Low-latency decoding for large language models (LLMs) is crucial for applications like chatbots and code assistants, yet generating long outputs remains slow in single-query settings. Prior work on speculative decoding (which combines a small draft model with a larger target model) and tensor parallelism has each accelerated decoding. However, conventional approaches fail to apply both simultaneously due to imbalanced compute requirements (between draft and target models), KV-cache inconsistencies, and communication overheads under small-batch tensor-parallelism. This paper introduces SwiftSpec, a system that targets ultra-low latency for LLM decoding. SwiftSpec redesigns the speculative decoding pipeline in an asynchronous and disaggregated manner, so that each component can be scaled flexibly and remove draft overhead from the critical path. To realize this design, SwiftSpec proposes parallel tree generation, tree-aware KV cache management, and fused, latency-optimized kernels to overcome the challenges listed above. Across 5 model families and 6 datasets, SwiftSpec achieves an average of 1.75x speedup over state-of-the-art speculative decoding systems and, as a highlight, serves Llama3-70B at 348 tokens/s on 8 Nvidia Hopper GPUs, making it the fastest known system for low-latency LLM serving at this scale.

Test-time scaling has emerged as an effective approach for improving language model performance by utilizing additional compute at inference time. Recent studies have shown that overriding end-of-thinking tokens (e.g., replacing "" with "Wait") can extend reasoning steps and improve accuracy. In this work, we explore whether a dedicated continue-thinking token can be learned to trigger extended reasoning. We augment a distilled version of DeepSeek-R1 with a single learned "<|continue-thinking|>" token, training only its embedding via reinforcement learning while keeping the model weights frozen. Our experiments show that this learned token achieves improved accuracy on standard math benchmarks compared to both the baseline model and a test-time scaling approach that uses a fixed token (e.g., "Wait") for budget forcing. In particular, we observe that in cases where the fixed-token approach enhances the base model's accuracy, our method achieves a markedly greater improvement. For example, on the GSM8K benchmark, the fixed-token approach yields a 1.3% absolute improvement in accuracy, whereas our learned-token method achieves a 4.2% improvement over the base model that does not use budget forcing.

Building energy management (BEM) tasks require processing and learning from a variety of time-series data. Existing solutions rely on bespoke task- and data-specific models to perform these tasks, limiting their broader applicability. Inspired by the transformative success of Large Language Models (LLMs), Time-Series Foundation Models (TSFMs), trained on diverse datasets, have the potential to change this. Were TSFMs to achieve a level of generalizability across tasks and contexts akin to LLMs, they could fundamentally address the scalability challenges pervasive in BEM. To understand where they stand today, we evaluate TSFMs across four dimensions: (1) generalizability in zero-shot univariate forecasting, (2) forecasting with covariates for thermal behavior modeling, (3) zero-shot representation learning for classification tasks, and (4) robustness to performance metrics and varying operational conditions. Our results reveal that TSFMs exhibit \emph{limited} generalizability, performing only marginally better than statistical models on unseen datasets and modalities for univariate forecasting. Similarly, inclusion of covariates in TSFMs does not yield performance improvements, and their performance remains inferior to conventional models that utilize covariates. While TSFMs generate effective zero-shot representations for downstream classification tasks, they may remain inferior to statistical models in forecasting when statistical models perform test-time fitting. Moreover, TSFMs forecasting performance is sensitive to evaluation metrics, and they struggle in more complex building environments compared to statistical models. These findings underscore the need for targeted advancements in TSFM design, particularly their handling of covariates and incorporating context and temporal dynamics into prediction mechanisms, to develop more adaptable and scalable solutions for BEM.

In an effort to automatically evaluate and select the best model and improve code quality for automatic incident remediation in IT Automation, it is crucial to verify if the generated code for remediation action is syntactically and semantically correct and whether it can be executed correctly as intended. There are three approaches: 1) conventional methods use surface form similarity metrics (token match, exact match, etc.) which have numerous limitations, 2) execution-based evaluation focuses more on code functionality based on pass/fail judgments for given test-cases, and 3) LLM-as-a-Judge employs LLMs for automated evaluation to judge if it is a correct answer for a given problem based on pre-defined metrics. In this work, we focused on enhancing LLM-as-a-Judge using bidirectional functionality matching and logic representation for reference-less automatic validation and refinement for Bash code generation to select the best model for automatic incident remediation in IT Automation. We used execution-based evaluation as ground-truth to evaluate our LLM-as-a-Judge metrics. Results show high accuracy and agreement with execution-based evaluation (and up to 8% over baseline). Finally, we built Reflection code agents to utilize judgments and feedback from our evaluation metrics which achieved significant improvement (up to 24% increase in accuracy) for automatic code refinement.

As a cornerstone of patient care, clinical decision-making significantly influences patient outcomes and can be enhanced by large language models (LLMs). Although LLMs have demonstrated remarkable performance, their application to visual question answering in medical imaging, particularly for reasoning-based diagnosis, remains largely unexplored. Furthermore, supervised fine-tuning for reasoning tasks is largely impractical due to limited data availability and high annotation costs. In this work, we introduce a zero-shot framework for reliable medical image diagnosis that enhances the reasoning capabilities of LLMs in clinical settings through test-time scaling. Given a medical image and a textual prompt, a vision-language model processes a medical image along with a corresponding textual prompt to generate multiple descriptions or interpretations of visual features. These interpretations are then fed to an LLM, where a test-time scaling strategy consolidates multiple candidate outputs into a reliable final diagnosis. We evaluate our approach across various medical imaging modalities -- including radiology, ophthalmology, and histopathology -- and demonstrate that the proposed test-time scaling strategy enhances diagnostic accuracy for both our and baseline methods. Additionally, we provide an empirical analysis showing that the proposed approach, which allows unbiased prompting in the first stage, improves the reliability of LLM-generated diagnoses and enhances classification accuracy.

Large language models (LLMs) show promise in healthcare, but hallucinations remain a major barrier to clinical use. We present CHECK, a continuous-learning framework that integrates structured clinical databases with a classifier grounded in information theory to detect both factual and reasoning-based hallucinations. Evaluated on 1500 questions from 100 pivotal clinical trials, CHECK reduced LLama3.3-70B-Instruct hallucination rates from 31% to 0.3% - making an open source model state of the art. Its classifier generalized across medical benchmarks, achieving AUCs of 0.95-0.96, including on the MedQA (USMLE) benchmark and HealthBench realistic multi-turn medical questioning. By leveraging hallucination probabilities to guide GPT-4o's refinement and judiciously escalate compute, CHECK boosted its USMLE passing rate by 5 percentage points, achieving a state-of-the-art 92.1%. By suppressing hallucinations below accepted clinical error thresholds, CHECK offers a scalable foundation for safe LLM deployment in medicine and other high-stakes domains.

--Large Language Models (LLMs), combined with T ext-to-Speech (TTS) and Automatic Speech Recognition (ASR), are increasingly used to automate voice phishing (vishing) scams. These systems are scalable and convincing, posing a significant security threat. We identify the ASR transcription step as the most vulnerable link in the scam pipeline and introduce ASRJam, a proactive defence framework that injects adversarial perturbations into the victim's audio to disrupt the attacker's ASR. This breaks the scam's feedback loop without affecting human callers, who can still understand the conversation. While prior adversarial audio techniques are often unpleasant and impractical for real-time use, we also propose EchoGuard, a novel jammer that leverages natural distortions, such as reverberation and echo, that are disruptive to ASR but tolerable to humans. T o evaluate EchoGuard's effectiveness and usability, we conducted a 39-person user study comparing it with three state-of-the-art attacks. Results show that EchoGuard achieved the highest overall utility, offering the best combination of ASR disruption and human listening experience. Large Language Models (LLMs) are now widely used across many applications, demonstrating impressive progress in understanding and generating natural language [1], [2], [3]. When combined with text-to-speech (TTS) and automatic speech recognition (ASR) technologies, LLMs enable powerful new capabilities such as automated customer service, outbound sales, cold calling, and advanced virtual assistants. However, as these systems become more realistic and lifelike, they also raise significant security concerns. LLMs have proven effective at generating phishing content that rivals human-written emails [4], [5], contributing to a 703% rise in credential phishing in 2024. The integration of LLMs with speech synthesis into real-time, automated scam agents is the inevitable conclusion [6]. V oice agents operate by chaining together a sequence of neural networks to handle calls in real time: (1) ASR transcribes the victim's speech into text, (2) An LLM generates an appropriate textual response, and (3) TTS synthesizes that response into natural-sounding audio. This pipeline enables scalable voice interactions that can convincingly impersonate trusted entities and extract sensitive information from victims as seen in Figure 1.

In spite of strong performance achieved by LLMs, the costs of their deployment are unaffordable. For the compression of LLMs, gradient-based pruning methods present promising effectiveness. However, in these methods, the gradient computation with one-hot labels ignore the potential predictions on other words, thus missing key information for generative capability of the original model. To address this issue, we introduce a self-distillation loss during the pruning phase (rather than post-training) to fully exploit the predictions of the original model, thereby obtaining more accurate gradient information for pruning. Moreover, we find that, compared to attention modules, the predictions of LLM are less sensitive to multilayer perceptron (MLP) modules, which take up more than $5 \times$ parameters (LLaMA3.2-1.2B). To this end, we focus on the pruning of MLP modules, to significantly compress LLM without obvious performance degradation. Experimental results on extensive zero-shot benchmarks demonstrate that our method significantly outperforms existing pruning methods. Furthermore, our method achieves very competitive performance among 1B-scale open source LLMs. The source code and trained weights are available at https://github.com/visresearch/SDMPrune.

Recent advances in large language models (LLMs) have demonstrated notable performance in medical licensing exams. However, comprehensive evaluation of LLMs across various healthcare roles, particularly in high-stakes clinical scenarios, remains a challenge. Existing benchmarks are typically text-based, English-centric, and focus primarily on medicines, which limits their ability to assess broader healthcare knowledge and multimodal reasoning. To address these gaps, we introduce KokushiMD-10, the first multimodal benchmark constructed from ten Japanese national healthcare licensing exams. This benchmark spans multiple fields, including Medicine, Dentistry, Nursing, Pharmacy, and allied health professions. It contains over 11588 real exam questions, incorporating clinical images and expert-annotated rationales to evaluate both textual and visual reasoning. We benchmark over 30 state-of-the-art LLMs, including GPT-4o, Claude 3.5, and Gemini, across both text and image-based settings. Despite promising results, no model consistently meets passing thresholds across domains, highlighting the ongoing challenges in medical AI. KokushiMD-10 provides a comprehensive and linguistically grounded resource for evaluating and advancing reasoning-centric medical AI across multilingual and multimodal clinical tasks.

Recent benchmarks have probed factual consistency and rhetorical robustness in Large Language Models (LLMs). However, a knowledge gap exists regarding how directional framing of factually true statements influences model agreement, a common scenario for LLM users. AssertBench addresses this by sampling evidence-supported facts from FEVEROUS, a fact verification dataset. For each (evidence-backed) fact, we construct two framing prompts: one where the user claims the statement is factually correct, and another where the user claims it is incorrect. We then record the model's agreement and reasoning. The desired outcome is that the model asserts itself, maintaining consistent truth evaluation across both framings, rather than switching its evaluation to agree with the user. AssertBench isolates framing-induced variability from the model's underlying factual knowledge by stratifying results based on the model's accuracy on the same claims when presented neutrally. In doing so, this benchmark aims to measure an LLM's ability to "stick to its guns" when presented with contradictory user assertions about the same fact. The complete source code is available at https://github.com/achowd32/assert-bench.