Diffusion large language models (dLLMs) are a new paradigm of non-autoregressive language models that are trained to predict multiple tokens in parallel and generate text via iterative unmasking. Recent works have successfully pretrained dLLMs to parity with autoregressive LLMs at the 8B scale, but dLLMs have yet to benefit from modern post-training techniques, e.g. reinforcement learning (RL), that have proven effective for autoregressive models. Crucially, algorithms designed for traditional LLMs aren't directly compatible with diffusion frameworks due to inherent differences in modeling assumptions. Moreover, existing attempts at dLLM post-training with RL rely on heuristic-based objectives with no theoretical grounding. In this work, we present Amortized Group Relative Policy Optimization (AGRPO), a principled on-policy RL algorithm designed specifically for dLLMs. AGRPO uses Monte Carlo sampling to compute an unbiased policy gradient estimate, making it the first tractable, faithful adaptation of policy gradient methods for dLLMs. We demonstrate AGRPO's effectiveness on different math/reasoning tasks, a common setting for RL with LLMs, achieving up to +7.6% absolute gain on GSM8K and 3.8x performance on the Countdown task over the baseline LLaDA-8B-Instruct model and 1.3x performance gains over comparable RL methods such as diffu-GRPO. Furthermore, these gains persist across different numbers of sampling steps at inference time, achieving better tradeoffs between compute and performance. Our results demonstrate that online RL algorithms can be extended to diffusion LLMs in principled ways, maintaining both theoretical soundness and practical effectiveness.

Foundation models for weather science are pre-trained on vast amounts of structured numerical data and outperform traditional weather forecasting systems. However, these models lack language-based reasoning capabilities, limiting their utility in interactive scientific workflows. Large language models (LLMs) excel at understanding and generating text but cannot reason about high-dimensional meteorological datasets. We bridge this gap by building a novel agentic framework for weather science. Our framework includes a Python code-based environment for agents (ZephyrusWorld) to interact with weather data, featuring tools like an interface to WeatherBench 2 dataset, geoquerying for geographical masks from natural language, weather forecasting, and climate simulation capabilities. We design Zephyrus, a multi-turn LLM-based weather agent that iteratively analyzes weather datasets, observes results, and refines its approach through conversational feedback loops. We accompany the agent with a new benchmark, ZephyrusBench, with a scalable data generation pipeline that constructs diverse question-answer pairs across weather-related tasks, from basic lookups to advanced forecasting, extreme event detection, and counterfactual reasoning. Experiments on this benchmark demonstrate the strong performance of Zephyrus agents over text-only baselines, outperforming them by up to 35 percentage points in correctness. However, on harder tasks, Zephyrus performs similarly to text-only baselines, highlighting the challenging nature of our benchmark and suggesting promising directions for future work.

Fluid voice-to-voice interaction requires reliable and low-latency detection of when a user has finished speaking. Traditional audio-silence end-pointers add hundreds of milliseconds of delay and fail under hesitations or language-specific phenomena. We present, to our knowledge, the first systematic study of Thai text-only end-of-turn (EOT) detection for real-time agents. We compare zero-shot and few-shot prompting of compact LLMs to supervised fine-tuning of lightweight transformers. Using transcribed subtitles from the YODAS corpus and Thai-specific linguistic cues (e.g., sentence-final particles), we formulate EOT as a binary decision over token boundaries. We report a clear accuracy-latency tradeoff and provide a public-ready implementation plan. This work establishes a Thai baseline and demonstrates that small, fine-tuned models can deliver near-instant EOT decisions suitable for on-device agents.

Although large language models (LLMs) have tremendous utility, trustworthiness is still a chief concern: models often generate incorrect information with high confidence. While contextual information can help guide generation, identifying when a query would benefit from retrieved context and assessing the effectiveness of that context remains challenging. In this work, we operationalize interpretability methods to ascertain whether we can predict the correctness of model outputs from the model's activations alone. We also explore whether model internals contain signals about the efficacy of external context. We consider correct, incorrect, and irrelevant context and introduce metrics to distinguish amongst them. Experiments on six different models reveal that a simple classifier trained on intermediate layer activations of the first output token can predict output correctness with about 75% accuracy, enabling early auditing. Our model-internals-based metric significantly outperforms prompting baselines at distinguishing between correct and incorrect context, guarding against inaccuracies introduced by polluted context. These findings offer a lens to better understand the underlying decision-making processes of LLMs. Our code is publicly available at https://github.com/jiarui-liu/LLM-Microscope

People often struggle to remember specific details of past experiences, which can lead to the need to revisit these memories. Consequently, lifelog retrieval has emerged as a crucial application. Various studies have explored methods to facilitate rapid access to personal lifelogs for memory recall assistance. In this paper, we propose a Captioning-Integrated Visual Lifelog (CIVIL) Retrieval System for extracting specific images from a user's visual lifelog based on textual queries. Unlike traditional embedding-based methods, our system first generates captions for visual lifelogs and then utilizes a text embedding model to project both the captions and user queries into a shared vector space. Visual lifelogs, captured through wearable cameras, provide a first-person viewpoint, necessitating the interpretation of the activities of the individual behind the camera rather than merely describing the scene. To address this, we introduce three distinct approaches: the single caption method, the collective caption method, and the merged caption method, each designed to interpret the life experiences of lifeloggers. Experimental results show that our method effectively describes first-person visual images, enhancing the outcomes of lifelog retrieval. Furthermore, we construct a textual dataset that converts visual lifelogs into captions, thereby reconstructing personal life experiences.

The meteoric rise of foundation models (FMs) has expanded their capabilities far beyond conventional tasks. Creativity, long regarded as a hallmark of human intelligence and a driver of innovation, is now increasingly recognized as a critical dimension of machine intelligence in the era of generative FMs, complementing traditional measures of accuracy. However, existing evaluation frameworks for creativity remain fragmented, relying on ad hoc metrics not firmly grounded in established theories. To address this gap, we introduce C^2-Eval, a holistic benchmark for unified assessment of creativity in FMs. C^2-Eval distinguishes between two complementary forms of creativity: convergent creativity, where tasks admit constrained solutions (e.g., code generation), and divergent creativity, where tasks are open-ended (e.g., storytelling). It evaluates both dimensions using fine-grained criteria derived from social-science theory, focusing on Usefulness, Originality, and Surprise (U-O-S). Through extensive experiments on leading proprietary and open-source models, we analyze trade-offs in their creative capabilities. Our results highlight both the strengths and challenges of current FMs in pursuing a creative machine mind, showing that C^2-Eval is an effective lens for examining the evolving landscape of creative AI.

Large language models (LLMs) are increasingly deployed in agentic systems where they map user intents to relevant external tools to fulfill a task. A critical step in this process is tool selection, where a retriever first surfaces candidate tools from a larger pool, after which the LLM selects the most appropriate one. This pipeline presents an underexplored attack surface where errors in selection can lead to severe outcomes like unauthorized data access or denial of service, all without modifying the agent's model or code. While existing evaluations measure task performance in benign settings, they overlook the specific vulnerabilities of the tool selection mechanism under adversarial conditions. To address this gap, we introduce ToolCert, the first statistical framework that formally certifies tool selection robustness. ToolCert models tool selection as a Bernoulli success process and evaluates it against a strong, adaptive attacker who introduces adversarial tools with misleading metadata, and are iteratively refined based on the agent's previous choices. By sampling these adversarial interactions, ToolCert produces a high-confidence lower bound on accuracy, formally quantifying the agent's worst-case performance. Our evaluation with ToolCert uncovers the severe fragility: under attacks injecting deceptive tools or saturating retrieval, the certified accuracy bound drops near zero, an average performance drop of over 60% compared to non-adversarial settings. For attacks targeting the retrieval and selection stages, the certified accuracy bound plummets to less than 20% after just a single round of adversarial adaptation. ToolCert thus reveals previously unexamined security threats inherent to tool selection and provides a principled method to quantify an agent's robustness to such threats, a necessary step for the safe deployment of agentic systems.

The Transformer architecture has revolutionized the field of sequence modeling and underpins the recent breakthroughs in large language models (LLMs). However, a comprehensive mathematical theory that explains its structure and operations remains elusive. In this work, we propose a novel continuous framework that rigorously interprets the Transformer as a discretization of a structured integro-differential equation. Within this formulation, the self-attention mechanism emerges naturally as a non-local integral operator, and layer normalization is characterized as a projection to a time-dependent constraint. This operator-theoretic and variational perspective offers a unified and interpretable foundation for understanding the architecture's core components, including attention, feedforward layers, and normalization. Our approach extends beyond previous theoretical analyses by embedding the entire Transformer operation in continuous domains for both token indices and feature dimensions. This leads to a principled and flexible framework that not only deepens theoretical insight but also offers new directions for architecture design, analysis, and control-based interpretations. This new interpretation provides a step toward bridging the gap between deep learning architectures and continuous mathematical modeling, and contributes a foundational perspective to the ongoing development of interpretable and theoretically grounded neural network models.

Distilling long reasoning traces (10K+ tokens) from stronger teacher models into smaller student LLMs via SFT has emerged as a standard paradigm. This approach is practical and efficient: it leverages the ease of generating abundant reasoning data from stronger models and provides a direct, data-driven way to teach less capable models better reasoning. While previous work has largely focused on prompt selection with responses from a single teacher, the equally important problem of choosing the best response when multiple teacher outputs are available for a single prompt remains underexplored. This challenge becomes important in a multi-teacher setting, where different students may benefit from the outputs of different teachers. This paper fills that gap with a systematic study of response selection for reasoning distillation. We first show that the current method, which picks responses the student assigns the highest global log-probability (global naturalness), fails when responses come from multiple teachers, i.e., global naturalness no longer correlates with downstream performance, especially as the reasoning traces from strong teachers become longer. To overcome this problem, we introduce Local Naturalness, which measures the student's log-probabilities over short, sequential reasoning steps conditioned only on a small local window. Local Naturalness enables two applications: 1) Teacher Selection: Aggregating local scores across prompts reliably identifies the most helpful teacher. 2) Response Selection from a Multiple Teachers: When mixing answers from many teachers, Local Naturalness boosts a 32B student's accuracy on math benchmarks by 9.4pp over global selection, also surpassing the performance achieved by training on data from the single best teacher. These results highlight the power of localized data quality evaluation and data mixing for more effective reasoning distillation.

Embedding vision-language models (VLMs) are typically pretrained with short text windows (<77 tokens), which forces the truncation of long-format captions. Yet, the distribution of biomedical captions from large-scale open source literature reveals that a huge portion of captions far exceed 77 tokens. To this end, we investigate the impact of pretraining on long-format biomedical captions by extending the context length of text encoders in VLMs. We find that longer context (thus, enabling additional supervision provided in long-format captions) correlates with better retrieval and classification performance. Given this finding, we introduce BIOMEDICA-LongCAP, a dataset of 1M image-caption pairs enriched with context-aware descriptions from full-text articles, providing longer and additional textual supervision. Using BIOMEDICA-LongCAP, we train BMC-LongCLIP, a long-context biomedical VLM with a text encoder supporting windows of up to 512 tokens. Our model extends context capacity by 6.6x, reducing token waste from 55% to just 2.2%. On long-caption retrieval benchmarks, BMC-LongCLIP achieves up to +30% absolute gains in Recall@1 and +2% average improvements in classification, while also converging faster than short-context. Our results demonstrate that long-context modeling is a promising direction for advancing biomedical VLMs.