Test-time compute scaling has demonstrated the ability to improve the performance of reasoning language models by generating longer chain-of-thought (CoT) sequences. However, this increase in performance comes with a significant increase in computational cost. In this work, we investigate two compute constraint strategies: (1) reasoning length constraint and (2) model quantization, as methods to reduce the compute demand of reasoning models and study their impact on their safety performance. Specifically, we explore two approaches to apply compute constraints to reasoning models: (1) fine-tuning reasoning models using a length controlled policy optimization (LCPO) based reinforcement learning method to satisfy a user-defined CoT reasoning length, and (2) applying quantization to maximize the generation of CoT sequences within a user-defined compute constraint. Furthermore, we study the trade-off between the computational efficiency and the safety of the model.

Most automatic speech processing systems operate in "open loop" mode without user feedback about who said what; yet, human-in-the-loop workflows can potentially enable higher accuracy. We propose an LLM-assisted speaker diarization correction system that lets users fix speaker attribution errors in real time. The pipeline performs streaming ASR and diarization, uses an LLM to deliver concise summaries to the users, and accepts brief verbal feedback that is immediately incorporated without disrupting interactions. Moreover, we develop techniques to make the workflow more effective: First, a split-when-merged (SWM) technique detects and splits multi-speaker segments that the ASR erroneously attributes to just a single speaker. Second, online speaker enrollments are collected based on users' diarization corrections, thus helping to prevent speaker diarization errors from occurring in the future. LLM-driven simulations on the AMI test set indicate that our system substantially reduces DER by 9.92% and speaker confusion error by 44.23%. We further analyze correction efficacy under different settings, including summary vs full transcript display, the number of online enrollments limitation, and correction frequency.

Grounding vision--language models in low-resource languages remains challenging, as they often produce fluent text about the wrong objects. This stems from scarce paired data, translation pivots that break alignment, and English-centric pretraining that ignores target-language semantics. We address this with a compute-aware Bengali captioning pipeline trained on LaBSE-verified EN--BN pairs and 110k bilingual-prompted synthetic images. A frozen MaxViT yields stable visual patches, a Bengali-native mBART-50 decodes, and a lightweight bridge links the modalities. Our core novelty is a tri-loss objective: Patch-Alignment Loss (PAL) aligns real and synthetic patch descriptors using decoder cross-attention, InfoNCE enforces global real--synthetic separation, and Sinkhorn-based OT ensures balanced fine-grained patch correspondence. This PAL+InfoNCE+OT synergy improves grounding, reduces spurious matches, and drives strong gains on Flickr30k-1k (BLEU-4 12.29, METEOR 27.98, BERTScore-F1 71.20) and MSCOCO-1k (BLEU-4 12.00, METEOR 28.14, BERTScore-F1 75.40), outperforming strong CE baselines and narrowing the real--synthetic centroid gap by 41%.

As large language models (LLMs) become increasingly powerful, the sequential nature of autoregressive generation creates a fundamental throughput bottleneck that limits the practical deployment. While Multi-Token Prediction (MTP) has demonstrated remarkable benefits for model training efficiency and performance, its inherent potential for inference acceleration remains largely unexplored. This paper introduces FastMTP, a simple yet effective method that improves multi-step draft quality by aligning MTP training with its inference pattern, significantly enhancing speculative decoding performance. Our approach fine-tunes a single MTP head with position-shared weights on self-distilled data, enabling it to capture dependencies among consecutive future tokens and maintain high acceptance rates across multiple recursive draft steps. By integrating language-aware dynamic vocabulary compression into the MTP head, we further reduce computational overhead in the drafting process. Experimental results across seven diverse benchmarks demonstrate that FastMTP achieves an average of 2.03x speedup compared to standard next token prediction with lossless output quality, outperforming vanilla MTP by 82%. FastMTP requires only lightweight training and seamlessly integrates with existing inference frameworks, offering a practical and rapidly deployable solution for accelerating LLM inference.

Evaluating developer satisfaction with conversational AI assistants at scale is critical but challenging. User studies provide rich insights, but are unscalable, while large-scale quantitative signals from logs or in-product ratings are often too shallow or sparse to be reliable. To address this gap, we propose and evaluate a new approach: using sentiment analysis of developer prompts to identify implicit signals of user satisfaction. With an analysis of industrial usage logs of 372 professional developers, we show that this approach can identify a signal in ~8% of all interactions, a rate more than 13 times higher than explicit user feedback, with reasonable accuracy even with an off-the-shelf sentiment analysis approach. This new practical approach to complement existing feedback channels would open up new directions for building a more comprehensive understanding of the developer experience at scale.

Large language models (LLMs) show promise for diagnostic reasoning but often lack reliable, knowledge grounded inference. Knowledge graphs (KGs), such as the Unified Medical Language System (UMLS), offer structured biomedical knowledge that can support trustworthy reasoning. Prior approaches typically integrate KGs via retrieval augmented generation or fine tuning, inserting KG content into prompts rather than enabling structured reasoning. We explore an alternative paradigm: treating the LLM as a reward model of KG reasoning paths, where the model learns to judge whether a candidate path leads to correct diagnosis for a given patient input. This approach is inspired by recent work that leverages reward training to enhance model reasoning abilities, and grounded in computational theory, which suggests that verifying a solution is often easier than generating one from scratch. It also parallels physicians' diagnostic assessment, where they judge which sequences of findings and intermediate conditions most plausibly support a diagnosis. We first systematically evaluate five task formulation for knowledge path judging and eight training paradigm. Second, we test whether the path judging abilities generalize to downstream diagnostic tasks, including diagnosis summarization and medical question answering. Experiments with three open source instruct-tuned LLMs reveal both promise and brittleness: while specific reward optimization and distillation lead to strong path-judging performance, the transferability to downstream tasks remain weak. Our finding provides the first systematic assessment of "reward model style" reasoning over clinical KGs, offering insights into how structured, reward-based supervision influences diagnostic reasoning in GenAI systems for healthcare.

Robotic manipulation policies often fail to generalize because they must simultaneously learn where to attend, what actions to take, and how to execute them. We argue that high-level reasoning about where and what can be offloaded to vision-language models (VLMs), leaving policies to specialize in how to act. We present PEEK (Policy-agnostic Extraction of Essential Keypoints), which fine-tunes VLMs to predict a unified point-based intermediate representation: 1. end-effector paths specifying what actions to take, and 2. task-relevant masks indicating where to focus. These annotations are directly overlaid onto robot observations, making the representation policy-agnostic and transferable across architectures. To enable scalable training, we introduce an automatic annotation pipeline, generating labeled data across 20+ robot datasets spanning 9 embodiments. In real-world evaluations, PEEK consistently boosts zero-shot generalization, including a 41.4x real-world improvement for a 3D policy trained only in simulation, and 2-3.5x gains for both large VLAs and small manipulation policies. By letting VLMs absorb semantic and visual complexity, PEEK equips manipulation policies with the minimal cues they need--where, what, and how. Website at https://peek-robot.github.io/.

Controlling desktop applications via software remains a fundamental yet under-served problem. Existing multi-modal large language models (MLLMs) ingest screenshots and task instructions to generate keystrokes and mouse events, but they suffer from prohibitive inference latency, poor sample efficiency on long-horizon sparse-reward tasks, and infeasible on-device deployment. We introduce a lightweight hierarchical reinforcement learning framework, ComputerAgent, that formulates OS control as a two-level option process (manager and subpolicy), employs a triple-modal state encoder (screenshot, task ID, numeric state) to handle visual and contextual diversity, integrates meta-actions with an early-stop mechanism to reduce wasted interactions, and uses a compact vision backbone plus small policy networks for on-device inference (15M parameters). On a suite of 135 real-world desktop tasks, ComputerAgent attains 92.1% success on simple tasks (<8 steps) and 58.8% on hard tasks (>=8 steps), matching or exceeding 200B-parameter MLLM baselines on simple scenarios while reducing model size by over four orders of magnitude and halving inference time. These results demonstrate that hierarchical RL offers a practical, scalable alternative to monolithic MLLM-based automation for computer control.

Personalized recipe recommendation faces challenges in handling fuzzy user intent, ensuring semantic accuracy, and providing sufficient detail coverage. We propose ChefMind, a hybrid architecture combining Chain of Exploration (CoE), Knowledge Graph (KG), Retrieval-Augmented Generation (RAG), and a Large Language Model (LLM). CoE refines ambiguous queries into structured conditions, KG offers semantic reasoning and interpretability, RAG supplements contextual culinary details, and LLM integrates outputs into coherent recommendations. We evaluate ChefMind on the Xiachufang dataset and manually annotated queries, comparing it with LLM-only, KG-only, and RAG-only baselines. Results show that ChefMind achieves superior performance in accuracy, relevance, completeness, and clarity, with an average score of 8.7 versus 6.4-6.7 for ablation models. Moreover, it reduces unprocessed queries to 1.6%, demonstrating robustness in handling fuzzy demands.

With the rising global burden of chronic diseases and the multimodal and heterogeneous clinical data (medical imaging, free-text recordings, wearable sensor streams, etc.), there is an urgent need for a unified multimodal AI framework that can proactively predict individual health risks. We propose VL-RiskFormer, a hierarchical stacked visual-language multimodal Transformer with a large language model (LLM) inference head embedded in its top layer. The system builds on the dual-stream architecture of existing visual-linguistic models (e.g., PaLM-E, LLaVA) with four key innovations: (i) pre-training with cross-modal comparison and fine-grained alignment of radiological images, fundus maps, and wearable device photos with corresponding clinical narratives using momentum update encoders and debiased InfoNCE losses; (ii) a time fusion block that integrates irregular visit sequences into the causal Transformer decoder through adaptive time interval position coding; (iii) a disease ontology map adapter that injects ICD-10 codes into visual and textual channels in layers and infers comorbid patterns with the help of a graph attention mechanism. On the MIMIC-IV longitudinal cohort, VL-RiskFormer achieved an average AUROC of 0.90 with an expected calibration error of 2.7 percent.