arXiv subjects

Xiaohui Huang

Publications and source records attributed to Xiaohui Huang.

7 recordsLinked to original sources

arXiv2026

Encoded but Not Decoded: Layer-Localized Evidence for a Three-Level Gap in LLM Syntax

A language model can fail a syntactic test in two distinct ways: by not encoding the relevant structure, or by encoding it but failing to use it at the output. Behavioral evaluation alone cannot tell these apart. We propose a three-level evaluation framework (behavioral deployment, LM-head readout, and probe recoverability) measured on the same items under the same binary decision. Using a compact trilingual (English, Chinese, German) control-dependency benchmark, we find that probe recoverability exceeds or equals LM-head readout, which in turn exceeds or equals behavioral deployment, across seven models and all three languages in the aggregate. The recoverability surplus is never negative across all 14 (model, task) conditions. The disconnect concentrates in subject-control, where a nearest-noun heuristic gives the wrong answer. The single largest gap (0.653) appears on Qwen3-0.6B Instruct in question answering. The gap persists at Qwen3-14B Instruct. Instruction tuning degrades deployment more than encoding in percentage terms. We rule out option-position bias, late-layer erasure, output-formatting artifacts, and probe-training variance. The pattern is consistent with decoding that favors surface shortcuts, and the behavior-probe gap measures the strength of that preference. Activation patching shows the gap is layer-localized. Under instruction tuning, the LM-head-decoded layer shifts approximately ten layers later than the probe-decoded layer. These findings argue that behavioral evaluation understates what models encode, while probing alone overstates what they deploy.

cs.CL↗

arXiv2026

ReviewGuard: Aligning LLM-Assisted Peer Review with Long-Term Scientific Impact

Peer review is central to scientific quality control, yet it can undervalue papers that later achieve substantial citation impact. While frontier large language models have shown promise in automating aspects of peer review, they primarily mimic human reviewer preferences rather than predict long-term scientific value. We introduce ReviewGuard, a two-stage framework that aligns LLM-generated reviews with citation-based estimates of long-term scientific impact rather than contemporaneous reviewer judgments. On 20,861 AI/ML papers from OpenReview augmented with Semantic Scholar citation data, ReviewGuard achieves a Spearman correlation of \r{ho} = 0.776 with future citations on rejected-then-published papers, outperforming human reviewers (\r{ho} = 0.492) and a supervised Expert model (\r{ho} = 0.681). Under the same decision threshold, ReviewGuard flags 10.2% of high-impact rejected papers, compared with 1.8% for human reviewers, corresponding to a 5.6x improvement. Our results demonstrate that impact-aligned reinforcement learning can provide editors with a complementary signal for identifying high-potential work, without replacing human judgment.

cs.DL↗

arXiv2025

Restoring Network Evolution from Static Structure

The dynamical evolution of complex networks underpins the structure-function relationships in natural and artificial systems. Yet, restoring a network's formation from a single static snapshot remains challenging. Here, we present a transferable machine learning framework that infers network evolutionary trajectories solely from present topology. By integrating graph neural networks with transformers, our approach unlocks a latent temporal dimension directly from the static topology. Evaluated across diverse domains, the framework achieves high transfer accuracy of up to 95.3%, demonstrating its robustness and transferability. Applied to the Drosophila brain connectome, it restores the formation times of over 2.6 million neural connections, revealing that early-forming links support essential behaviors such as mating and foraging, whereas later-forming connections underpin complex sensory and social functions. These results demonstrate that a substantial fraction of evolutionary information is encoded within static network architecture, offering a powerful, general tool for elucidating the hidden temporal dynamics of complex systems.

physics.soc-ph↗

arXiv2019

Intelligent Intersection: Two-Stream Convolutional Networks for Real-time Near Accident Detection in Traffic Video

In Intelligent Transportation System, real-time systems that monitor and analyze road users become increasingly critical as we march toward the smart city era. Vision-based frameworks for Object Detection, Multiple Object Tracking, and Traffic Near Accident Detection are important applications of Intelligent Transportation System, particularly in video surveillance and etc. Although deep neural networks have recently achieved great success in many computer vision tasks, a uniformed framework for all the three tasks is still challenging where the challenges multiply from demand for real-time performance, complex urban setting, highly dynamic traffic event, and many traffic movements. In this paper, we propose a two-stream Convolutional Network architecture that performs real-time detection, tracking, and near accident detection of road users in traffic video data. The two-stream model consists of a spatial stream network for Object Detection and a temporal stream network to leverage motion features for Multiple Object Tracking. We detect near accidents by incorporating appearance features and motion features from two-stream networks. Using aerial videos, we propose a Traffic Near Accident Dataset (TNAD) covering various types of traffic interactions that is suitable for vision-based traffic analysis tasks. Our experiments demonstrate the advantage of our framework with an overall competitive qualitative and quantitative performance at high frame rates on the TNAD dataset.

cs.CV↗

arXiv2017

Approximation Algorithm for Fault-Tolerant Virtual Backbone in Wireless Sensor Networks

To save energy and alleviate interferences in a wireless sensor network, the usage of virtual backbone was proposed. Because of accidental damages or energy depletion, it is desirable to construct a fault tolerant virtual backbone, which can be modeled as a $k$-connected $m$-fold dominating set (abbreviated as $(k,m)$-CDS) in a graph. A node set $C\subseteq V(G)$ is a $(k,m)$-CDS of graph $G$ if every node in $V(G)\backslash C$ is adjacent with at least $m$ nodes in $C$ and the subgraph of $G$ induced by $C$ is $k$-connected. In this paper, we present an approximation algorithm for the minimum $(3,m)$-CDS problem with $m\geq3$. The performance ratio is at most $γ$, where $γ=α+8+2\ln(2α-6)$ for $α\geq4$ and $γ=3α+2\ln2$ for $α<4$, and $α$ is the performance ratio for the minimum $(2,m)$-CDS problem. Using currently best known value of $α$, the performance ratio is $\lnδ+o(\lnδ)$, where $δ$ is the maximum degree of the graph, which is asymptotically best possible in view of the non-approximability of the problem. This is the first performance-guaranteed algorithm for the minimum $(3,m)$-CDS problem on a general graph. Furthermore, applying our algorithm on a unit disk graph which models a homogeneous wireless sensor network, the performance ratio is less than 27, improving previous ratio 62.3 by a large amount for the $(3,m)$-CDS problem on a unit disk graph.

cs.DM↗

arXiv2016

Performance Guaranteed Approximation Algorithm for Minimum $k$-Connected $m$-Fold Dominating Set

To achieve an efficient routing in a wireless sensor network, connected dominating set (CDS) is used as virtual backbone. A fault-tolerant virtual backbone can be modeled as a $(k,m)$-CDS. For a connected graph $G=(V,E)$ and two fixed integers $k$ and $m$, a node set $C\subseteq V$ is a $(k,m)$-CDS of $G$ if every node in $V\setminus C$ has at least $m$ neighbors in $C$, and the subgraph of $G$ induced by $C$ is $k$-connected. Previous to this work, approximation algorithms with guaranteed performance ratio in a general graph were know only for $k\leq 3$. This paper makes a significant progress by presenting a $(2k-1)α_0$ approximation algorithm for general $k$ and $m$ with $m\geq k$, where $α_0$ is the performance ratio for the minimum CDS problem. Using currently best known ratio for $α_0$, our algorithm has performance ratio $O(\lnΔ)$, where $Δ$ is the maximum degree of the graph.

cs.DM↗

arXiv2014

Application of Multilayer Feedforward Neural Networks in Predicting Tree Height and Forest Stock Volume of Chinese Fir

Wood increment is critical information in forestry management. Previous studies used mathematics models to describe complex growing pattern of forest stand, in order to determine the dynamic status of growing forest stand in multiple conditions. In our research, we aimed at studying non-linear relationships to establish precise and robust Artificial Neural Networks (ANN) models to predict the precise values of tree height and forest stock volume based on data of Chinese fir. Results show that Multilayer Feedforward Neural Networks with 4 nodes (MLFN-4) can predict the tree height with the lowest RMS error (1.77); Multilayer Feedforward Neural Networks with 7 nodes (MLFN-7) can predict the forest stock volume with the lowest RMS error (4.95). The training and testing process have proved that our models are precise and robust.

cs.CE↗