Source author record

Jinghui Zhang

Jinghui Zhang appears in the imported research catalog. Authorship, coauthor and topic links are available while profile ownership is still unclaimed.

ResearcherUnclaimed source record

Applications Artificial Intelligence Computation Computation and Language Genomics Networking and Internet Architecture Quantitative Methods

Catalog footprint

What is connected

4works

7topics

4close collaborators

Actions

Connect this record

Open graph Browse works

Inspect adjacent papers, topics, institutions and collaborators without losing the researcher page.

Building this map preview

BZPEER is loading the nearby papers, people, topics and institutions for this page.

preprint2026arXiv

M3MAD-Bench: Are Multi-Agent Debates Really Effective Across Domains and Modalities?

As an agent-level reasoning and coordination paradigm, Multi-Agent Debate (MAD) orchestrates multiple agents through structured debate to improve answer quality and support complex reasoning. However, existing research on MAD suffers from two fundamental limitations: evaluations are conducted under fragmented and inconsistent settings, hindering fair comparison, and are largely restricted to single-modality scenarios that rely on textual inputs only. To address these gaps, we introduce M3MAD-Bench, a unified and extensible benchmark for evaluating MAD methods across Multi-domain tasks, Multi-modal inputs, and Multi-dimensional metrics. M3MAD-Bench establishes standardized protocols over five core task domains: Knowledge, Mathematics, Medicine, Natural Sciences, and Complex Reasoning, and systematically covers both pure text and vision-language datasets, enabling controlled cross-modality comparison. We evaluate MAD methods on nine base models spanning different architectures, scales, and modality capabilities. Beyond accuracy, M3MAD-Bench incorporates efficiency-oriented metrics such as token consumption and inference time, providing a holistic view of performance--cost trade-offs. Extensive experiments yield systematic insights into the effectiveness, robustness, and efficiency of MAD across text-only and multimodal scenarios. We believe M3MAD-Bench offers a reliable foundation for future research on standardized MAD evaluation. The code is available at http://github.com/liaolea/M3MAD-Bench.

preprint2026arXiv

The Cylindrical Representation Hypothesis for Language Model Steering

Steering is a widely used technique for controlling large language models, yet its effects are often unstable and hard to predict. Existing theoretical accounts are largely based on the Linear Representation Hypothesis (LRH). While LRH assumes that concepts can be orthogonalized for lossless control, this idealized mapping fails in real representations and cannot account for the observed unpredictability of steering. By relaxing LRH's orthogonality assumption while preserving linear representations, we show that overlapping concept contributions naturally yield a sample-specific axis-orthogonal structure. We formalize this as the Cylindrical Representation Hypothesis (CRH). In CRH, a central axis captures the main difference between concept absence and presence and drives concept generation. A surrounding normal plane controls steering sensitivity by determining how easily the axis can activate the target concept. Within this plane, only specific sensitive sectors strongly facilitate concept activation, while other sectors can suppress or delay it. While the surrounding normal plane can be reliably identified from difference vectors, the sensitive sector cannot, introducing intrinsic uncertainty at the sector level. This uncertainty provides a principled explanation for why steering outcomes often fluctuate even when using well-aligned directions. Our experiments verify the existence of the cylindrical structure and demonstrate that CRH provides a valid and practical way to interpret model steering behavior in real settings: https://github.com/mbzuai-nlp/CRH.

preprint2013arXiv

Birkhoff-von-Neumann Switches with Deflection-Compensated Mechanism

Despite the high throughput and low complexity achieved by input scheduling based on Birkhoff-von-Neumann (BvN) decomposition; the performance of the BvN switch becomes less predictable when the input traffic is bursty. In this paper, we propose a deflection-compensated BvN (D-BvN) switch architecture to enhance the quasi-static scheduling based on BvN decomposition. The D-BvN switches provide capacity guarantee for virtual circuits (VCs) and deflect bursty traffic when overflow occurs. The deflection scheme is devised to offset the excessive buffer requirement of each VC when input traffic is bursty. The design of our conditional deflection mechanism is based on the fact that it is unlikely that the traffic input to VCs is all bursty at the same time; most likely some starving VCs have spare capacities when some other VCs are in the overflow state. The proposed algorithm makes full use of the spare capacities of those starving VCs to deflect the overflow traffic to other inputs and provide bandwidth for the deflected traffic to re-access the desired VC. Our analysis and simulation show that this deflection-compensated mechanism can support BvN switches to achieve close to 100% throughput of offered load even with bursty input traffic, and reduces the average end-to-end delay and delay jitter. Also, our result indicates that the packet out-of-sequence probability due to deflection of overflow traffic is negligible, thus only a small re-sequencing buffer is needed at each output port.

preprint2009arXiv

Needles in the Haystack: Identifying Individuals Present in Pooled Genomic Data

Recent publications have described and applied a novel metric that quantifies the genetic distance of an individual with respect to two population samples, and have suggested that the metric makes it possible to infer the presence of an individual of known genotype in a sample for which only the marginal allele frequencies are known. However, the assumptions, limitations, and utility of this metric remained incompletely characterized. Here we present an exploration of the strengths and limitations of that method. In addition to analytical investigations of the underlying assumptions, we use both real and simulated genotypes to test empirically the method's accuracy. The results reveal that, when used as a means by which to identify individuals as members of a population sample, the specificity is low in several circumstances. We find that the misclassifications stem from violations of assumptions that are crucial to the technique yet hard to control in practice, and we explore the feasibility of several methods to improve the sensitivity. Additionally, we find that the specificity may still be lower than expected even in ideal circumstances. However, despite the metric's inadequacies for identifying the presence of an individual in a sample, our results suggest potential avenues for future research on tuning this method to problems of ancestry inference or disease prediction. By revealing both the strengths and limitations of the proposed method, we hope to elucidate situations in which this distance metric may be used in an appropriate manner. We also discuss the implications of our findings in forensics applications and in the protection of GWAS participant privacy.