Faster and Better LLMs via Latency-Aware Test-Time Scaling

Test-Time Scaling (TTS) has proven effective in improving the performance of Large Language Models (LLMs) during inference. However, existing research has overlooked the efficiency of TTS from a latency-sensitive perspective. Through a latency-aware evaluation of representative TTS methods, we demonstrate that a compute-optimal TTS does not always result in the lowest latency in scenarios where latency is critical. To address this gap and achieve latency-optimal TTS, we propose two key approaches by optimizing the concurrency configurations: (1) branch-wise parallelism, which leverages multiple concurrent inference branches, and (2) sequence-wise parallelism, enabled by speculative decoding. By integrating these two approaches and allocating computational resources properly to each, our latency-optimal TTS enables a 32B model to reach 82.3% accuracy on MATH-500 within 1 minute and a smaller 3B model to achieve 72.4% within 10 seconds. Our work emphasizes the importance of latency-aware TTS and demonstrates its ability to deliver both speed and accuracy in latency-sensitive scenarios.

MoESD: Unveil Speculative Decoding's Potential for Accelerating Sparse MoE

Large Language Models (LLMs) have achieved remarkable success across many applications, with Mixture of Experts (MoE) models demonstrating great potential. Compared to traditional dense models, MoEs achieve better performance with less computation. Speculative decoding (SD) is a widely used technique to accelerate LLM inference without accuracy loss, but it has been considered efficient only for dense models. In this work, we first demonstrate that, under medium batch sizes, MoE surprisingly benefits more from SD than dense models. Furthermore, as MoE becomes sparser -- the prevailing trend in MoE designs -- the batch size range where SD acceleration is expected to be effective becomes broader. To quantitatively understand tradeoffs involved in SD, we develop a reliable modeling based on theoretical analyses. While current SD research primarily focuses on improving acceptance rates of algorithms, changes in workload and model architecture can still lead to degraded SD acceleration even with high acceptance rates. To address this limitation, we introduce a new metric 'target efficiency' that characterizes these effects, thus helping researchers identify system bottlenecks and understand SD acceleration more comprehensively. For scenarios like private serving, this work unveils a new perspective to speed up MoE inference, where existing solutions struggle. Experiments on different GPUs show up to 2.29x speedup for Qwen2-57B-A14B at medium batch sizes and validate our theoretical predictions.

STRICT: Stress Test of Rendering Images Containing Text

While diffusion models have revolutionized text-to-image generation with their ability to synthesize realistic and diverse scenes, they continue to struggle to generate consistent and legible text within images. This shortcoming is commonly attributed to the locality bias inherent in diffusion-based generation, which limits their ability to model long-range spatial dependencies. In this paper, we introduce $\textbf{STRICT}$, a benchmark designed to systematically stress-test the ability of diffusion models to render coherent and instruction-aligned text in images. Our benchmark evaluates models across multiple dimensions: (1) the maximum length of readable text that can be generated; (2) the correctness and legibility of the generated text, and (3) the ratio of not following instructions for generating text. We evaluate several state-of-the-art models, including proprietary and open-source variants, and reveal persistent limitations in long-range consistency and instruction-following capabilities. Our findings provide insights into architectural bottlenecks and motivate future research directions in multimodal generative modeling. We release our entire evaluation pipeline at https://github.com/tianyu-z/STRICT-Bench.

FairZK: A Scalable System to Prove Machine Learning Fairness in Zero-Knowledge

With the rise of machine learning techniques, ensuring the fairness of decisions made by machine learning algorithms has become of great importance in critical applications. However, measuring fairness often requires full access to the model parameters, which compromises the confidentiality of the models. In this paper, we propose a solution using zero-knowledge proofs, which allows the model owner to convince the public that a machine learning model is fair while preserving the secrecy of the model. To circumvent the efficiency barrier of naively proving machine learning inferences in zero-knowledge, our key innovation is a new approach to measure fairness only with model parameters and some aggregated information of the input, but not on any specific dataset. To achieve this goal, we derive new bounds for the fairness of logistic regression and deep neural network models that are tighter and better reflecting the fairness compared to prior work. Moreover, we develop efficient zero-knowledge proof protocols for common computations involved in measuring fairness, including the spectral norm of matrices, maximum, absolute value, and fixed-point arithmetic. We have fully implemented our system, FairZK, that proves machine learning fairness in zero-knowledge. Experimental results show that FairZK is significantly faster than the naive approach and an existing scheme that use zero-knowledge inferences as a subroutine. The prover time is improved by 3.1x--1789x depending on the size of the model and the dataset. FairZK can scale to a large model with 47 million parameters for the first time, and generates a proof for its fairness in 343 seconds. This is estimated to be 4 orders of magnitude faster than existing schemes, which only scale to small models with hundreds to thousands of parameters.

A Scalable System to Prove Machine Learning Fairness in Zero-Knowledge

With the rise of machine learning techniques, ensuring the fairness of decisions made by machine learning algorithms has become of great importance in critical applications. However, measuring fairness often requires full access to the model parameters, which compromises the confidentiality of the models. In this paper, we propose a solution using zero-knowledge proofs, which allows the model owner to convince the public that a machine learning model is fair while preserving the secrecy of the model. To circumvent the efficiency barrier of naively proving machine learning inferences in zero-knowledge, our key innovation is a new approach to measure fairness only with model parameters and some aggregated information of the input, but not on any specific dataset. To achieve this goal, we derive new bounds for the fairness of logistic regression and deep neural network models that are tighter and better reflecting the fairness compared to prior work. Moreover, we develop efficient zero-knowledge proof protocols for common computations involved in measuring fairness, including the spectral norm of matrices, maximum, absolute value, and fixed-point arithmetic. We have fully implemented our system, FairZK, that proves machine learning fairness in zero-knowledge. Experimental results show that FairZK is significantly faster than the naive approach and an existing scheme that use zero-knowledge inferences as a subroutine. The prover time is improved by 3.1x--1789x depending on the size of the model and the dataset. FairZK can scale to a large model with 47 million parameters for the first time, and generates a proof for its fairness in 343 seconds. This is estimated to be 4 orders of magnitude faster than existing schemes, which only scale to small models with hundreds to thousands of parameters.

Neural Co-Optimization of Structural Topology, Manufacturable Layers, and Path Orientations for Fiber-Reinforced Composites

We propose a neural network-based computational framework for the simultaneous optimization of structural topology, curved layers, and path orientations to achieve strong anisotropic strength in fiber-reinforced thermoplastic composites while ensuring manufacturability. Our framework employs three implicit neural fields to represent geometric shape, layer sequence, and fiber orientation. This enables the direct formulation of both design and manufacturability objectives - such as anisotropic strength, structural volume, machine motion control, layer curvature, and layer thickness - into an integrated and differentiable optimization process. By incorporating these objectives as loss functions, the framework ensures that the resultant composites exhibit optimized mechanical strength while remaining its manufacturability for filament-based multi-axis 3D printing across diverse hardware platforms. Physical experiments demonstrate that the composites generated by our co-optimization method can achieve an improvement of up to 33.1% in failure loads compared to composites with sequentially optimized structures and manufacturing sequences.

Lessons from Deploying Learning-based CSI Localization on a Large-Scale ISAC Platform

In recent years, Channel State Information (CSI), recognized for its fine-grained spatial characteristics, has attracted increasing attention in WiFi-based indoor localization. However, despite its potential, CSI-based approaches have yet to achieve the same level of deployment scale and commercialization as those based on Received Signal Strength Indicator (RSSI). A key limitation lies in the fact that most existing CSI-based systems are developed and evaluated in controlled, small-scale environments, limiting their generalizability. To bridge this gap, we explore the deployment of a large-scale CSI-based localization system involving over 400 Access Points (APs) in a real-world building under the Integrated Sensing and Communication (ISAC) paradigm. We highlight two critical yet often overlooked factors: the underutilization of unlabeled data and the inherent heterogeneity of CSI measurements. To address these challenges, we propose a novel CSI-based learning framework for WiFi localization, tailored for large-scale ISAC deployments on the server side. Specifically, we employ a novel graph-based structure to model heterogeneous CSI data and reduce redundancy. We further design a pretext pretraining task that incorporates spatial and temporal priors to effectively leverage large-scale unlabeled CSI data. Complementarily, we introduce a confidence-aware fine-tuning strategy to enhance the robustness of localization results. In a leave-one-smartphone-out experiment spanning five floors and 25, 600 m2, we achieve a median localization error of 2.17 meters and a floor accuracy of 99.49%. This performance corresponds to an 18.7% reduction in mean absolute error (MAE) compared to the best-performing baseline.

PHYBench: Holistic Evaluation of Physical Perception and Reasoning in Large Language Models

We introduce PHYBench, a novel, high-quality benchmark designed for evaluating reasoning capabilities of large language models (LLMs) in physical contexts. PHYBench consists of 500 meticulously curated physics problems based on real-world physical scenarios, designed to assess the ability of models to understand and reason about realistic physical processes. Covering mechanics, electromagnetism, thermodynamics, optics, modern physics, and advanced physics, the benchmark spans difficulty levels from high school exercises to undergraduate problems and Physics Olympiad challenges. Additionally, we propose the Expression Edit Distance (EED) Score, a novel evaluation metric based on the edit distance between mathematical expressions, which effectively captures differences in model reasoning processes and results beyond traditional binary scoring methods. We evaluate various LLMs on PHYBench and compare their performance with human experts. Our results reveal that even state-of-the-art reasoning models significantly lag behind human experts, highlighting their limitations and the need for improvement in complex physical reasoning scenarios. Our benchmark results and dataset are publicly available at https://phybench-official.github.io/phybench-demo/.

Aligning Constraint Generation with Design Intent in Parametric CAD

We adapt alignment techniques from reasoning LLMs to the task of generating engineering sketch constraints found in computer-aided design (CAD) models. Engineering sketches consist of geometric primitives (e.g. points, lines) connected by constraints (e.g. perpendicular, tangent) that define the relationships between them. For a design to be easily editable, the constraints must effectively capture design intent, ensuring the geometry updates predictably when parameters change. Although current approaches can generate CAD designs, an open challenge remains to align model outputs with design intent, we label this problem `design alignment'. A critical first step towards aligning generative CAD models is to generate constraints which fully-constrain all geometric primitives, without over-constraining or distorting sketch geometry. Using alignment techniques to train an existing constraint generation model with feedback from a constraint solver, we are able to fully-constrain 93% of sketches compared to 34% when using a na\"ive supervised fine-tuning (SFT) baseline and only 8.9% without alignment. Our approach can be applied to any existing constraint generation model and sets the stage for further research bridging alignment strategies between the language and design domains.

V2Flow: Unifying Visual Tokenization and Large Language Model Vocabularies for Autoregressive Image Generation

We propose V2Flow, a novel tokenizer that produces discrete visual tokens capable of high-fidelity reconstruction, while ensuring structural and latent distribution alignment with the vocabulary space of large language models (LLMs). Leveraging this tight visual-vocabulary coupling, V2Flow enables autoregressive visual generation on top of existing LLMs. Our approach formulates visual tokenization as a flow-matching problem, aiming to learn a mapping from a standard normal prior to the continuous image distribution, conditioned on token sequences embedded within the LLMs vocabulary space. The effectiveness of V2Flow stems from two core designs. First, we propose a Visual Vocabulary resampler, which compresses visual data into compact token sequences, with each represented as a soft categorical distribution over LLM's vocabulary. This allows seamless integration of visual tokens into existing LLMs for autoregressive visual generation. Second, we present a masked autoregressive Rectified-Flow decoder, employing a masked transformer encoder-decoder to refine visual tokens into contextually enriched embeddings. These embeddings then condition a dedicated velocity field for precise reconstruction. Additionally, an autoregressive rectified-flow sampling strategy is incorporated, ensuring flexible sequence lengths while preserving competitive reconstruction quality. Extensive experiments show that V2Flow outperforms mainstream VQ-based tokenizers and facilitates autoregressive visual generation on top of existing. https://github.com/zhangguiwei610/V2Flow