arXiv subjects

Haixu Tang

Publications and source records attributed to Haixu Tang.

At least 19 recordsLinked to original sources

arXiv2026

When Safety Routing Breaks: Understanding Alignment Fragility under Benign Fine-Tuning

Benign fine-tuning severely weakens the safety alignment of large language models (LLMs), so we study why refusal behavior is so fragile. While prior work often attributes this failure to gradient conflict, we propose a fundamentally different Fisher-geometric explanation: safety Fisher is low-rank, and alignment makes the safety geometry flatter while preserving an output-routing pathway. After 100 benign fine-tuning examples, this pathway is selectively re-sharpened in output-side MLP modules, explaining the asymmetric fragility: safety can collapse to high attack success rates, while general utility degrades mildly. The routing view also explains why few safety examples can restore refusal behavior, indicating that internal safety-relevant representations are preserved. Finally, we show that LoRA and ASAM mitigate early collapse by suppressing output-side sharpness, but their protection weakens at larger fine-tuning scales. Overall, safety failure is best understood as a disruption of a low-rank output-routing mechanism

cs.CR↗

arXiv2026

How to Compress KV Cache in RL Post-Training? Shadow Mask Distillation for Memory-Efficient Alignment

Reinforcement Learning (RL) has emerged as a crucial paradigm for unlocking the advanced reasoning capabilities of Large Language Models (LLMs), encompassing frameworks like RLHF and RLAIF. Regardless of the specific optimization algorithm (e.g., PPO, GRPO, or Online DPO), online RL inherently requires an exploratory trajectory generation (rollout) phase. However, for long-context reasoning tasks, this rollout phase imposes a severe ``memory wall'' due to the exorbitant Key-Value (KV) cache footprint. While applying KV cache compression during rollouts mitigates this memory overhead, it induces a critical off-policy bias. Although modern KV compression is often nearly lossless during standard inference, even minuscule approximation errors are drastically amplified by the inherent instability of RL optimization. Specifically, the sampler generates responses under a sparse context, whereas the learner updates parameters using the full, dense context. Existing statistical solutions, such as importance reweighting, struggle to correct this magnified bias, suffering from high gradient variance and severe sample inefficiency.

cs.LG↗

arXiv2026

Can LLMs Predict Polymer Physics Just by Reading Synthesis and Processing Prose?

Can large language models predict physical and mechanical polymer properties simply by reading unstructured scientific prose? Polymer performance is rarely determined by chemical structure alone; identical nominal polymers can exhibit drastically different behaviors depending on their synthesis route, processing history, morphology, and testing conditions. Yet, state-of-the-art polymer property models typically rely on structure-only representations -- such as SMILES or molecular graphs -- which strip away this vital experimental context. In this work, we introduce \textbf{PolyLM}, a natural-language-only, process- and condition-aware framework that predicts materials performance directly from full-text literature. By circumventing structural inputs entirely, PolyLM preserves the nuanced, unstructured descriptions of synthesis and processing reported by domain scientists. To train this framework, we curated an unprecedented, literature-scale dataset encompassing 185,000 scientific papers and over 276,400 unique polymer samples across 22 physical, mechanical, and thermal properties. We fine-tuned a massive 9-billion-parameter language model (Qwen3.5-9B) using Low-Rank Adaptation (LoRA) and task-level uncertainty weighting. Evaluated on 68,283 held-out observations, the model achieves remarkably high predictive accuracy, establishing new state-of-the-art benchmarks for complex properties. Across the 22 diverse targets, the model achieves a median $R^2$ of 0.74, with predictions for key thermal, mechanical, and physicochemical properties frequently surpassing an $R^2$ of 0.80. These results unequivocally demonstrate that natural language is a powerful, highly scalable interface for realistic materials performance prediction.

cs.LG↗

arXiv2026

Beyond Local vs. External: A Game-Theoretic Framework for Trustworthy Knowledge Acquisition

Cloud-hosted Large Language Models (LLMs) offer unmatched reasoning capabilities and dynamic knowledge, yet submitting raw queries to these external services risks exposing sensitive user intent. Conversely, relying exclusively on trusted local models preserves privacy but often compromises answer quality due to limited parameter scale and knowledge. To resolve this dilemma, we propose Game-theoretic Trustworthy Knowledge Acquisition (GTKA), a framework that formulates the trade-off between knowledge utility and privacy as a strategic game. GTKA consists of three components: (i) a privacy-aware sub-query generator that decomposes sensitive intent into generalized, low-risk fragments; (ii) an adversarial reconstruction attacker that attempts to infer the original query from these fragments, providing adaptive leakage signals; and (iii) a trusted local integrator that synthesizes external responses within a secure boundary. By training the generator and attacker in an alternating adversarial manner, GTKA optimizes the sub-query generation policy to maximize knowledge acquisition accuracy while minimizing the reconstructability of the original sensitive intent. To validate our approach, we construct two sensitive-domain benchmarks in the biomedical and legal fields. Extensive experiments demonstrate that GTKA significantly reduces intent leakage compared to state-of-the-art baselines while maintaining high-fidelity answer quality.

cs.CL↗

arXiv2026

PrivUn: Unveiling Latent Ripple Effects and Shallow Forgetting in Privacy Unlearning

Large language models (LLMs) often memorize private information during training, raising serious privacy concerns. While machine unlearning has emerged as a promising solution, its true effectiveness against privacy attacks remains unclear. To address this, we propose PrivUn, a new evaluation framework that systematically assesses unlearning robustness through three-tier attack scenarios: direct retrieval, in-context learning recovery, and fine-tuning restoration; combined with quantitative analysis using forgetting scores, association metrics, and forgetting depth assessment. Our study exposes significant weaknesses in current unlearning methods, revealing two key findings: 1) unlearning exhibits gradient-driven ripple effects: unlike traditional forgetting which follows semantic relations (e.g., knowledge graphs), privacy unlearning propagates across latent gradient-based associations; and 2) most methods suffer from shallow forgetting, failing to remove private information distributed across multiple deep model layers. To validate these insights, we explore two strategies: association-aware core-set selection that leverages gradient similarity, and multi-layer deep intervention through representational constraints. These strategies represent a paradigm shift from shallow forgetting to deep forgetting.

cs.LG↗

arXiv2026

A Large-Scale Comprehensive Measurement of AI-Generated Code in Real-World Repositories

Large language models (LLMs) are rapidly transforming software engineering by enabling developers to generate code ranging from small snippets to entire projects. As AI-assisted code becomes increasingly integrated into real-world systems, understanding its characteristics and impact is critical. Existing study on AI-generated code is usually limited in the lab setting with synthetic benchmarks and small-scale coding tasks and covers limited metrics. AI-assisted code's manifestation in real-world codebases and its differences between human-written one remain unclear. To close this gap, we perform a first large-scale measurement study of AI-assisted code, in comparison with the human-written, in real-world repositories. We study a comprehensive set of metrics including both code-level aspects (e.g., structural and graph-level complexity, coding style, security quality, etc.) and commit-level characteristics (e.g., commit size, frequency, post-commit stability, etc.). Our results provide new findings and insights: some contrast previous observations in the lab setting (e.g., we conclude that real-world AI-Human differences on code-level metrics are rather small instead of more pronounced), some extend prior results with finer-grained observations (e.g., the variance of security quality across different programming languages), yet more are presented for the first time on aspects not covered before (e.g., code duplication rate, commit size and stability, etc.). Based on these comprehensive real-world results, we also discuss the practical implications of AI-assisted programming.

cs.SE↗

arXiv2026

Trojans in Artificial Intelligence (TrojAI) Final Report

The Intelligence Advanced Research Projects Activity (IARPA) launched the TrojAI program to confront an emerging vulnerability in modern artificial intelligence: the threat of AI Trojans. These AI trojans are malicious, hidden backdoors intentionally embedded within an AI model that can cause a system to fail in unexpected ways, or allow a malicious actor to hijack the AI model at will. This multi-year initiative helped to map out the complex nature of the threat, pioneered foundational detection methods, and identified unsolved challenges that require ongoing attention by the burgeoning AI security field. This report synthesizes the program's key findings, including methodologies for detection through weight analysis and trigger inversion, as well as approaches for mitigating Trojan risks in deployed models. Comprehensive test and evaluation results highlight detector performance, sensitivity, and the prevalence of "natural" Trojans. The report concludes with lessons learned and recommendations for advancing AI security research.

cs.CR↗

arXiv2025

Adversarial Attack-Defense Co-Evolution for LLM Safety Alignment via Tree-Group Dual-Aware Search and Optimization

Large Language Models (LLMs) have developed rapidly in web services, delivering unprecedented capabilities while amplifying societal risks. Existing works tend to focus on either isolated jailbreak attacks or static defenses, neglecting the dynamic interplay between evolving threats and safeguards in real-world web contexts. To mitigate these challenges, we propose ACE-Safety (Adversarial Co-Evolution for LLM Safety), a novel framework that jointly optimize attack and defense models by seamlessly integrating two key innovative procedures: (1) Group-aware Strategy-guided Monte Carlo Tree Search (GS-MCTS), which efficiently explores jailbreak strategies to uncover vulnerabilities and generate diverse adversarial samples; (2) Adversarial Curriculum Tree-aware Group Policy Optimization (AC-TGPO), which jointly trains attack and defense LLMs with challenging samples via curriculum reinforcement learning, enabling robust mutual improvement. Evaluations across multiple benchmarks demonstrate that our method outperforms existing attack and defense approaches, and provides a feasible pathway for developing LLMs that can sustainably support responsible AI ecosystems.

cs.CR↗

arXiv2025

Near-Lossless Model Compression Enables Longer Context Inference in DNA Large Language Models

Trained on massive cross-species DNA corpora, DNA large language models (LLMs) learn the fundamental "grammar" and evolutionary patterns of genomic sequences. This makes them powerful priors for DNA sequence modeling, particularly over long ranges. However, two major constraints hinder their use in practice: the quadratic computational cost of self-attention and the growing memory required for key-value (KV) caches during autoregressive decoding. These constraints force the use of heuristics such as fixed-window truncation or sliding windows, which compromise fidelity on ultra-long sequences by discarding distant information. We introduce FOCUS (Feature-Oriented Compression for Ultra-long Self-attention), a progressive context-compression module that can be plugged into pretrained DNA LLMs. FOCUS combines the established k-mer representation in genomics with learnable hierarchical compression: it inserts summary tokens at k-mer granularity and progressively compresses attention key and value activations across multiple Transformer layers, retaining only the summary KV states across windows while discarding ordinary-token KV. A shared-boundary windowing scheme yields a stationary cross-window interface that propagates long-range information with minimal loss. We validate FOCUS on an Evo-2-based DNA LLM fine-tuned on GRCh38 chromosome 1 with self-supervised training and randomized compression schedules to promote robustness across compression ratios. On held-out human chromosomes, FOCUS achieves near-lossless fidelity: compressing a 1 kb context into only 10 summary tokens (about 100x) shifts the average per-nucleotide probability by only about 0.0004. Compared to a baseline without compression, FOCUS reduces KV-cache memory and converts effective inference scaling from O(N^2) to near-linear O(N), enabling about 100x longer inference windows on commodity GPUs with near-lossless fidelity.

q-bio.GN↗

arXiv2025

Characterizing Trust Boundary Vulnerabilities in TEE Containers: An Empirical Study

Trusted Execution Environments (TEEs) have become a cornerstone of confidential computing, attracting significant attention from academia and industry. To support secure and scalable application deployment on confidential clouds, TEE containers (Tcons) have been introduced as middleware to shield applications from malicious operating systems and orchestration layers while preserving usability. In this paper, we present the first comprehensive analysis of Tcons, focusing on three critical layers: OS interfaces, encrypted I/O, and orchestration mechanisms. To enable systematic evaluation, we design TBouncer, an automated analyzer that precisely exercises and benchmarks Tcon isolation boundaries. Our study uncovers fundamental flaws in existing Tcons, leading to exploitable vulnerabilities such as code execution, denial-of-service, and information leakage. In total, we identify six attack vectors, twelve new bugs, and three CVEs. These findings provide new insights into the underestimated attack surface of Tcons and highlight key directions for building more secure and trustworthy container solutions.

cs.CR↗

arXiv2025

Hey, That's My Data! Token-Only Dataset Inference in Large Language Models

Large Language Models (LLMs) rely on massive training datasets, often including proprietary data, which raises concerns about unauthorized usage and copyright infringement. Existing dataset inference methods typically require access to log probabilities or other internal signals, but many modern LLMs restrict such access, motivating token-only inference approaches. We propose CatShift, a token-only dataset inference framework based on catastrophic forgetting, where models overwrite prior knowledge when trained on new data. Fine-tuning an LLM on a subset of its training data induces larger output shifts than fine-tuning on unseen data. CatShift compares these shifts against those from a known non-member validation set to infer whether a dataset was included in training. Experiments on both open-source and API-based LLMs show that CatShift remains effective without logit access, enabling practical protection of proprietary datasets.

cs.CL↗

arXiv2025

Strategic priorities for transformative progress in advancing biology with proteomics and artificial intelligence

Artificial intelligence (AI) is transforming scientific research, including proteomics. Advances in mass spectrometry (MS)-based proteomics data quality, diversity, and scale, combined with groundbreaking AI techniques, are unlocking new challenges and opportunities in biological discovery. Here, we highlight key areas where AI is driving innovation, from data analysis to new biological insights. These include developing an AI-friendly ecosystem for proteomics data generation, sharing, and analysis; improving peptide and protein identification and quantification; characterizing protein-protein interactions and protein complexes; advancing spatial and perturbation proteomics; integrating multi-omics data; and ultimately enabling AI-empowered virtual cells.

q-bio.OT↗

arXiv2025

ByzSFL: Achieving Byzantine-Robust Secure Federated Learning with Zero-Knowledge Proofs

The advancement of AI models, especially those powered by deep learning, faces significant challenges in data-sensitive industries like healthcare and finance due to the distributed and private nature of data. Federated Learning (FL) and Secure Federated Learning (SFL) enable collaborative model training without data sharing, enhancing privacy by encrypting shared intermediate results. However, SFL currently lacks effective Byzantine robustness, a critical property that ensures model performance remains intact even when some participants act maliciously. Existing Byzantine-robust methods in FL are incompatible with SFL due to the inefficiency and limitations of encryption operations in handling complex aggregation calculations. This creates a significant gap in secure and robust model training. To address this gap, we propose ByzSFL, a novel SFL system that achieves Byzantine-robust secure aggregation with high efficiency. Our approach offloads aggregation weight calculations to individual parties and introduces a practical zero-knowledge proof (ZKP) protocol toolkit. This toolkit supports widely used operators for calculating aggregation weights, ensuring correct computations without compromising data privacy. Not only does this method maintain aggregation integrity, but it also significantly boosts computational efficiency, making ByzSFL approximately 100 times faster than existing solutions. Furthermore, our method aligns with open-source AI trends, enabling plaintext publication of the final model without additional information leakage, thereby enhancing the practicality and robustness of SFL in real-world applications.

cs.CR↗

arXiv2025

Enhancing Patient-Centric Communication: Leveraging LLMs to Simulate Patient Perspectives

Large Language Models (LLMs) have demonstrated impressive capabilities in role-playing scenarios, particularly in simulating domain-specific experts using tailored prompts. This ability enables LLMs to adopt the persona of individuals with specific backgrounds, offering a cost-effective and efficient alternative to traditional, resource-intensive user studies. By mimicking human behavior, LLMs can anticipate responses based on concrete demographic or professional profiles. In this paper, we evaluate the effectiveness of LLMs in simulating individuals with diverse backgrounds and analyze the consistency of these simulated behaviors compared to real-world outcomes. In particular, we explore the potential of LLMs to interpret and respond to discharge summaries provided to patients leaving the Intensive Care Unit (ICU). We evaluate and compare with human responses the comprehensibility of discharge summaries among individuals with varying educational backgrounds, using this analysis to assess the strengths and limitations of LLM-driven simulations. Notably, when LLMs are primed with educational background information, they deliver accurate and actionable medical guidance 88% of the time. However, when other information is provided, performance significantly drops, falling below random chance levels. This preliminary study shows the potential benefits and pitfalls of automatically generating patient-specific health information from diverse populations. While LLMs show promise in simulating health personas, our results highlight critical gaps that must be addressed before they can be reliably used in clinical settings. Our findings suggest that a straightforward query-response model could outperform a more tailored approach in delivering health information. This is a crucial first step in understanding how LLMs can be optimized for personalized health communication while maintaining accuracy.

cs.AI↗

arXiv2024

LLM-Enhanced Software Patch Localization

Open source software (OSS) is integral to modern product development, and any vulnerability within it potentially compromises numerous products. While developers strive to apply security patches, pinpointing these patches among extensive OSS updates remains a challenge. Security patch localization (SPL) recommendation methods are leading approaches to address this. However, existing SPL models often falter when a commit lacks a clear association with its corresponding CVE, and do not consider a scenario that a vulnerability has multiple patches proposed over time before it has been fully resolved. To address these challenges, we introduce LLM-SPL, a recommendation-based SPL approach that leverages the capabilities of the Large Language Model (LLM) to locate the security patch commit for a given CVE. More specifically, we propose a joint learning framework, in which the outputs of LLM serves as additional features to aid our recommendation model in prioritizing security patches. Our evaluation on a dataset of 1,915 CVEs associated with 2,461 patches demonstrates that LLM-SPL excels in ranking patch commits, surpassing the state-of-the-art method in terms of Recall, while significantly reducing manual effort. Notably, for vulnerabilities requiring multiple patches, LLM-SPL significantly improves Recall by 22.83\%, NDCG by 19.41\%, and reduces manual effort by over 25\% when checking up to the top 10 rankings. The dataset and source code are available at \url{https://anonymous.4open.science/r/LLM-SPL-91F8}.

cs.CR↗

arXiv2024

DPAdapter: Improving Differentially Private Deep Learning through Noise Tolerance Pre-training

Recent developments have underscored the critical role of \textit{differential privacy} (DP) in safeguarding individual data for training machine learning models. However, integrating DP oftentimes incurs significant model performance degradation due to the perturbation introduced into the training process, presenting a formidable challenge in the {differentially private machine learning} (DPML) field. To this end, several mitigative efforts have been proposed, typically revolving around formulating new DPML algorithms or relaxing DP definitions to harmonize with distinct contexts. In spite of these initiatives, the diminishment induced by DP on models, particularly large-scale models, remains substantial and thus, necessitates an innovative solution that adeptly circumnavigates the consequential impairment of model utility. In response, we introduce DPAdapter, a pioneering technique designed to amplify the model performance of DPML algorithms by enhancing parameter robustness. The fundamental intuition behind this strategy is that models with robust parameters are inherently more resistant to the noise introduced by DP, thereby retaining better performance despite the perturbations. DPAdapter modifies and enhances the sharpness-aware minimization (SAM) technique, utilizing a two-batch strategy to provide a more accurate perturbation estimate and an efficient gradient descent, thereby improving parameter robustness against noise. Notably, DPAdapter can act as a plug-and-play component and be combined with existing DPML algorithms to further improve their performance. Our experiments show that DPAdapter vastly enhances state-of-the-art DPML algorithms, increasing average accuracy from 72.92\% to 77.09\% with a privacy budget of $\epsilon=4$.

cs.LG↗

arXiv2023

The Janus Interface: How Fine-Tuning in Large Language Models Amplifies the Privacy Risks

The rapid advancements of large language models (LLMs) have raised public concerns about the privacy leakage of personally identifiable information (PII) within their extensive training datasets. Recent studies have demonstrated that an adversary could extract highly sensitive privacy data from the training data of LLMs with carefully designed prompts. However, these attacks suffer from the model's tendency to hallucinate and catastrophic forgetting (CF) in the pre-training stage, rendering the veracity of divulged PIIs negligible. In our research, we propose a novel attack, Janus, which exploits the fine-tuning interface to recover forgotten PIIs from the pre-training data in LLMs. We formalize the privacy leakage problem in LLMs and explain why forgotten PIIs can be recovered through empirical analysis on open-source language models. Based upon these insights, we evaluate the performance of Janus on both open-source language models and two latest LLMs, i.e., GPT-3.5-Turbo and LLaMA-2-7b. Our experiment results show that Janus amplifies the privacy risks by over 10 times in comparison with the baseline and significantly outperforms the state-of-the-art privacy extraction attacks including prefix attacks and in-context learning (ICL). Furthermore, our analysis validates that existing fine-tuning APIs provided by OpenAI and Azure AI Studio are susceptible to our Janus attack, allowing an adversary to conduct such an attack at a low cost.

cs.CR↗

arXiv2023

Large Language Model Soft Ideologization via AI-Self-Consciousness

Large language models (LLMs) have demonstrated human-level performance on a vast spectrum of natural language tasks. However, few studies have addressed the LLM threat and vulnerability from an ideology perspective, especially when they are increasingly being deployed in sensitive domains, e.g., elections and education. In this study, we explore the implications of GPT soft ideologization through the use of AI-self-consciousness. By utilizing GPT self-conversations, AI can be granted a vision to "comprehend" the intended ideology, and subsequently generate finetuning data for LLM ideology injection. When compared to traditional government ideology manipulation techniques, such as information censorship, LLM ideologization proves advantageous; it is easy to implement, cost-effective, and powerful, thus brimming with risks.

cs.CL↗