arXiv subjects

Junwei Zhang

Publications and source records attributed to Junwei Zhang.

At least 19 recordsLinked to original sources

arXiv2026

HalluTracer: Hallucination Detection via Depth-Averaging Truth Signals

Even well-aligned large language models confidently generate factually incorrect text, making hallucination a persistent reliability risk in high-stakes deployments. These models nonetheless carry linearly separable truthfulness signals in their internal representations. Existing white-box detectors, however, collapse this evidence to isolated components or a single depth, discarding discriminative information distributed across the full forward pass. We introduce HalluTracer, a detection framework that reads and aggregates truthfulness evidence across every layer of the forward pass before the model emits any answer token. A geometric analysis reveals that the per-layer signals are weakly correlated, so that simple depth averaging suppresses layer-specific noise and captures nearly all linearly accessible information. Across six open-source language models and five hallucination benchmarks, HalluTracer consistently outperforms matched white-box baselines, with gains ranging from one to fourteen points. Collectively, our work recasts hallucination detection from a layer-selection problem into a depth-aggregation problem governed by the geometric sparsity of the truthfulness signal.

cs.CL↗

arXiv2026

Beyond Reasoning: Reinforcement Learning Unlocks Parametric Knowledge in LLMs

Reinforcement learning (RL) has achieved remarkable success in LLM reasoning, but whether it can also improve direct recall of parametric knowledge remains an open question. We study this question in a controlled zero-shot, one-hop, closed-book QA setting with no chain-of-thought, training only on binary correctness rewards and applying fact-level train-test deduplication to ensure gains reflect improved recall rather than reasoning or memorization. Across three model families and multiple factual QA benchmarks, RL yields ~27% average relative gains, surpassing both training- and inference-time baselines alike. Mechanistically, RL primarily redistributes probability mass over existing knowledge rather than acquiring new facts, moving correct answers from the low-probability tail into reliable greedy generations. Our data-attribution study reveals that the hardest examples are the most informative: those whose answers never appear in 128 pre-RL samples (only ~18% of training data) drive ~83% of the gain, since rare correct rollouts still emerge during training and get reinforced. Together, these findings broaden the role of RL beyond reasoning, repositioning it as a tool for unlocking rather than acquiring latent parametric knowledge.

cs.CL↗

arXiv2026

Polymorphic crystallites model for monolayer amorphous materials

Modeling the atomic structure of amorphous materials has long been a critical challenge in materials science. Recent advances in monolayer amorphous materials enable direct observation of their atomic structures, paving the way for a better understanding of their atomic-scale models. Here, we investigate amorphous multielement monolayers using machine learning potential from first-principles total energies via energy-driven kinetic Monte Carlo based active-learning framework. A polymorphic crystallite model is proposed to describe the atomic configuration of monolayer amorphous boron nitride, as it consists of coexisting crystallite of $o-B_2N_2$ and $o-B_4N_4$ structural motifs. Generality of the polymorphic crystallite model is further validated in two other multielement monolayer amorphous systems. Monolayer amorphous LiCl shows coexisting hexagonal and tetragonal crystallites, while monolayer amorphous BCN contains a combination of graphene-like, h-BN-like, and borophene-like crystallites. These findings expand the classical picture of amorphous structure models and offer new insight into the microscopic structure of amorphous materials.

cond-mat.mtrl-sci↗

arXiv2026

Origin of anomalous p-type conductivity in monolayer Fe-doped MoS2

Substitutional doping effectively modulates carrier polarity of semiconducting two-dimensional (2D) transition metal dichalcogenides (TMDs) like MoS2. Although Fe doping typically induces n-type conductivity in monolayer MoS2, anomalous p-type behavior has also been experimentally reported, the origin of which remains unresolved. Here, we prove that this anomalous p-type conductivity originates from defect associates formed through interactions between Fe dopants and S atoms, which consists of three Fe substituting Mo (FeMo) point defects arranged into an equilateral triangle with a central S atom, denoted as 3FeMo-S associate. Its p-type effect is directly verified through scanning tunneling microscopy/scanning tunneling spectroscopy (STM/STS) measurement, in sharp contrast to the n-type behavior induced by isolated FeMo point defects, and the conclusion is further supported by electrical transport measurements and first-principles calculations. Similar 3FeW-S associates and their p-type doping effect are also identified in monolayer Fe-doped WS2. This work resolves a longstanding controversy and highlights the critical role of defect associates in modulating properties of 2D TMDs.

cond-mat.mtrl-sci↗

arXiv2026

ChronosAudio: A Comprehensive Long-Audio Benchmark for Evaluating Audio-Large Language Models

Although Audio Large Language Models (ALLMs) have witnessed substantial advancements, their long audio understanding capabilities remain unexplored. A plethora of benchmarks have been proposed for general audio tasks, they predominantly focus on short-form clips, leaving without a consensus on evaluating ALLMs over extended durations. This paper proposes ChronosAudio, the first multi-task benchmark tailored for long-audio understanding in ALLMs. It encompasses six major task categories and comprises 36,000 test instances totaling over 200 hours audio, stratified into short, middle, and long-form categories to comprehensively evaluate length generalization. Extensive experiments on 16 state-of-the-art models using ChronosAudio yield three critical findings: 1.Precipitous Long-Context Collapse: ALLMs exhibit a severe inability to sustain performance, with the transition from short to long contexts triggering a staggering performance degradation of over 90% in specific tasks. 2.Structural Attention Dilution: Performance degradation stems from a fundamental failure in maintaining temporal locality; attention mechanisms suffer from significant diffusion in later sequences. 3.Restorative Ceiling of Mitigation: Current strategies only offer 50% recovery. These findings reveal significant challenges in long-audio, underscoring the urgent need for approaches to achieve robust, document-level audio reasoning.

cs.SD↗

arXiv2026

BaseCal: Unsupervised Confidence Calibration via Base Model Signals

Reliable confidence is essential for trusting the outputs of LLMs, yet widely deployed post-trained LLMs (PoLLMs) typically compromise this trust with severe overconfidence. In contrast, we observe that their corresponding base LLMs often remain well-calibrated. This naturally motivates us to calibrate PoLLM confidence using the base LLM as a reference. This work proposes two ways to achieve this. A straightforward solution, BaseCal-ReEval, evaluates PoLLM's responses by feeding them into the base LLM to get average probabilities as confidence. While effective, this approach introduces additional inference overhead. To address this, we propose BaseCal-Proj, which trains a lightweight projection to map the final-layer hidden states of PoLLMs back to those of their base LLMs. These projected states are then processed by the base LLM's output layer to derive base-calibrated confidence for PoLLM's responses. Notably, BaseCal is an unsupervised, plug-and-play solution that operates without human labels or LLM modifications. Experiments across five datasets and three LLM families demonstrate the effectiveness of BaseCal, reducing Expected Calibration Error (ECE) by an average of 42.90\% compared to the best unsupervised baselines.

cs.CL↗

arXiv2025

PolarStore: High-Performance Data Compression for Large-Scale Cloud-Native Databases

In recent years, resource elasticity and cost optimization have become essential for RDBMSs. While cloud-native RDBMSs provide elastic computing resources via disaggregated computing and storage, storage costs remain a critical user concern. Consequently, data compression emerges as an effective strategy to reduce storage costs. However, existing compression approaches in RDBMSs present a stark trade-off: software-based approaches incur significant performance overheads, while hardware-based alternatives lack the flexibility required for diverse database workloads. In this paper, we present PolarStore, a compressed shared storage system for cloud-native RDBMSs. PolarStore employs a dual-layer compression mechanism that combines in-storage compression in PolarCSD hardware with lightweight compression in software. This design leverages the strengths of both approaches. PolarStore also incorporates database-oriented optimizations to maintain high performance on critical I/O paths. Drawing from large-scale deployment experiences, we also introduce hardware improvements for PolarCSD to ensure host-level stability and propose a compression-aware scheduling scheme to improve cluster-level space efficiency. PolarStore is currently deployed on thousands of storage servers within PolarDB, managing over 100 PB of data. It achieves a compression ratio of 3.55 and reduces storage costs by approximately 60%. Remarkably, these savings are achieved while maintaining performance comparable to uncompressed clusters.

cs.DC↗

arXiv2025

The Shawshank Redemption of Embodied AI: Understanding and Benchmarking Indirect Environmental Jailbreaks

The adoption of Vision-Language Models (VLMs) in embodied AI agents, while being effective, brings safety concerns such as jailbreaking. Prior work have explored the possibility of directly jailbreaking the embodied agents through elaborated multi-modal prompts. However, no prior work has studied or even reported indirect jailbreaks in embodied AI, where a black-box attacker induces a jailbreak without issuing direct prompts to the embodied agent. In this paper, we propose, for the first time, indirect environmental jailbreak (IEJ), a novel attack to jailbreak embodied AI via indirect prompt injected into the environment, such as malicious instructions written on a wall. Our key insight is that embodied AI does not ''think twice'' about the instructions provided by the environment -- a blind trust that attackers can exploit to jailbreak the embodied agent. We further design and implement open-source prototypes of two fully-automated frameworks: SHAWSHANK, the first automatic attack generation framework for the proposed attack IEJ; and SHAWSHANK-FORGE, the first automatic benchmark generation framework for IEJ. Then, using SHAWSHANK-FORGE, we automatically construct SHAWSHANK-BENCH, the first benchmark for indirectly jailbreaking embodied agents. Together, our two frameworks and one benchmark answer the questions of what content can be used for malicious IEJ instructions, where they should be placed, and how IEJ can be systematically evaluated. Evaluation results show that SHAWSHANK outperforms eleven existing methods across 3,957 task-scene combinations and compromises all six tested VLMs. Furthermore, current defenses only partially mitigate our attack, and we have responsibly disclosed our findings to all affected VLM vendors.

cs.CR↗

arXiv2025

Reconstruction of the Effective Energy-deposition Vertex of Muon Showers using PMT Waveform in a Large-scale Liquid Scintillator Detector

Cosmogenic muon-induced radioactive isotopes pose a significant background source for deep-underground low-background experiments. Although rock overburdens at underground sites substantially attenuate the cosmogenic muon flux, residual muon-induced backgrounds still require active suppression. For future multi-kiloton liquid scintillator (LS) detectors, such as the Jiangmen Underground Neutrino Observatory (JUNO), shower muons contribute to more than 88\% of all muon-induced isotopes. Consequently, precise reconstruction of shower vertices is essential for implementing localized spatial vetoes. We propose a novel waveform-based method to reconstruct the shower vertex, defined as the energy-deposition centroid. By subtracting the track contributions from non-shower muons in the recorded waveforms, the isolated shower component is extracted. Subsequently, combined with a photon propagation model and an iterative optimization algorithm, the shower vertex positions are reconstructed. Simulations show that for 68\% of events, the single shower vertex resolution is better than 0.16~m, 0.15~m, and 0.26~m along X, Y, and Z respectively. Furthermore, the reconstruction efficiency exceeds 96\% when requiring the distance between the reconstructed and true vertices to be less than 3.0 m. This method provides a critical technical foundation for muon-induced background suppression in JUNO and other large-scale LS detectors.

hep-ex↗

arXiv2025

A Dynamic Knowledge Update-Driven Model with Large Language Models for Fake News Detection

As the Internet and social media evolve rapidly, distinguishing credible news from a vast amount of complex information poses a significant challenge. Due to the suddenness and instability of news events, the authenticity labels of news can potentially shift as events develop, making it crucial for fake news detection to obtain the latest event updates. Existing methods employ retrieval-augmented generation to fill knowledge gaps, but they suffer from issues such as insufficient credibility of retrieved content and interference from noisy information. We propose a dynamic knowledge update-driven model for fake news detection (DYNAMO), which leverages knowledge graphs to achieve continuous updating of new knowledge and integrates with large language models to fulfill dual functions: news authenticity detection and verification of new knowledge correctness, solving the two key problems of ensuring the authenticity of new knowledge and deeply mining news semantics. Specifically, we first construct a news-domain-specific knowledge graph. Then, we use Monte Carlo Tree Search to decompose complex news and verify them step by step. Finally, we extract and update new knowledge from verified real news texts and reasoning paths. Experimental results demonstrate that DYNAMO achieves the best performance on two real-world datasets.

cs.CL↗

arXiv2025

STEP: Stepwise Curriculum Learning for Context-Knowledge Fusion in Conversational Recommendation

Conversational recommender systems (CRSs) aim to proactively capture user preferences through natural language dialogue and recommend high-quality items. To achieve this, CRS gathers user preferences via a dialog module and builds user profiles through a recommendation module to generate appropriate recommendations. However, existing CRS faces challenges in capturing the deep semantics of user preferences and dialogue context. In particular, the efficient integration of external knowledge graph (KG) information into dialogue generation and recommendation remains a pressing issue. Traditional approaches typically combine KG information directly with dialogue content, which often struggles with complex semantic relationships, resulting in recommendations that may not align with user expectations. To address these challenges, we introduce STEP, a conversational recommender centered on pre-trained language models that combines curriculum-guided context-knowledge fusion with lightweight task-specific prompt tuning. At its heart, an F-Former progressively aligns the dialogue context with knowledge-graph entities through a three-stage curriculum, thus resolving fine-grained semantic mismatches. The fused representation is then injected into the frozen language model via two minimal yet adaptive prefix prompts: a conversation prefix that steers response generation toward user intent and a recommendation prefix that biases item ranking toward knowledge-consistent candidates. This dual-prompt scheme allows the model to share cross-task semantics while respecting the distinct objectives of dialogue and recommendation. Experimental results show that STEP outperforms mainstream methods in the precision of recommendation and dialogue quality in two public datasets.

cs.AI↗

arXiv2025

CPA-RAG:Covert Poisoning Attacks on Retrieval-Augmented Generation in Large Language Models

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by incorporating external knowledge, but its openness introduces vulnerabilities that can be exploited by poisoning attacks. Existing poisoning methods for RAG systems have limitations, such as poor generalization and lack of fluency in adversarial texts. In this paper, we propose CPA-RAG, a black-box adversarial framework that generates query-relevant texts capable of manipulating the retrieval process to induce target answers. The proposed method integrates prompt-based text generation, cross-guided optimization through multiple LLMs, and retriever-based scoring to construct high-quality adversarial samples. We conduct extensive experiments across multiple datasets and LLMs to evaluate its effectiveness. Results show that the framework achieves over 90\% attack success when the top-k retrieval setting is 5, matching white-box performance, and maintains a consistent advantage of approximately 5 percentage points across different top-k values. It also outperforms existing black-box baselines by 14.5 percentage points under various defense strategies. Furthermore, our method successfully compromises a commercial RAG system deployed on Alibaba's BaiLian platform, demonstrating its practical threat in real-world applications. These findings underscore the need for more robust and secure RAG frameworks to defend against poisoning attacks.

cs.CR↗

arXiv2025

Assessing and Advancing Benchmarks for Evaluating Large Language Models in Software Engineering Tasks

Large language models (LLMs) are gaining increasing popularity in software engineering (SE) due to their unprecedented performance across various applications. These models are increasingly being utilized for a range of SE tasks, including requirements engineering and design, code analysis and generation, software maintenance, and quality assurance. As LLMs become more integral to SE, evaluating their effectiveness is crucial for understanding their potential in this field. In recent years, substantial efforts have been made to assess LLM performance in various SE tasks, resulting in the creation of several benchmarks tailored to this purpose. This paper offers a thorough review of 291 benchmarks, addressing three main aspects: what benchmarks are available, how benchmarks are constructed, and the future outlook for these benchmarks. We begin by examining SE tasks such as requirements engineering and design, coding assistant, software testing, AIOPs, software maintenance, and quality management. We then analyze the benchmarks and their development processes, highlighting the limitations of existing benchmarks. Additionally, we discuss the successes and failures of LLMs in different software tasks and explore future opportunities and challenges for SE-related benchmarks. We aim to provide a comprehensive overview of benchmark research in SE and offer insights to support the creation of more effective evaluation tools.

cs.SE↗

arXiv2025

The High Voltage Splitter board for the JUNO SPMT system

The Jiangmen Underground Neutrino Observatory (JUNO) in southern China is designed to study neutrinos from nuclear reactors and natural sources to address fundamental questions in neutrino physics. Achieving its goals requires continuous operation over a 20-year period. The small photomultiplier tube (small PMT or SPMT) system is a subsystem within the experiment composed of 25600 3-inch PMTs and their associated readout electronics. The High Voltage Splitter (HVS) is the first board on the readout chain of the SPMT system and services the PMTs by providing high voltage for biasing and by decoupling the generated physics signal from the high-voltage bias for readout, which is then fed to the front-end board. The necessity to handle high voltage, manage a large channel count, and operate stably for 20 years imposes significant constraints on the physical design of the HVS. This paper serves as a comprehensive documentation of the HVS board: its role in the SPMT readout system, the challenges in its design, performance and reliability metrics, and the methods employed for production and quality control.

physics.ins-det↗

arXiv2025

Materials discovery acceleration by using condition generative methodology

With the rapid advancement of AI technologies, generative models have been increasingly employed in the exploration of novel materials. By integrating traditional computational approaches such as density functional theory (DFT) and molecular dynamics (MD), existing generative models, including diffusion models and autoregressive models, have demonstrated remarkable potential in the discovery of novel materials. However, their efficiency in goal-directed materials design remains suboptimal. In this work we developed a highly transferable, efficient and robust conditional generation framework, PODGen, by integrating a general generative model with multiple property prediction models. Based on PODGen, we designed a workflow for the high-throughput crystals conditional generation which is used to search new topological insulators (TIs). Our results show that the success rate of generating TIs using our framework is 5.3 times higher than that of the unconstrained approach. More importantly, while general methods rarely produce gapped TIs, our framework succeeds consistently, highlighting an effectively $\infty$ improvement. This demonstrates that conditional generation significantly enhances the efficiency of targeted material discovery. Using this method, we generated tens of thousands of new topological materials and conducted further first-principles calculations on those with promising application potential. Furthermore, we identified promising, synthesizable topological (crystalline) insulators such as CsHgSb, NaLaB$_{12}$, Bi$_4$Sb$_2$Se$_3$, Be$_3$Ta$_2$Si and Be$_2$W.

cond-mat.mtrl-sci↗

arXiv2025

Lifts of Brauer characters in characteristic two, II

In 2007, J. P. Cossey conjectured that if $G$ is a finite $p$-solvable group and $\varphi$ is an irreducible Brauer character of $G$ with vertex $Q$, then the number of lifts of $\varphi$ is at most $|Q:Q'|$. In this paper we revisited Cossey's conjecture for $p=2$ from the perspective of Navarro vertices and obtained a new way to count the number of lifts of $\varphi$. Some applications were given.

math.GR↗

arXiv2025

Less is More: On the Importance of Data Quality for Unit Test Generation

Unit testing is crucial for software development and maintenance. Effective unit testing ensures and improves software quality, but writing unit tests is time-consuming and labor-intensive. Recent studies have proposed deep learning (DL) techniques or large language models (LLMs) to automate unit test generation. These models are usually trained or fine-tuned on large-scale datasets. Despite growing awareness of the importance of data quality, there has been limited research on the quality of datasets used for test generation. To bridge this gap, we systematically examine the impact of noise on the performance of learning-based test generation models. We first apply the open card sorting method to analyze the most popular and largest test generation dataset, Methods2Test, to categorize eight distinct types of noise. Further, we conduct detailed interviews with 17 domain experts to validate and assess the importance, reasonableness, and correctness of the noise taxonomy. Then, we propose CleanTest, an automated noise-cleaning framework designed to improve the quality of test generation datasets. CleanTest comprises three filters: a rule-based syntax filter, a rule-based relevance filter, and a model-based coverage filter. To evaluate its effectiveness, we apply CleanTest on two widely-used test generation datasets, i.e., Methods2Test and Atlas. Our findings indicate that 43.52% and 29.65% of datasets contain noise, highlighting its prevalence. Finally, we conduct comparative experiments using four LLMs (i.e., CodeBERT, AthenaTest, StarCoder, and CodeLlama7B) to assess the impact of noise on test generation performance. The results show that filtering noise positively influences the test generation ability of the models.

cs.SE↗

arXiv2025

Counting lifts of irreducible Brauer characters

Let $p$ be an odd prime, and suppose that $G$ is a $p$-solvable group and $\varphi\in {\rm IBr}(G)$ has vertex $Q$. In 2011, Cossey, Lewis and Navarro proved that the number of lifts of $\varphi$ is at most $|Q:Q'|$ whenever $Q$ is normal in $G$. In this paper, we present an explicit description of the set of lifts of $\varphi$ with a given vertex pair $(Q,\delta)$ under a weaker condition on $Q$, and thus generalize their result.

math.GR↗