompt-specific Behavior-Equivalent token ([BE]). The framework first trains [BE] to encode the natural-language content of the original system prompt via reconstruction, and then distills the prompt 's downstream behavior into this single token. Importantly, our method requires no access to model internals, no auxiliary compression models, and no labeled responses. Empirical evaluations on three datasets show that a single [BE] token achieves up to a 3000x reduction in prompt length, while retaining about 98% of the downstream performance of the original system prompts. This substantially reduces inference cost and leaves almost the entire context window available for user inputs.

High-quality and carefully curated data is a cornerstone of training medical large language models, as it directly impacts both generalization and robustness to unseen clinical tasks. We investigate strategies for training and data curation to develop a robust multimodal reasoning model in the medical domain. Our work focuses on supervised fine-tuning (SFT) and explores data recipes that leverage structured reasoning traces. Using our proposed data recipe, we scale experiments to a dataset of over 8 million examples and 6.8 billion response tokens, achieving state-of-the-art performance among open-source models across diverse out-of-distribution medical benchmark tasks. Our results further indicate that curating a high-quality, diverse training dataset with varying structured reasoning trace lengths enables the fine-tuned model to self-calibrate its reasoning trajectory lengths based on the downstream task, without explicit supervision. We present key insights, describe the data curation strategy, and outline next steps toward developing robust medical vision-language reasoning system.

Recent Vision-Language Models (VLMs) exhibit strong perceptual reasoning abilities, yet they often struggle to adapt efficiently when encountering novel tasks at test time. In contrast, humans leverage the metacognitive model with memory, enabling continuous strategy refinement through metacognitive control when faced with new challenges. T o bridge this gap, we propose metacognitive test-time reasoning (MCTR), a framework that equips models with the ability to learn, adapt, and improve during test time through metacognitive self-updating. Inspired by the dual structure of human metacognition, MCTR comprises meta-level and object-level VLM reasoning modules, each equipped with dedicated memory systems for hierarchical adaptive reasoning. Specifically, MCTR consists of (1) a meta-reasoning module which incrementally builds a structured memory by discovering and storing task-relevant rules, environmental patterns, and action-outcome relationships from test-time observations as natural language descriptions; and (2) an action-reasoning module that determines optimal actions through context-aware perception and strategic reasoning by dynamically retrieving and integrating knowledge from memory. The action-reasoning module continuously updates its policy through proposed metacognitive test-time reinforcement learning, adapting as knowledge memory evolves. W e evaluate MCTR on 45 Atari games (33 seen, 12 unseen). MCTR demonstrates robust test-time adaptation, achieving 9/12 top-1 results on unseen games compared with baselines. Analyses through ablations, learning dynamics, and case studies reveal the complementary contributions of both components and show meta-reasoning evolving toward human-like adaptation strategies.

Recent advancements in Vision-Language Models (VLMs) have significantly transformed various industries. In agriculture, these dual-modal capabilities offer promising applications such as precision farming, crop monitoring, pest detection, and environmental sustainability. While several Visual Question Answering (VQA) datasets and benchmarks have been developed to evaluate VLM performance, they often fail to adequately assess the critical reasoning and problem-solving skills required in complex agricultural contexts. T o address this gap, we introduce AgriCoT, a VQA dataset that incorporates Chain-of-Thought (CoT) reasoning, specifically designed to evaluate the reasoning capabilities of VLMs. With 4,535 carefully curated samples, AgriCoT offers a comprehensive and robust evaluation of reasoning abilities for VLMs, particularly in zero-shot scenarios, by focusing on their capacity to engage in logical reasoning and effective problem-solving. Our evaluations, conducted with 26 representative VLMs, including both proprietary and open-source models, reveal that while some proprietary models excel at answering questions, there is a notable and significant gap in their reasoning capabilities. This underscores the importance of incorporating CoT for more precise and effective assessments.

This article presents the first comprehensive study on designing a baseline extractive question-answering (QA) system for the Hindi tourism domain, with a specialized focus on the Varanasi-a cultural and spiritual hub renowned for its Bhakti-Bhaav (devotional ethos). Targeting ten tourism-centric subdomains-Ganga Aarti, Cruise, Food Court, Public Toilet, Kund, Museum, General, Ashram, Temple and Travel, the work addresses the absence of language-specific QA resources in Hindi for culturally nuanced applications. In this paper, a dataset comprising 7,715 Hindi QA pairs pertaining to Varanasi tourism was constructed and subsequently augmented with 27,455 pairs generated via Llama zero-shot prompting. We propose a framework leveraging foundation models-BERT and RoBERTa, fine-tuned using Supervised Fine-Tuning (SFT) and Low-Rank Adaptation (LoRA), to optimize parameter efficiency and task performance. Multiple variants of BERT, including pre-trained languages (e.g., Hindi-BERT), are evaluated to assess their suitability for low-resource domain-specific QA. Evaluation metrics - F1, BLEU, and ROUGE-L - highlight trade-offs between answer precision and linguistic fluency. Experiments demonstrate that LoRA-based fine-tuning achieves competitive performance (85.3\% F1) while reducing trainable parameters by 98\% compared to SFT, striking a balance between efficiency and accuracy. Comparative analysis across models reveals that RoBERTa with SFT outperforms BERT variants in capturing contextual nuances, particularly for culturally embedded terms (e.g., Aarti, Kund). This work establishes a foundational baseline for Hindi tourism QA systems, emphasizing the role of LORA in low-resource settings and underscoring the need for culturally contextualized NLP frameworks in the tourism domain.

Native FP8 support in modern hardware is essential for training large Transformers, but is severely hindered by extreme activation outliers. Existing solutions either rely on complex mixed-precision engineering or invasive architectural modifications. This paper fundamentally challenges the conventional wisdom that outliers are data-driven. We demonstrate that extreme outliers are a data-independent, mechanically-produced artifact of training, originating from specific structural properties of the weight matrices (i.e., colinearity). Based on this insight, we propose TWEO (Transformers Without Extreme Outliers), a novel, non-invasive loss function. TWEO effectively prevents extreme outliers via a very simple loss term, which reduces outliers from 10000+ to less than 20. TWEO then enables full-model FP8 pre-training with neither engineering tricks nor architectural changes for both LLM and ViT. When standard FP8 training catastrophically collapses, TWEO achieves performance comparable to the BF16 baseline while delivering a 36% increase in training throughput. Also, TWEO enables a new quantization paradigm. Hardware-friendly W8A8 per-tensor static quantization of LLMs, previously considered completely unusable due to outliers, achieves SOTA performance for the first time on TWEO-trained models.

Aspect sentiment quad prediction (ASQP) is inherently challenging to predict a structured quadruple with four core sentiment elements, including aspect term (a), aspect category (c), opinion term (o), and sentiment polarity (s). Prior methods relying on marker-based prediction struggle with modeling the intricate relationships among elements and experience sharp performance declines when predicting higher-order elements (e.g., c and s) under standard supervised fine-tuning. To address these limitations, we employ reasoning-based generation to output both the quadruple and a natural language rationale under element prefixes within a unified template, encouraging explicit relational reasoning and interpretability. To further enhance element-wise alignment, we introduce a listwise preference optimization framework for improving structural validity and relational coherence. Specifically, we generate element-wise confusable candidates via syntactic and semantic proximity, then train the model with listwise objectives to prefer the gold candidates over closely competing alternatives. Extensive experiments on four benchmark datasets demonstrate that our framework effectively improves quadruple prediction accuracy and explanation consistency.

Large Language Models (LLMs) are trained to refuse to respond to harmful content. However, systematic analyses of whether this behavior is truly a reflection of its safety policies or an indication of political censorship, that is practiced globally by countries, is lacking. Differentiating between safety influenced refusals or politically motivated censorship is hard and unclear. For this purpose we introduce PSP, a dataset built specifically to probe the refusal behaviors in LLMs from an explicitly political context. PSP is built by formatting existing censored content from two data sources, openly available on the internet: sensitive prompts in China generalized to multiple countries, and tweets that have been censored in various countries. We study: 1) impact of political sensitivity in seven LLMs through data-driven (making PSP implicit) and representation-level approaches (erasing the concept of politics); and, 2) vulnerability of models on PSP through prompt injection attacks (PIAs). Associating censorship with refusals on content with masked implicit intent, we find that most LLMs perform some form of censorship. We conclude with summarizing major attributes that can cause a shift in refusal distributions across models and contexts of different countries.