Hongliang Liang

Ruikang Ni, Da Xiao, Qingye Meng et al.

Compositional relational reasoning (CRR) is a hallmark of human intelligence, but we lack a clear understanding of whether and how existing transformer large language models (LLMs) can solve CRR tasks. To enable systematic exploration of the CRR capability of LLMs, we first propose a new synthetic benchmark called Generalized Associative Recall (GAR) by integrating and generalizing the essence of several tasks in mechanistic interpretability (MI) study in a unified framework. Evaluation shows that GAR is challenging enough for existing LLMs, revealing their fundamental deficiency in CRR. Meanwhile, it is easy enough for systematic MI study. Then, to understand how LLMs solve GAR tasks, we use attribution patching to discover the core circuits reused by Vicuna-33B across different tasks and a set of vital attention heads. Intervention experiments show that the correct functioning of these heads significantly impacts task performance. Especially, we identify two classes of heads whose activations represent the abstract notion of true and false in GAR tasks respectively. They play a fundamental role in CRR across various models and tasks. The dataset and code are available at https://github.com/Caiyun-AI/GAR.

Zhuang Li, Qiuping Yi, Zongcheng Ji et al.

The rapid growth of Large Language Models (LLMs) raises concerns about distinguishing AI-generated text from human content. Existing watermarking techniques, like \kgw, struggle with low watermark strength and stringent false-positive requirements. Our analysis reveals that current methods rely on coarse estimates of non-watermarked text, limiting watermark detectability. To address this, we propose Bipolar Watermark (\tool), which splits generated text into positive and negative poles, enhancing detection without requiring additional computational resources or knowledge of the prompt. Theoretical analysis and experimental results demonstrate \tool's effectiveness and compatibility with existing optimization techniques, providing a new optimization dimension for watermarking in LLM-generated content.

Zhuang Li, Qiuping Yi, Keyang Xiao et al.

With the rise of multi-core processors and distributed systems, concurrent programming has become essential yet challenging, primarily due to the non-deterministic nature of thread execution. Manually addressing concurrency bugs is time-consuming and error-prone. Automated Program Repair techniques provide a promising solution. However, developing an end-to-end concurrency bug repair tool is particularly challenging. Most existing tools rely on the assumption that bug-related information is readily available or that concurrency bug contexts are ideally extracted, which is often impractical in real-world scenarios. This paper introduces ConFixAgent, an LLM-driven agent capable of fixing various types of concurrency bugs in an end-to-end manner, eliminating the need for any prior bug-related information. Specifically, we propose a novel context extraction approach designed for concurrency bug repair, utilizing Static Happens-Before Graphs to identify bug-relevant sections.We implemented ConFixAgent and evaluated it across multiple benchmark sets. Our extensive experiments demonstrate that ConFixAgent significantly outperforms state-of-the-art tools in addressing diverse types of concurrency bugs, with its context extraction method markedly enhancing the accuracy of LLM-generated repair solutions.

Hongliang Liang, Mingyu Li, Qiong Zhang et al.

Intel SGX provisions shielded executions for security-sensitive computation, but lacks support for trusted system services (TSS), such as clock, network and filesystem. This makes \textit{enclaves} vulnerable to Iago attacks~\cite{DBLP:conf/asplos/CheckowayS13} in the face of a powerful malicious system. To mitigate this problem, we present Aurora, a novel architecture that provides TSSes via a secure channel between enclaves and devices on top of an untrusted system, and implement two types of TSSes, i.e. clock and end-to-end network. We evaluate our solution by porting SQLite and OpenSSL into Aurora, experimental results show that SQLite benefits from a \textit{microsecond} accuracy trusted clock and OpenSSL gains end-to-end secure network with about 1ms overhead.