The rise of disaggregated AI GPUs has exposed a critical bottleneck in large-scale attention workloads: non-uniform memory access (NUMA). As multi-chiplet designs become the norm for scaling compute capabilities, memory latency and bandwidth vary sharply across compute regions, undermining the performance of traditional GPU kernel scheduling strategies that assume uniform memory access. We identify how these NUMA effects distort locality in multi-head attention (MHA) and present Swizzled Head-first Mapping, a spatially-aware scheduling strategy that aligns attention heads with GPU NUMA domains to exploit intra-chiplet cache reuse. On AMD's MI300X architecture, our method achieves up to 50% higher performance over state-of-the-art attention algorithms using conventional scheduling techniques and sustains consistently high L2 cache hit rates of 80-97%. These results demonstrate that NUMA-aware scheduling is now fundamental to achieving full efficiency on next-generation disaggregated GPUs, offering a path forward for scalable AI training and inference.

We propose Re-FORC, an adaptive reward prediction method that, given a context, enables prediction of the expected future rewards as a function of the number of future thinking tokens. Re-FORC trains a lightweight adapter on reasoning models, demonstrating improved prediction with longer reasoning and larger models. Re-FORC enables: 1) early stopping of unpromising reasoning chains, reducing compute by 26% while maintaining accuracy, 2) optimized model and thinking length selection that achieves 4% higher accuracy at equal compute and 55% less compute at equal accuracy compared to the largest model, 3) adaptive test-time scaling, which increases accuracy by 11% in high compute regime, and 7% in low compute regime. Re-FORC allows dynamic reasoning with length control via cost-per-token thresholds while estimating computation time upfront.

Flood insurance is an effective strategy for individuals to mitigate disaster-related losses. However, participation rates among at-risk populations in the United States remain strikingly low. This gap underscores the need to understand and model the behavioral mechanisms underlying insurance decisions. Large language models (LLMs) have recently exhibited human-like intelligence across wide-ranging tasks, offering promising tools for simulating human decision-making. This study constructs a benchmark dataset to capture insurance purchase probabilities across factors. Using this dataset, the capacity of LLMs is evaluated: while LLMs exhibit a qualitative understanding of factors, they fall short in estimating quantitative probabilities. To address this limitation, InsurAgent, an LLM-empowered agent comprising five modules including perception, retrieval, reasoning, action, and memory, is proposed. The retrieval module leverages retrieval-augmented generation (RAG) to ground decisions in empirical survey data, achieving accurate estimation of marginal and bivariate probabilities. The reasoning module leverages LLM common sense to extrapolate beyond survey data, capturing contextual information that is intractable for traditional models. The memory module supports the simulation of temporal decision evolutions, illustrated through a roller coaster life trajectory. Overall, InsurAgent provides a valuable tool for behavioral modeling and policy analysis.

Contrast-Consistent Search (CCS) is an unsupervised probing method able to test whether large language models represent binary features, such as sentence truth, in their internal activations. While CCS has shown promise, its two-term objective has been only partially understood. In this work, we revisit CCS with the aim of clarifying its mechanisms and extending its applicability. We argue that what should be optimized for, is relative contrast consistency. Building on this insight, we reformulate CCS as an eigenproblem, yielding closed-form solutions with interpretable eigenvalues and natural extensions to multiple variables. We evaluate these approaches across a range of datasets, finding that they recover similar performance to CCS, while avoiding problems around sensitivity to random initialization. Our results suggest that relativizing contrast consistency not only improves our understanding of CCS but also opens pathways for broader probing and mechanistic interpretability methods.

Watermarking has emerged as a promising technique to track AI-generated content and differentiate it from authentic human creations. While prior work extensively studies watermarking for autoregressive large language models (LLMs) and image diffusion models, none address discrete diffusion language models, which are becoming popular due to their high inference throughput. In this paper, we introduce the first watermarking method for discrete diffusion models by applying the distribution-preserving Gumbel-max trick at every diffusion step and seeding the randomness with the sequence index to enable reliable detection. We experimentally demonstrate that our scheme is reliably detectable on state-of-the-art diffusion language models and analytically prove that it is distortion-free with an exponentially decaying probability of false detection in the token sequence length.

Scientific experiment and manufacture rely on complex, multi-step procedures that demand continuous human expertise for precise execution and decision-making. Despite advances in machine learning and automation, conventional models remain confined to virtual domains, while real-world experiment and manufacture still rely on human supervision and expertise. This gap between machine intelligence and physical execution limits reproducibility, scalability, and accessibility across scientific and manufacture workflows. Here, we introduce human-AI co-embodied intelligence, a new form of physical AI that unites human users, agentic AI, and wearable hardware into an integrated system for real-world experiment and intelligent manufacture. In this paradigm, humans provide precise execution and control, while agentic AI contributes memory, contextual reasoning, adaptive planning, and real-time feedback. The wearable interface continuously captures the experimental and manufacture processes, facilitates seamless communication between humans and AI for corrective guidance and interpretable collaboration. As a demonstration, we present Agentic-Physical Experimentation (APEX) system, coupling agentic reasoning with physical execution through mixed-reality. APEX observes and interprets human actions, aligns them with standard operating procedures, provides 3D visual guidance, and analyzes every step. Implemented in a cleanroom for flexible electronics fabrication, APEX system achieves context-aware reasoning with accuracy exceeding general multimodal large language models, corrects errors in real time, and transfers expertise to beginners. These results establish a new class of agentic-physical-human intelligence that extends agentic reasoning beyond computation into the physical domain, transforming scientific research and manufacturing into autonomous, traceable, interpretable, and scalable processes.

Creation of large-scale databases for Visual Question Answering tasks pertaining to the text data in a scene (text-VQA) involves skilful human annotation, which is tedious and challenging. With the advent of foundation models that handle vision and language modalities, and with the maturity of OCR systems, it is the need of the hour to establish an end-to-end pipeline that can synthesize Question-Answer (QA) pairs based on scene-text from a given image. We propose a pipeline for automated synthesis for text-VQA dataset that can produce faithful QA pairs, and which scales up with the availability of scene text data. Our proposed method harnesses the capabilities of multiple models and algorithms involving OCR detection and recognition (text spotting), region of interest (ROI) detection, caption generation, and question generation. These components are streamlined into a cohesive pipeline to automate the synthesis and validation of QA pairs. To the best of our knowledge, this is the first pipeline proposed to automatically synthesize and validate a large-scale text-VQA dataset comprising around 72K QA pairs based on around 44K images.

Recent work has discovered that large language models can develop broadly misaligned behaviors after being fine-tuned on narrowly harmful datasets, a phenomenon known as emergent misalignment (EM). However, the fundamental mechanisms enabling such harmful generalization across disparate domains remain poorly understood. In this work, we adopt a geometric perspective to study EM and demonstrate that it exhibits a fundamental cross-task linear structure in how harmful behavior is encoded across different datasets. Specifically, we find a strong convergence in EM parameters across tasks, with the fine-tuned weight updates showing relatively high cosine similarities, as well as shared lower-dimensional subspaces as measured by their principal angles and projection overlaps. Furthermore, we also show functional equivalence via linear mode connectivity, wherein interpolated models across narrow misalignment tasks maintain coherent, broadly misaligned behavior. Our results indicate that EM arises from different narrow tasks discovering the same set of shared parameter directions, suggesting that harmful behaviors may be organized into specific, predictable regions of the weight landscape. By revealing this fundamental connection between parametric geometry and behavioral outcomes, we hope our work catalyzes further research on parameter space interpretability and weight-based interventions.

Speculative decoding accelerates LLMs by using a lightweight draft model to generate tokens autoregressively before verifying them in parallel with a larger target model. However, determining the optimal number of tokens to draft remains a key challenge limiting the approach's effectiveness. Dynamic speculative decoding aims to intelligently decide how many tokens to draft to achieve maximum speedups. Existing methods often rely on hand-tuned, sensitive thresholds (e.g., token entropy), which are costly to set and generalize poorly across models and domains. We propose TapOut, an online, training-free, plug-and-play algorithm for dynamic speculation policy selection using multi-armed bandits. Our approach employs a meta-algorithm that selects among multiple parameter-free dynamic speculation strategies based on past reward and exploration. We conduct extensive experiments across diverse model pairs and datasets, showing that TapOut achieves competitive or superior speedups compared to well-established dynamic speculation baselines without any hyperparameter tuning.

In recent years, querying semantic web data using SPARQL has remained challenging, especially for non-expert users, due to the language's complex syntax and the prerequisite of understanding intricate data structures. To address these challenges, we propose InteracSPARQL, an interactive SPARQL query generation and refinement system that leverages natural language explanations (NLEs) to enhance user comprehension and facilitate iterative query refinement. InteracSPARQL integrates LLMs with a rule-based approach to first produce structured explanations directly from SPARQL abstract syntax trees (ASTs), followed by LLM-based linguistic refinements. Users can interactively refine queries through direct feedback or LLM-driven self-refinement, enabling the correction of ambiguous or incorrect query components in real time. We evaluate InteracSPARQL on standard benchmarks, demonstrating significant improvements in query accuracy, explanation clarity, and overall user satisfaction compared to baseline approaches. Our experiments further highlight the effectiveness of combining rule-based methods with LLM-driven refinements to create more accessible and robust SPARQL interfaces.