Searcharxiv⌕ Search

arXiv subjects

Yu Huang

Publications and source records attributed to Yu Huang.

At least 19 recordsLinked to original sources
arXiv2026

Accelerating LLM Pretraining with Zeroth-Order Curvature on Idle Serving Compute

LLM pretraining is increasingly compute-intensive, while substantial inference serving compute remains idle during off-peak periods. However, such serving hardware is difficult to use for conventional first-order training due to its memory, networking and other system requirements. Zeroth-order (ZO) optimization offers a natural interface to this forward-only compute resource, but its gradient estimates break down in pretraining. Unlike fine-tuning, which can confine updates to a small parameter subspace, pretraining must update every parameter, and the alignment of a ZO gradient estimate with the true gradient shrinks as the number of updated parameters grows. We make a different observation: noisy ZO estimates can still provide useful curvature information, even when they are too inaccurate to serve as gradients. Based on this insight, we propose ZO-Curv, which leverages idle forward-only devices to estimate curvature and uses it to adapt and refine the update used during first-order training, thereby accelerating the process. We provide convergence guarantees that characterize the effect of finite-query curvature estimation. Experiments on LLMs from 125M to 1.5B parameters show that ZO-Curv consistently accelerates pretraining. At 770M parameters, ZO-Curv reaches AdamW's 20K-step validation loss in 3.9 hours instead of 7.5 hours, while achieving a lower final loss of 2.70 versus 2.84 under the same step budget. These results show that idle model serving compute can accelerate LLM pretraining when zeroth-order queries are used to estimate curvature rather than gradients directly.

cs.LG↗
arXiv2026

TLAPS-Bench: Formally Proving Correctness of Real-World System Specifications with TLA+ Proof System

TLAPS-Bench is a benchmark and execution harness for constructing machine-checked proofs for real-world system specifications in TLA+ with agentic AI. Proving correctness of complex system specifications requires discovering inductive invariants, reasoning about intricate state transitions, and decomposing proofs into tractable obligations. TLAPS-Bench provides a proof harness that supports iterative proof development atop the TLA+ Proof System (TLAPS), where agents are equipped with proof tools and libraries, with safeguards against reward hacking. Using the harness, we have employed AI agents to prove desired properties for important system specifications that were not proved before, spanning cache coherence, Byzantine consensus, and write-ahead logs. TLAPS-Bench contains 265 proof targets (each encodes a property to prove) from 95 TLA+ specifications of real-world systems. We evaluated frontier AI agents, represented by Codex with GPT models, on a subset of 70 proof targets across 12 specifications. The highest-scoring model, GPT-6.1 Sol (max), proved 31 (44.3%) proof targets. The results show challenges and opportunities in end-to-end proof development.

cs.AI↗
arXiv2026

A Neuroprosthetic SoC Featuring Rapid Local Sensory Feedback with On-Chip Unsupervised Multi-Class Spike Sorting

This letter presents a 65 nm neuroprosthetic system-on-chip (SoC) featuring a rapid feedback pathway from motor recording to sensory stimulation. To enable low-latency sensory mapping modulated by individual-neuron firing rates, the SoC integrates a mixed-signal unsupervised spike-sorting module with stimulation-artifact rejection, a programmable sliding-window firing-rate calculator, and a closed-loop controller for adapting stimulation parameters. The spike-sorting module combines event-driven analog feature extraction with Manhattan-distance-based digital clustering while consuming 1.6 $μ$W/channel. On the Quiroga Difficult-1 benchmark, it achieves 94.58% three-class accuracy, with only a 0.4% reduction when stimulation artifacts are introduced at 20% of the spike count. The closed-loop controller supports polynomial and logarithmic mapping functions and incorporates Mitchell's approximation for power saving. The SoC is fully characterized on the bench and in vivo.

eess.SP↗
arXiv2026

KnowMT-Bench: Benchmarking Knowledge-Intensive Long-Form Question Answering in Multi-Turn Dialogues

Multi-Turn Long-Form Question Answering (MT-LFQA) is a key application paradigm of Large Language Models (LLMs) in knowledge-intensive domains. However, existing benchmarks are limited to single-turn dialogue, while multi-turn dialogue benchmarks typically assess other orthogonal capabilities rather than knowledge-intensive factuality. To bridge this critical gap, we introduce \textbf{KnowMT-Bench}, a new benchmark designed to systematically evaluate MT-LFQA for LLMs across knowledge-intensive fields, including medicine, finance, and law. To faithfully assess the model's real-world performance, KnowMT-Bench employs a dynamic evaluation setting where models generate their own multi-turn dialogue histories given logically progressive question sequences. Our experiments on a diverse suite of LLMs show a clear degradation in both factual capability and information delivery efficiency in multi-turn dialogue. We further probe the underlying causes and investigate mitigation strategies spanning domain-specific finetuning, active context denoising, and RAG. The results show domain-dependent gains from finetuning, while active context denoising and RAG more consistently mitigate these issues, with RAG notably reversing this performance degradation. These findings underscore the importance of our benchmark for evaluating LLMs' conversational factual capabilities in real-world applications. Code and data are publicly available.\footnote{\href{https://github.com/Sven-LI-sankyuu/KnowMT-Bench}{\texttt{https://github.com/Sven-LI-sankyuu/KnowMT-Bench}}}

cs.CL↗
arXiv2026

Can AI Guide Testing Without Generating Tests? A Comparative Study of AI-Assisted Unit Testing

Recent advances in generative AI have enabled assistants to generate increasingly sophisticated executable unit tests. However, novices may accept these outputs without independently identifying what to test, evaluating test-suite completeness, or verifying correctness, potentially limiting opportunities to develop essential testing skills. Although prior work has explored AI-assisted test generation and feedback-oriented programming tutors, little is known about whether AI can support novices' testing practices without generating tests for them. To investigate this, we introduce TestCompass, a generative AI tutor that provides targeted feedback on weaknesses in students' test suites without generating executable tests. In a classroom study with 18 students, we compared TestCompass with a general AI assistant across two Java unit-testing tasks, examining students' testing behaviors, test-suite quality, and subjective experiences. We found that participants using the general-purpose AI assistant produced more tests, whereas TestCompass participants spent more time developing and verifying tests themselves. However, most test-suite quality measures were comparable between conditions, and TestCompass participants achieved similar requirement coverage with significantly fewer test methods ($p=0.012$) in one task. Neither group reported short-term improvements in perceived testing knowledge or confidence. These findings suggest that feedback-oriented AI can support meaningful testing outcomes while preserving opportunities for hands-on testing practice, offering a complementary approach to AI-assisted testing education.

cs.SE↗
arXiv2026

NosRacer: Dynamic Detection of Race Conditions in On-Device Network Operating Systems

Commercial on-device network operating systems (NOSes) run the complex control planes of production routers and switches, where configuration update tasks are executed by multiple loosely coupled components through asynchronous message passing. Such executions are prone to race conditions: the same ordered input commands may produce different outcomes when messages are delivered in different orders. The race conditions are difficult to expose because they often manifest only as subtle, delayed malfunctions. Existing static analysis techniques lack scalability and precision for large-scale industrial NOSes, while dynamic ones incur substantial system-execution cost when attempting to cover the space of asynchronous message interleavings. To address the challenges above, we present NosRacer, a dynamic analysis framework for race condition detection in industrial-grade on-device NOSes. NosRacer uses a two-phase design. The concentration phase reduces analysis scope by leveraging the locality of configuration update tasks. Race condition detection is then limited to a small subset of involved components and crucial state variables, thereby reducing the detection cost. The perturbation phase proactively perturbs message asynchrony to increase the likelihood of executions with race-condition-exposing message interleavings. It keeps perturbation practical by grouping compatible perturbations for parallel execution and adaptively strengthening perturbations. We implement NosRacer in a commercial, actively developed NOS. NosRacer is integrated into the existing testing factory and used to detect race conditions in 5 key control-plane components. NosRacer achieves 66% precision and detects 21 race conditions confirmed by developers as severe bugs, while keeping the detection overhead below 10%.

cs.DC↗
arXiv2026

JevCoder: When Does Decision Delegation Pay Off in Code Repair?

A coding agent repeatedly makes decisions as it repairs code. These decisions shape repair trajectories and affect the token cost of successive model requests. We ask which repair decisions can be delegated to a low-cost probabilistic judge to reduce token use while maintaining repair quality. The judges answer questions with probabilities over predefined answers, without generating explanations. We evaluate judgment quality across repair stages and test how individual interventions change subsequent repair work. We compare these interventions with unguided repair, which uses no judge, and with the same workflow using the repair model itself as the judge. The benefit depends on the work an intervention changes, rather than the cost of the judgment alone. Based on these studies, we propose JevCoder, a workflow built around Jev, a low-cost probabilistic judge. It combines code-snippet recommendations with action-type guidance and uses judgments to determine when to submit. We evaluate three repair models, each on the same 400 tasks: 200 from SWE-bench Verified and 200 from SWE-Bench Pro Verified. Compared to unguided repair, JevCoder reduces mean repair-model token use by 21.6%-50.6%, with higher observed repair rates across all three models.

cs.SE↗
arXiv2026

Generative modelling for multivariate spatio-temporal downscaling of monthly climate fields

Inferring climate fields at higher spatio-temporal resolutions from coarse representations is essential for accessing socioeconomic and environmental impacts. The downscaling of modelled climate fields faces a trade-off between data storage constraints, simulation efficiency, and resolution requirements. While many methods for spatial downscaling have been proposed, joint spatio-temporal downscaling has received comparable little attention. When deriving, for example, daily fields from climate datasets that are available only as monthly averages at coarser spatial resolution, it is highly challenging to faithfully reproduce extreme event statistics and spatiotemporal consistency. To address this task, we introduce a generative machine learning approach, using a diffusion model to jointly downscale monthly temperature and precipitation fields into temporally consistent daily fields at higher spatial resolution. We demonstrate that our approach can generate high-dimensional multivariate climate fields at daily temporal and 0.25° spatial resolution from monthly input fields at 1° spatial resolution, exhibiting coherent temporal dynamics and physically plausible spatial structures. In particular, by incorporating a conditional guidance technique, we show that a pre-trained diffusion model can be directly leveraged to efficiently generate daily high-resolution fields conditioned on monthly inputs, without requiring additional fine-tuning or retraining. Our probabilistic approach thus provides an efficient and generalizable framework for spatio-temporal climate downscaling, enabling flexible climate fields generation for climate impact assessments and digital Earth simulator constructions.

physics.ao-ph↗
arXiv2026

UltraText Bench: A Comprehensive Bilingual Benchmark for Evaluating Visual Text Rendering in Image Generation

Dense visual text requires image generators to reproduce long strings across multiple regions with correct placement and legibility. As short-string rendering improves, evaluation must test sustained performance across more demanding scenes. We introduce UltraText Bench, a bilingual benchmark for prompt-only generation of dense visual text. It contains 432 prompts spanning 24 real-world scene categories and three difficulty levels, split equally between English and Chinese. Each human-reviewed prompt supplies exact strings for four to twelve text regions, paired with structured references for their content, placement, and visual attributes. We use the Q-Judger vision-language model to assess each image against the complete reference, reporting text fidelity, text clarity, spatial quality, and scene quality. Across 24 model configurations, these dimensions reveal different strengths: Z-Image-Turbo gains 3.81 clarity points over Z-Image-Base while losing 14.76 fidelity points under the reported settings. Performance also varies with workload; Qwen-Image-2512's English composite falls from 86.50 at L1 to 42.86 at L3. Ten participants took part in human evaluation of the automatic scores. Repository: https://github.com/LINs-lab/UltraText_Bench.

cs.CV↗
arXiv2026

CogAdapt: Cognition-informed Sparse Adaptation of Code LLMs

Large language models (LLMs) have become increasingly capable of generating code. However, achieving stronger code-generation performance still often relies on costly model adaptation, i.e., fine-tuning pretrained model parameters. Prior studies have shown correspondence between human code processing and neural models' attention or internal computation. Human-aligned learning approaches use cognitive signals to guide training, but typically adapt a large portion of the model, leaving training costs largely unchanged. Human cognitive signals may indicate not only what the model must learn from, but also where adaptation is most useful. We investigate whether human responses during code reading correspond to code-model behavior and can guide selective adaptation without sacrificing performance. We present CogAdapt, a cognition-informed framework for task-dependent sparse adaptation of code models. CogAdapt first learns transferable program-level and token-level priors from human Electroencephalography (EEG) and attention data, then combines these priors with the frozen model's response to each coding task to determine how much adaptation to allocate and which transformer blocks should receive updates. During fine-tuning, only the selected blocks are updated, while no new human recordings are required for inference. Across Qwen and GLM, we find consistent correspondence between human reading behavior and Mixture-of-Experts (MoE) computation. CogAdapt achieves the best pass@1 across both LiveCodeBench and BigCodeBench, including gains of 10.86 and 6.29 percentage points over matched regular fine-tuning on LiveCodeBench, while reducing gradient-eligible adaptation parameters by 86.21-87.21%. These results suggest that human comprehension signals can provide useful guidance for making code-model adaptation both more selective and more effective.

cs.SE↗
arXiv2026

VTR-Bench: A Systematic Benchmark for Evaluating Visual Text Rendering in Video Generation

Recent video generation models can produce highly realistic videos from natural language instructions, with visual quality approaching cinematic standards. Existing evaluation benchmarks, however, predominantly assess visual quality, aesthetic appeal and physical plausibility, while paying limited attention to text, an essential medium for conveying information in everyday scenes. A generated video may appear visually compelling and feature lifelike subjects, yet still render the text within the scene incorrectly. To address this overlooked dimension, we introduce \textbf{VTR-Bench}, a systematic benchmark for evaluating the \textbf{V}isual \textbf{T}ext \textbf{R}endering capabilities of video generation models. VTR-Bench situates text within concrete application scenarios, such as advertisements and scientific videos, with 300 carefully constructed prompts spanning five scenario categories. We develop an automated evaluation pipeline with human alignments that separately assesses text fidelity through carrier-specific transcription and scene and motion requirements through a prompt-specific chain of query. Beyond evaluation, we introduce a \textbf{Keyframe-Guided Agentic Framework} in which a Director agent coordinates image and video generation with visual evaluation, guiding iterative refinement and candidate selection through visual feedback. Experiments on 11 state-of-the-art models reveal widespread difficulties in accurately rendering scene text, with the best-performing model recording an overall word error rate (WER) of 0.250. We further analyze text rendering failures to characterize the challenges faced by current video generation models. These findings highlight visual text rendering as a key challenge for video generation and demonstrate a practical path toward improvement. Code is available at https://github.com/hardenyu21/VTR-Bench.

cs.CV↗
arXiv2026

Provable Test-Time Scaling for Beam Search in LLM Reasoning

Beam-search-based test-time methods provide an effective way to improve large language model (LLM) performance on long-horizon generation by pruning invalid reasoning paths early, leading to significantly improved reasoning efficiency and more favorable test-time cost scaling. Despite strong empirical success, the theoretical understanding of beam search remains limited. In this paper, we study the test-time compute guarantee of the commonly used beam search framework that uses the model's internal log-likelihood for intermediate scoring, while relying on an external reward model only after a complete response is generated. We first establish a lower bound for vanilla beam search, showing that at least $Ω(C^\star(x)^2)$ samples are required for the optimal response to survive, where $C^\star(x)$ is the token-level coverage coefficient for prompt $x$. This motivates our modified confidence-filtered beam search (CF-Beam), which reduces the sufficient coverage dependence from quadratic to nearly linear under prefix competitiveness, for fixed horizon, gap, and target accuracy. We then show that the regret of CF-Beam is upper-bounded by the probability of rare failure events and the reward estimation error scaled by a path-level coverage coefficient, where the rare-failure term vanishes as per-step sampling increases. Our results highlight a fundamental advantage of beam search over sequence-level inference methods such as Best-of-N and Best-of-Majority. While the guarantees of these approaches typically involve coverage coefficients that grow exponentially with the horizon $L$, CF-Beam controls the dominant search-induced term through a token-level coverage coefficient that scales polynomially with $L$. Our numerical experiments further confirm that beam search is more robust on hard instances and under increasing reasoning horizons.

cs.LG↗
arXiv2026

Towards Understanding On-Policy Distillation through the Lens of Test-Time Scaling

On-policy distillation (OPD) has emerged as a promising post-training technique for enhancing LLM reasoning. Under the reverse KL objective, the idealized optimum of OPD aligns the student distribution with that of the teacher. When the teacher consistently outperforms the student, this naturally suggests that OPD should yield broad improvements over the pre-OPD student. However, do such improvements extend across the entire range of test-time sampling budgets? In this work, we revisit this expectation through the lens of test-time scaling by varying the sampling budget $K$ and evaluating performance with pass@$K$. Across multiple settings, we observe two distinct patterns: OPD can improve pass@$K$ at both small and large sampling budgets, but it can also improve small-budget performance while reducing large-budget pass@$K$. We show one condition that guarantees such a reversal and an idealized reverse KL counterexample where it occurs even when the teacher has higher accuracy on every problem. To choose between two candidate teachers at a target sampling budget, we propose the \textit{Teacher Advantage Score at $K$} (TAS@$K$), which can be computed before OPD training to predict which teacher will lead to a larger improvement in pass@$K$. Across three domains and thirteen benchmarks, the ordering predicted by TAS@$K$ agrees with the observed pass@$K$ improvements of the resulting OPD models in 83.6\% of experiments, providing a useful signal for teacher selection at the target pass@$K$.

cs.LG↗
arXiv2026

TCSAlgBench: Benchmarking Automated Proving for Research-Level Theoretical Computer Science

Large language models perform strongly on competition mathematics, but their research-level reasoning remains difficult to evaluate systematically. Theoretical computer science (TCS) connects algorithm design to explicit guarantees and fundamental limits, providing a setting for evaluating whether models can justify computational improvements with arguments humans can inspect. We introduce TCSAlgBench, a benchmark and reusable pipeline for natural-language proof discovery, comprising 398 theorem-level challenges from 138 STOC and COLT 2026 papers. Expert-designed rules complete paper-specific context, preserve computational assumptions and quantitative guarantees, and withhold constructions when discovering an algorithm is part of the task. For each task, prover systems receive theorem statements and access to cited prior work. The pipeline supports fresh, versioned challenge batches from newly released papers. We evaluate ten model configurations from four families under direct inference and prover-verifier discussion, and compare four agent workflows under matched model-call opportunities. All evaluations use the full benchmark. In the model comparison, GPT-5.6 Sol max achieves the highest five-run verifier-accepted coverage at 23.6% after 10-round discussion. Discussion and repeated sampling improve coverage. In the separate agent comparison using GPT-5.5 xhigh, decomposition improves coverage over discussion, and agentic planning achieves the highest five-run verifier-accepted coverage at 25.4%. TCSAlgBench provides a refreshable testbed for measuring progress in model reasoning and studying how agent workflows support research-level proof discovery.

cs.AI↗
arXiv2026

InterTab: Interleaved Visual-Structure Alignment for Multi-Modal Table Reasoning

Table images preserve structural information that are often lost in text serialization, and reasoning over them requires locating relevant rows, columns, and cells step by step. Current multimodal large language models (MLLMs) encode the whole image once before reasoning, so they cannot pick up row-, column-, and cell-level evidence as the question unfolds. Encoder-side table structure and generic interleaved visual chain-of-thought still do not bind each reasoning step to that structure. We propose \textbf{InterTab}, an \textbf{Inter}leaved structure-aware framework for CoT reasoning over \textbf{Tab}le images, interleaves chain-of-thought with tool calls that crop structure-aligned table regions. First, we build InterTab-22K, includes reasoning trajectories in which each step is tied to both a structural location and a bounding box. InterTab is trained in two stages: supervised structure-aware alignment (SSA) on InterTab-22K teaches the model to interleave reasoning with structure-aligned crops, and active localization optimization (ALO) further optimizes answer correctness, localization IoU, and output format, while penalizing missing or excessive tool calls. Experiments on nine table benchmarks show that InterTab improves the average accuracy of its backbone from 68.28% to 73.17% and achieves the best average performance among all compared methods. Code and data will be released soon.

cs.CV↗
arXiv2026

CausalEmbed: Auto-Regressive Multi-Vector Generation in Latent Space for Visual Document Embedding

Although Multimodal Large Language Models (MLLMs) have shown remarkable potential in Visual Document Retrieval (VDR) through generating high-quality multi-vector embeddings, the substantial storage overhead caused by representing a page with thousands of visual tokens limits their practicality in real-world applications. To address this challenge, we propose an auto-regressive generation approach, CausalEmbed, for constructing multi-vector embeddings. By incorporating iterative margin loss during contrastive training, CausalEmbed encourages the embedding models to learn compact and well-structured representations. Our method enables efficient VDR tasks using only dozens of visual tokens, achieving a 30-155x reduction in token count while maintaining highly competitive performance across various backbones and benchmarks. Theoretical analysis and empirical results demonstrate the unique advantages of auto-regressive embedding generation in terms of training efficiency and scalability at test time. As a result, CausalEmbed introduces a flexible test-time scaling strategy for multi-vector VDR representations and sheds light on the generative paradigm within multimodal document retrieval. Our code is available at https://github.com/Z1zs/Causal-Embed.

cs.CL↗
arXiv2026

HyperCLIP++: Fine-tuning CLIP forOpen-vocabulary Semantic Segmentation in Hyperbolic Space

CLIP, a foundational vision-language model, has emerged as a powerful tool for open-vocabulary semantic segmentation. While freezing CLIP's text encoder is known to preserve its generalization capability, recent studies show that fine-tuning both CLIP's text and image encoders jointly significantly enhances segmentation performance, especially for classes from open sets. In this work, we explain this phenomenon from the perspective of hierarchy alignment, since during fine-tuning, the hierarchical level of image embeddings shifts from image-level to pixel-level. We achieve this by leveraging hyperbolic space, which naturally encodes hierarchical structures. Our key observation is that, during fine-tuning, the hyperbolic radius of CLIP's text embeddings decreases, facilitating better alignment with the pixel-level granularity of visual data. Building on this, we propose HyperCLIP++, a novel and parameter-efficient adaptation strategy. HyperCLIP++ directly adjusts the hyperbolic radius of CLIP's embeddings via scaling transformations to achieve a hierarchy alignment to the target task, i.e., segmentation. To ensure this hierarchy alignment is effected consistently across both modalities and preserves their cross-modal alignment during training, HyperCLIP++ integrates a Dual Cross-Relation Communication (DCRC) module that synchronizes these adjustments between the vision and text pathways. Our experiments show that HyperCLIP++ achieves state-of-the-art performance across three benchmarks while fine-tuning only approximately 5% of CLIP's total parameters. More importantly, we observe that after adjustment, CLIP's text embeddings exhibit a relatively fixed hyperbolic radius across datasets, suggesting that the hierarchical level required for this segmentation task might be quantified using the hyperbolic radius.

cs.CV↗
arXiv2026

PhysBrain 1.5: From Vision-Language Models to Physical Foundation Models

We present PhysBrain 1.5, a unified model for understanding physical environments, generating actions, and predicting future states. Motivated by the physical loop of observation, interaction, and environmental change, we bring these capabilities into a common learning framework. Starting from a general vision--language model, we encode language responses, end-effector motion, and dense visual targets as discrete sequences and jointly optimize them with autoregressive next-token prediction. Pre-training draws its embodied supervision entirely from human interaction videos, using task-centered episodes to pair semantic and spatial context with recovered motion and subsequent observations. We then adapt the model through supervised fine-tuning on a mixture of human demonstrations, robot trajectories, and simulated experience. Across 28 embodied understanding benchmarks, our 8B model achieves an average score of 72.5, setting a new open-source state of the art and performing on par with leading proprietary models such as GPT-6-Astra and Gemini 3.6 Flash. It achieves the best open-source results on 14 benchmarks while retaining general multimodal capabilities. Beyond these understanding evaluations, qualitative examples show the model's ability to produce end-effector trajectories and predict future scenes through spatially aligned RGB, depth, and robot-mask outputs.

cs.CV↗