Millions of users rely on a market of cloud-based services to obtain access to state-of-the-art large language models. However, it has been very recently shown that the de facto pay-per-token pricing mechanism used by providers creates a financial incentive for them to strategize and misreport the (number of) tokens a model used to generate an output. In this paper, we develop an auditing framework based on martingale theory that enables a trusted third-party auditor who sequentially queries a provider to detect token misreporting. Crucially, we show that our framework is guaranteed to always detect token misreporting, regardless of the provider's (mis-)reporting policy, and not falsely flag a faithful provider as unfaithful with high probability. To validate our auditing framework, we conduct experiments across a wide range of (mis-)reporting policies using several large language models from the $\texttt{Llama}$, $\texttt{Gemma}$ and $\texttt{Ministral}$ families, and input prompts from a popular crowdsourced benchmarking platform. The results show that our framework detects an unfaithful provider after observing fewer than $\sim 70$ reported outputs, while maintaining the probability of falsely flagging a faithful provider below $α= 0.05$.

KV cache in autoregressive LLMs eliminates redundant recomputation but has emerged as the dominant memory and bandwidth bottleneck during inference, notably with long contexts and test-time scaling. KV quantization is a key lever for reducing cache cost, but accuracy drops sharply as the native KV distribution lacks flatness and thus maintains a wide quantization range. Prior work focuses on isolating outliers, which caps their error but fails to flatten the overall distribution, leaving performance fragile under low-bit settings. In this work, we show that the K cache maintains a stable structure that evolves gradually with context, while the V cache carries latent semantic regularities. Building on these insights, we propose PatternKV, a pattern-aligned residual quantization scheme. It mines representative pattern vectors online, aligns each KV vector to its nearest pattern, and quantizes only the residual. This reshaping of the KV distribution flattens the quantization target and narrows its range, thereby improving the fidelity of low-bit KV quantization. Across long-context and test-time scaling settings on multiple backbones, PatternKV delivers consistent 2-bit gains, with a 0.08% average 4-bit drop relative to FP16, improves test-time scaling accuracy by 10% on average, and raises throughput by 1.4x while supporting 1.25x larger batches.

Large language models (LLMs) demonstrate remarkable performance across diverse tasks, yet their effectiveness frequently depends on costly commercial APIs or cloud services. Model selection thus entails a critical trade-off between performance and cost: high-performing LLMs typically incur substantial expenses, whereas budget-friendly small language models (SLMs) are constrained by limited capabilities. Current research primarily proposes two routing strategies: pre-generation routing and cascade routing. Both approaches have distinct characteristics, with cascade routing typically offering superior cost-effectiveness and accuracy despite its higher latency. To further address the limitations of both approaches, we introduce SATER, a dual-mode compatible approach that fine-tunes models through shortest-response preference optimization and a confidence-aware rejection mechanism. SATER significantly reduces redundant outputs and response times, while improving both the performance of pre-generation routing and the efficiency of cascade routing. Experiments across three SLMs and six datasets, varying in type and complexity, demonstrate that SATER achieves comparable performance while consistently reducing computational costs by over 50\% and cascade latency by over 80\%.

Physics-informed neural networks (PINNs)provide a powerful approach for solving partial differential equations (PDEs), but constructing a usable PINN remains labor-intensive and error-prone. Scientists must interpret problems as PDE formulations, design architectures and loss functions, and implement stable training pipelines. Existing large language model (LLM)approaches address isolated steps such as code generation or architecture suggestion, but typically assume a formal PDE is already specified and therefore lack an end-to-end perspective. We present Lang-PINN, an LLM-driven multi-agent system that builds trainable PINNs directly from natural language task descriptions. Lang-PINN coordinates four complementary agents: a PDE Agent that parses task descriptions into symbolic PDEs, a PINN Agent that selects architectures, a Code Agent that generates modular implementations, and a Feedback Agent that executes and diagnoses errors for iterative refinement. This design transforms informal task statements into executable and verifiable PINN code. Experiments show that Lang-PINN achieves substantially lower errors and greater robustness than competitive baselines: mean squared error (MSE)is reduced by up to 3-5 orders of magnitude, end-to-end execution success improves by more than 50%, and reduces time overhead by up to 74%. Partial differential equations (PDEs)are central to scientific computing, underpinning applications in physics, engineering, and materials science. Physics-informed neural networks (PINNs)(Raissi et al., 2019) have emerged as a flexible framework that embeds governing equations into trainable neural models, offering a unified approach for forward, inverse, and data-scarce problems (Karni-adakis et al., 2021; Lu et al., 2021). Although libraries and benchmarks such as DeepXDE (Lu et al., 2021), PINNacle (Hao et al., 2023), and PDEBench (Takamoto et al., 2022) have been developed, deploying a trainable PINN still requires expert-level manual effort in PDE specification, architecture design, and optimization tuning.

The deployment of autonomous AI agents in sensitive domains, such as healthcare, introduces critical risks to safety, security, and privacy. These agents may deviate from user objectives, violate data handling policies, or be compromised by adversarial attacks. Mitigating these dangers necessitates a mechanism to formally guarantee that an agent's actions adhere to predefined safety constraints, a challenge that existing systems do not fully address. We introduce VeriGuard, a novel framework that provides formal safety guarantees for LLM-based agents through a dual-stage architecture designed for robust and verifiable correctness. The initial offline stage involves a comprehensive validation process. It begins by clarifying user intent to establish precise safety specifications. VeriGuard then synthesizes a behavioral policy and subjects it to both testing and formal verification to prove its compliance with these specifications. This iterative process refines the policy until it is deemed correct. Subsequently, the second stage provides online action monitoring, where VeriGuard operates as a runtime monitor to validate each proposed agent action against the pre-verified policy before execution. This separation of the exhaustive offline validation from the lightweight online monitoring allows formal guarantees to be practically applied, providing a robust safeguard that substantially improves the trustworthiness of LLM agents.

Common Large Language model (LLM) evaluations rely on demonstration examples to steer models' responses to the desired style. While the number of examples used has been studied and standardized, the choice of how to format examples is less investigated. In evaluation protocols and real world usage, users face the choice how to separate in-context examples: use a comma? new line? semi-colon? hashtag? etc.? Surprisingly, we find this seemingly minor choice can dramatically alter model response quality. Across leading model families (Llama, Qwen, Gemma), performance on MMLU for example can vary by $\pm 23\%$ depending on the choice of delimiter. In fact, one can manipulate model rankings to put any model in the lead by only modifying the single character separating examples. We find LLMs' brittleness pervades topics, model families, and doesn't improve with scale. By probing attention head scores, we find that good-performing delimiters steer attention towards key tokens in the input. Finally, we explore methods to improve LLMs' robustness to the choice of delimiter. We find specifying the selected delimiter in the prompt boosts robustness and offer practical recommendations for the best-performing delimiters to select.

The analysis centers on FinMA, a model created within the PIXIU framework, which is evaluated for its performance in specialized financial tasks. Recognizing the critical demands of accuracy, reliability, and domain adaptation in financial applications, this study examines FinMA's model architecture, its instruction tuning process utilizing the Financial Instruction Tuning (FIT) dataset, and its evaluation under the FLARE benchmark. Findings indicate that FinMA performs well in sentiment analysis and classification, but faces notable challenges in tasks involving numerical reasoning, entity recognition, and summarization. This work aims to advance the understanding of how financial LLMs can be effectively designed and evaluated to assist in finance-related decision-making processes. Keywords: Large Language Models (LLMs); Financial NLP; FinLLMs; FinMA; FLARE Benchmark; FIT Dataset; Sentiment Analysis; Financial Question Answering; Stock Movement Prediction; Named Entity Recognition; Financial Text Summarization; Instruction Tuning; Financial Reasoning; Domain Adaptation.

Discrete Diffusion Large Language Models (dLLMs) have recently emerged as a competitive paradigm for non-autoregressive language modeling. Their distinctive decoding mechanism enables faster inference speed and strong performance in code generation and mathematical tasks. In this work, we show that the decoding mechanism of dLLMs not only enhances model utility but also can be used as a powerful tool for model attribution. A key challenge in this problem lies in the diversity of attribution scenarios, including distinguishing between different models as well as between different checkpoints or backups of the same model. To ensure broad applicability, we identify two fundamental problems: what information to extract from the decoding trajectory, and how to utilize it effectively. We first observe that relying directly on per-step model confidence yields poor performance. This is mainly due to the bidirectional decoding nature of dLLMs: each newly decoded token influences the confidence of other decoded tokens, making model confidence highly redundant and washing out structural signal regarding decoding order or dependencies. To overcome this, we propose a novel information extraction scheme called the Directed Decoding Map (DDM), which captures structural relationships between decoding steps and better reveals model-specific behaviors. Furthermore, to make full use of the extracted structural information during attribution, we propose Gaussian-Trajectory Attribution (GTA), where we fit a cell-wise Gaussian distribution at each decoding position for each target model, and define the likelihood of a trajectory as the attribution score: if a trajectory exhibits higher log-likelihood under the distribution of a specific model, it is more likely to have been generated by that model. Extensive experiments under different settings validate the utility of our methods.

Deep research agents, which synthesize information across diverse sources, are significantly constrained by their sequential reasoning processes. To overcome this, we introduce FlashResearch, a novel framework for efficient deep research that transforms sequential processing into parallel, runtime orchestration by dynamically decomposing complex queries into tree-structured sub-tasks. Our core contributions are threefold: (1) an adaptive planner that dynamically allocates computational resources by determining research breadth and depth based on query complexity; (2) a real-time orchestration layer that monitors research progress and prunes redundant paths to reallocate resources and optimize efficiency; and (3) a multi-dimensional parallelization framework that enables concurrency across both research breadth and depth. Experiments show that FlashResearch consistently improves final report quality within fixed time budgets, and can deliver up to a 5 speedup while maintaining comparable quality. Deep research tasks, which involve synthesizing information from diverse sources and navigating complex, interdependent concepts, pose significant challenges for existing AI systems. These tasks often demand knowledge retrieval, advanced reasoning, sophisticated tool use, and dynamic planning over multiple steps under structural uncertainty and evolving objectives (Du et al., 2025). Applications include literature review (Haman & ˇ Skoln ık, 2025), open-domain question answering, and policy analysis (Gambrell, 2025), where the ability to evaluate conflicting perspectives, explore hypotheses, and revise beliefs as new evidence emerges is essential.

Data-driven materials discovery requires large-scale experimental datasets, yet most of the information remains trapped in unstructured literature. Existing extraction efforts often focus on a limited set of features and have not addressed the integrated composition-processing-microstructure-property relationships essential for understanding materials behavior, thereby posing challenges for building comprehensive databases. To address this gap, we propose a multi-stage information extraction pipeline powered by large language models, which captures 47 features spanning composition, processing, microstructure, and properties exclusively from experimentally reported materials. The pipeline integrates iterative extraction with source tracking to enhance both accuracy and reliability. Evaluations at the feature level (independent attributes) and tuple level (interdependent features) yielded F1 scores around 0.96. Compared with single-pass extraction without source tracking, our approach improved F1 scores of microstructure category by 10.0% (feature level) and 13.7% (tuple level), and reduced missed materials from 49 to 13 out of 396 materials in 100 articles on precipitate-containing multi-principal element alloys (miss rate reduced from 12.4% to 3.3%). The pipeline enables scalable and efficient literature mining, producing databases with high precision, minimal omissions, and zero false positives. These datasets provide trustworthy inputs for machine learning and materials informatics, while the modular design generalizes to diverse material classes, enabling comprehensive materials information extraction.