arXiv subjects

Zhengheng Li

Publications and source records attributed to Zhengheng Li.

6 recordsLinked to original sources

arXiv2026

Rubric-Guided Process Reward for Stepwise Model Routing

Stepwise model routing improves the efficiency of Large Reasoning Models (LRMs) by assigning each reasoning step to a suitable model. Recent methods formulate routing as a sequential decision process and train the router with reinforcement learning. However, although they model routing as a process, they still supervise the router with outcome rewards. Such rewards only reflect final answer correctness and fail to evaluate intermediate routing decisions, which can weaken performance and generalization. To address this gap, we propose RoRo, a rubric-guided process reward framework for stepwise model routing. RoRo first collects diverse routing trajectories and constructs preference pairs based on outcome, cost, and process quality. It then trains a Rubricor to generate a query-specific evaluation rubric and a Judge to score routing trajectories under this rubric through alternating optimization. The resulting process rewards are combined with outcome rewards to optimize the routing policy via GRPO. Experiments on five reasoning benchmarks under both same-family and cross-family settings show that RoRo consistently outperforms strong baselines and achieves better accuracy and cost trade-offs.

cs.AI↗

arXiv2026

VulTriage: Triple-Path Context Augmentation for LLM-Based Vulnerability Detection

Automated vulnerability detection is a fundamental task in software security, yet existing learning-based methods still struggle to capture the structural dependencies, domain-specific vulnerability knowledge, and complex program semantics required for accurate detection. Recent Large Language Models (LLMs) have shown strong code understanding ability, but directly prompting them with raw source code often leads to missed vulnerabilities or false alarms, especially when vulnerable and benign functions differ only in subtle semantic details. To address this, we propose VulTriage, a triple-path context augmentation framework for LLM-based vulnerability detection. VulTriage enhances the LLM input through three complementary paths: a Control Path that extracts and verbalizes AST, CFG, and DFG information to expose control and data dependencies; a Knowledge Path that retrieves relevant CWE-derived vulnerability patterns and examples through hybrid dense--sparse retrieval; and a Semantic Path that summarizes the functional behavior of the code before the final judgment. These contexts are integrated into a unified instruction to guide the LLM toward more reliable vulnerability reasoning. Experiments on the PrimeVul pair test set show that VulTriage achieves state-of-the-art performance, outperforming existing deep learning and LLM-based baselines on key pair-wise and classification metrics. Further ablation studies verify the effectiveness of each path, and additional experiments on the Kotlin dataset demonstrate the generalization ability of VulTriage under low-resource and class-imbalanced settings. Our code is available at https://github.com/vinsontang1/VulTriage

cs.AI↗

arXiv2026

Where to Place the Query? Unveiling and Mitigating Positional Bias in In-Context Learning for Diffusion LLMs via Decoding Dynamics

While In-Context Learning (ICL) is extensively studied in Autoregressive (AR) LLMs, its mechanism within Diffusion Large Language Models (dLLMs) remains largely unexplored. Unlike AR models restricted by unidirectional causal masking, dLLMs intrinsically utilize bidirectional attention, offering extensive spatial flexibility for query placement. Unfortunately, current practices conventionally inherit AR-style trailing-query templates, often overlooking the structural paradigm shift. This paper presents a comprehensive analysis unveiling that query position is actually a first-order variable in dLLMs. Through empirical decoupling, we demonstrate that positional variance impacts generation quality on par with example semantic quality. Internally, this positional sensitivity stems from a spatial ``Recency Effect'' in attention flow and task-dependent shifts in decoding trajectories. To mitigate this instability without ground-truth labels, we reveal that traditional single-step confidence ($C_{decoded}$) fails in dLLMs. Instead, we propose Average Confidence ($\overline{C}$), a novel metric tracking the iterative decoding process. By establishing the foundational spatial ICL baselines, we introduce Auto-ICL, a training-free adaptive routing strategy that dynamically optimizes query placement, robustly approaching oracle performance across heterogeneous reasoning and perception tasks.

cs.CL↗

arXiv2018

In-flight energy calibration of the space-borne Compton polarimeter POLAR

POLAR is a compact wide-field space-borne detector for precise measurements of the linear polarisation of hard X-rays emitted by transient sources in the energy range from 50 keV to 500 keV. It consists of a 40$\times$40 array of plastic scintillator bars used as a detection material. The bars are grouped in 25 detector modules. The energy range sensitivity of POLAR is optimized to match with the prompt emission photons from the gamma-ray bursts (GRBs). Polarization measurements of the prompt emission would probe source geometries, emission mechanisms and magnetic structures in GRB jets. The instrument can also detect hard X-rays from solar flares and be used for precise measurement of their polarisation. POLAR was launched into a low Earth orbit on-board the Chinese space-lab TG-2 on September 15th, 2016. To achieve high accuracies in polarisation measurements it is essential to assure both before and during the flight a precise energy calibration. Such calibrations are performed with four low activity $^{22}$Na radioactive sources placed inside the instrument. Energy conversion factors are related to Compton edge positions from the collinear annihilation photons from the sources. This paper presents main principles of the in-flight calibration, describes studies of the method based on Monte Carlo simulations and its laboratory verification and finally provides some observation results based on the in-flight data analysis.

physics.ins-det↗

arXiv2018

In-Orbit Instrument Performance Study and Calibration for POLAR Polarization Measurements

POLAR is a compact space-borne detector designed to perform reliable measurements of the polarization for transient sources like Gamma-Ray Bursts in the energy range 50-500keV. The instrument works based on the Compton Scattering principle with the plastic scintillators as the main detection material along with the multi-anode photomultiplier tube. POLAR has been launched successfully onboard the Chinese space laboratory TG-2 on 15th September, 2016. In order to reliably reconstruct the polarization information a highly detailed understanding of the instrument is required for both data analysis and Monte Carlo studies. For this purpose a full study of the in-orbit performance was performed in order to obtain the instrument calibration parameters such as noise, pedestal, gain nonlinearity of the electronics, threshold, crosstalk and gain, as well as the effect of temperature on the above parameters. Furthermore the relationship between gain and high voltage of the multi-anode photomultiplier tube has been studied and the errors on all measurement values are presented. Finally the typical systematic error on polarization measurements of Gamma-Ray Bursts due to the measurement error of the calibration parameters are estimated using Monte Carlo simulations.

astro-ph.IM↗

arXiv2018

A Low-latency Pipeline for GRB Light Curve and Spectrum using Fermi/GBM Near Real-time Data

Rapid response and short time latency are very important for Time Domain Astronomy, such as the observations of Gamma-ray Bursts (GRBs) and electromagnetic (EM) counterparts of gravitational waves (GWs). Based on the near real-time Fermi/GBM data, we developed a low-latency pipeline to automatically calculate the temporal and spectral properties of GRBs. With this pipeline, some important parameters can be obtained, such as T90 and fluence, within ~20 minutes after the GRB trigger. For ~90% GRBs, T90 and fluence are consistent with the GBM catalog results within 2 sigma errors. This pipeline has been used by the Gamma-ray Bursts Polarimeter (POLAR) and the Insight Hard X-ray Modulation Telescope (Insight-HXMT) to follow up the bursts of interest. For GRB 170817A, the first EM counterpart of GW events detected by Fermi/GBM and INTEGRAL/SPI-ACS, the pipeline gave T90 and spectral information in 21 minutes after the GBM trigger, providing important information for POLAR and Insight-HXMT observations.

astro-ph.IM↗