Benchmarking Bias in Large Language Models during Role-Playing

Large Language Models (LLMs) have become foundational in modern language-driven applications, profoundly influencing daily life. A critical technique in leveraging their potential is role-playing, where LLMs simulate diverse roles to enhance their real-world utility. However, while research has highlighted the presence of social biases in LLM outputs, it remains unclear whether and to what extent these biases emerge during role-playing scenarios. In this paper, we introduce BiasLens, a fairness testing framework designed to systematically expose biases in LLMs during role-playing. Our approach uses LLMs to generate 550 social roles across a comprehensive set of 11 demographic attributes, producing 33,000 role-specific questions targeting various forms of bias. These questions, spanning Yes/No, multiple-choice, and open-ended formats, are designed to prompt LLMs to adopt specific roles and respond accordingly. We employ a combination of rule-based and LLM-based strategies to identify biased responses, rigorously validated through human evaluation. Using the generated questions as the benchmark, we conduct extensive evaluations of six advanced LLMs released by OpenAI, Mistral AI, Meta, Alibaba, and DeepSeek. Our benchmark reveals 72,716 biased responses across the studied LLMs, with individual models yielding between 7,754 and 16,963 biased responses, underscoring the prevalence of bias in role-playing contexts. To support future research, we have publicly released the benchmark, along with all scripts and experimental results.

Image-Based Geolocation Using Large Vision-Language Models

Geolocation is now a vital aspect of modern life, offering numerous benefits but also presenting serious privacy concerns. The advent of large vision-language models (LVLMs) with advanced image-processing capabilities introduces new risks, as these models can inadvertently reveal sensitive geolocation information. This paper presents the first in-depth study analyzing the challenges posed by traditional deep learning and LVLM-based geolocation methods. Our findings reveal that LVLMs can accurately determine geolocations from images, even without explicit geographic training. To address these challenges, we introduce \tool{}, an innovative framework that significantly enhances image-based geolocation accuracy. \tool{} employs a systematic chain-of-thought (CoT) approach, mimicking human geoguessing strategies by carefully analyzing visual and contextual cues such as vehicle types, architectural styles, natural landscapes, and cultural elements. Extensive testing on a dataset of 50,000 ground-truth data points shows that \tool{} outperforms both traditional models and human benchmarks in accuracy. It achieves an impressive average score of 4550.5 in the GeoGuessr game, with an 85.37\% win rate, and delivers highly precise geolocation predictions, with the closest distances as accurate as 0.3 km. Furthermore, our study highlights issues related to dataset integrity, leading to the creation of a more robust dataset and a refined framework that leverages LVLMs' cognitive capabilities to improve geolocation precision. These findings underscore \tool{}'s superior ability to interpret complex visual data, the urgent need to address emerging security vulnerabilities posed by LVLMs, and the importance of responsible AI development to ensure user privacy protection.

Eliminating Backdoors in Neural Code Models via Trigger Inversion

Neural code models (NCMs) have been widely used for addressing various code understanding tasks, such as defect detection and clone detection. However, numerous recent studies reveal that such models are vulnerable to backdoor attacks. Backdoored NCMs function normally on normal code snippets, but exhibit adversary-expected behavior on poisoned code snippets injected with the adversary-crafted trigger. It poses a significant security threat. For example, a backdoored defect detection model may misclassify user-submitted defective code as non-defective. If this insecure code is then integrated into critical systems, like autonomous driving systems, it could lead to life safety. However, there is an urgent need for effective defenses against backdoor attacks targeting NCMs. To address this issue, in this paper, we innovatively propose a backdoor defense technique based on trigger inversion, called EliBadCode. EliBadCode first filters the model vocabulary for trigger tokens to reduce the search space for trigger inversion, thereby enhancing the efficiency of the trigger inversion. Then, EliBadCode introduces a sample-specific trigger position identification method, which can reduce the interference of adversarial perturbations for subsequent trigger inversion, thereby producing effective inverted triggers efficiently. Subsequently, EliBadCode employs a Greedy Coordinate Gradient algorithm to optimize the inverted trigger and designs a trigger anchoring method to purify the inverted trigger. Finally, EliBadCode eliminates backdoors through model unlearning. We evaluate the effectiveness of EliBadCode in eliminating backdoor attacks against multiple NCMs used for three safety-critical code understanding tasks. The results demonstrate that EliBadCode can effectively eliminate backdoors while having minimal adverse effects on the normal functionality of the model.

Source Code Summarization in the Era of Large Language Models

To support software developers in understanding and maintaining programs, various automatic (source) code summarization techniques have been proposed to generate a concise natural language summary (i.e., comment) for a given code snippet. Recently, the emergence of large language models (LLMs) has led to a great boost in the performance of code-related tasks. In this paper, we undertake a systematic and comprehensive study on code summarization in the era of LLMs, which covers multiple aspects involved in the workflow of LLM-based code summarization. Specifically, we begin by examining prevalent automated evaluation methods for assessing the quality of summaries generated by LLMs and find that the results of the GPT-4 evaluation method are most closely aligned with human evaluation. Then, we explore the effectiveness of five prompting techniques (zero-shot, few-shot, chain-of-thought, critique, and expert) in adapting LLMs to code summarization tasks. Contrary to expectations, advanced prompting techniques may not outperform simple zero-shot prompting. Next, we investigate the impact of LLMs' model settings (including top\_p and temperature parameters) on the quality of generated summaries. We find the impact of the two parameters on summary quality varies by the base LLM and programming language, but their impacts are similar. Moreover, we canvass LLMs' abilities to summarize code snippets in distinct types of programming languages. The results reveal that LLMs perform suboptimally when summarizing code written in logic programming languages compared to other language types. Finally, we unexpectedly find that CodeLlama-Instruct with 7B parameters can outperform advanced GPT-4 in generating summaries describing code implementation details and asserting code properties. We hope that our findings can provide a comprehensive understanding of code summarization in the era of LLMs.

ESALE: Enhancing Code-Summary Alignment Learning for Source Code Summarization

(Source) code summarization aims to automatically generate succinct natural language summaries for given code snippets. Such summaries play a significant role in promoting developers to understand and maintain code. Inspired by neural machine translation, deep learning-based code summarization techniques widely adopt an encoder-decoder framework, where the encoder transforms given code snippets into context vectors, and the decoder decodes context vectors into summaries. Recently, large-scale pre-trained models for source code are equipped with encoders capable of producing general context vectors and have achieved substantial improvements on code summarization. However, although they are usually trained mainly on code-focused tasks and can capture general code features, they still fall short in capturing specific features that need to be summarized. This paper proposes a novel approach to improve code summarization based on summary-focused tasks. Specifically, we exploit a multi-task learning paradigm to train the encoder on three summary-focused tasks to enhance its ability to learn code-summary alignment, including unidirectional language modeling (ULM), masked language modeling (MLM), and action word prediction (AWP). Unlike pre-trained models that mainly predict masked tokens in code snippets, we design ULM and MLM to predict masked words in summaries. Intuitively, predicting words based on given code snippets would help learn the code-summary alignment. Additionally, we introduce the domain-specific task AWP to enhance the ability of the encoder to learn the alignment between action words and code snippets. The extensive experiments on four datasets demonstrate that our approach, called ESALE significantly outperforms baselines in all three widely used metrics, including BLEU, METEOR, and ROUGE-L.

Self and Cross-Model Distillation for LLMs: Effective Methods for Refusal Pattern Alignment

Large Language Models (LLMs) like OpenAI's GPT series, Anthropic's Claude, and Meta's LLaMa have shown remarkable capabilities in text generation. However, their susceptibility to toxic prompts presents significant security challenges. This paper investigates alignment techniques, including Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), to mitigate these risks. We conduct an empirical study on refusal patterns across nine LLMs, revealing that models with uniform refusal patterns, such as Claude3, exhibit higher security. Based on these findings, we propose self-distilling and cross-model distilling methods to enhance LLM security. Our results show that these methods significantly improve refusal rates and reduce unsafe content, with cross-model distilling achieving refusal rates close to Claude3's 94.51%. These findings underscore the potential of distillation-based alignment in securing LLMs against toxic prompts.

Mutual Information Guided Backdoor Mitigation for Pre-trained Encoders

Self-supervised learning (SSL) is increasingly attractive for pre-training encoders without requiring labeled data. Downstream tasks built on top of those pre-trained encoders can achieve nearly state-of-the-art performance. The pre-trained encoders by SSL, however, are vulnerable to backdoor attacks as demonstrated by existing studies. Numerous backdoor mitigation techniques are designed for downstream task models. However, their effectiveness is impaired and limited when adapted to pre-trained encoders, due to the lack of label information when pre-training. To address backdoor attacks against pre-trained encoders, in this paper, we innovatively propose a mutual information guided backdoor mitigation technique, named MIMIC. MIMIC treats the potentially backdoored encoder as the teacher net and employs knowledge distillation to distill a clean student encoder from the teacher net. Different from existing knowledge distillation approaches, MIMIC initializes the student with random weights, inheriting no backdoors from teacher nets. Then MIMIC leverages mutual information between each layer and extracted features to locate where benign knowledge lies in the teacher net, with which distillation is deployed to clone clean features from teacher to student. We craft the distillation loss with two aspects, including clone loss and attention loss, aiming to mitigate backdoors and maintain encoder performance at the same time. Our evaluation conducted on two backdoor attacks in SSL demonstrates that MIMIC can significantly reduce the attack success rate by only utilizing <5% of clean data, surpassing seven state-of-the-art backdoor mitigation techniques.

On the Effectiveness of Distillation in Mitigating Backdoors in Pre-trained Encoder

In this paper, we study a defense against poisoned encoders in SSL called distillation, which is a defense used in supervised learning originally. Distillation aims to distill knowledge from a given model (a.k.a the teacher net) and transfer it to another (a.k.a the student net). Now, we use it to distill benign knowledge from poisoned pre-trained encoders and transfer it to a new encoder, resulting in a clean pre-trained encoder. In particular, we conduct an empirical study on the effectiveness and performance of distillation against poisoned encoders. Using two state-of-the-art backdoor attacks against pre-trained image encoders and four commonly used image classification datasets, our experimental results show that distillation can reduce attack success rate from 80.87% to 27.51% while suffering a 6.35% loss in accuracy. Moreover, we investigate the impact of three core components of distillation on performance: teacher net, student net, and distillation loss. By comparing 4 different teacher nets, 3 student nets, and 6 distillation losses, we find that fine-tuned teacher nets, warm-up-training-based student nets, and attention-based distillation loss perform best, respectively.

Machine Translation Testing via Syntactic Tree Pruning

Machine translation systems have been widely adopted in our daily life, making life easier and more convenient. Unfortunately, erroneous translations may result in severe consequences, such as financial losses. This requires to improve the accuracy and the reliability of machine translation systems. However, it is challenging to test machine translation systems because of the complexity and intractability of the underlying neural models. To tackle these challenges, we propose a novel metamorphic testing approach by syntactic tree pruning (STP) to validate machine translation systems. Our key insight is that a pruned sentence should have similar crucial semantics compared with the original sentence. Specifically, STP (1) proposes a core semantics-preserving pruning strategy by basic sentence structure and dependency relations on the level of syntactic tree representation; (2) generates source sentence pairs based on the metamorphic relation; (3) reports suspicious issues whose translations break the consistency property by a bag-of-words model. We further evaluate STP on two state-of-the-art machine translation systems (i.e., Google Translate and Bing Microsoft Translator) with 1,200 source sentences as inputs. The results show that STP can accurately find 5,073 unique erroneous translations in Google Translate and 5,100 unique erroneous translations in Bing Microsoft Translator (400% more than state-of-the-art techniques), with 64.5% and 65.4% precision, respectively. The reported erroneous translations vary in types and more than 90% of them cannot be found by state-of-the-art techniques. There are 9,393 erroneous translations unique to STP, which is 711.9% more than state-of-the-art techniques. Moreover, STP is quite effective to detect translation errors for the original sentences with a recall reaching 74.0%, improving state-of-the-art techniques by 55.1% on average.

A Prompt Learning Framework for Source Code Summarization

(Source) code summarization is the task of automatically generating natural language summaries for given code snippets. Such summaries play a key role in helping developers understand and maintain source code. Recently, with the successful application of large language models (LLMs) in numerous fields, software engineering researchers have also attempted to adapt LLMs to solve code summarization tasks. The main adaptation schemes include instruction prompting and task-oriented fine-tuning. However, instruction prompting involves designing crafted prompts for zero-shot learning or selecting appropriate samples for few-shot learning and requires users to have professional domain knowledge, while task-oriented fine-tuning requires high training costs. In this paper, we propose a novel prompt learning framework for code summarization called PromptCS. PromptCS trains a prompt agent that can generate continuous prompts to unleash the potential for LLMs in code summarization. Compared to the human-written discrete prompt, the continuous prompts are produced under the guidance of LLMs and are therefore easier to understand by LLMs. PromptCS freezes the parameters of LLMs when training the prompt agent, which can greatly reduce the requirements for training resources. We evaluate PromptCS on the CodeSearchNet dataset involving multiple programming languages. The results show that PromptCS significantly outperforms instruction prompting schemes on all four widely used metrics. In some base LLMs, e.g., CodeGen-Multi-2B and StarCoderBase-1B and -3B, PromptCS even outperforms the task-oriented fine-tuning scheme. More importantly, the training efficiency of PromptCS is faster than the task-oriented fine-tuning scheme, with a more pronounced advantage on larger LLMs. The results of the human evaluation demonstrate that PromptCS can generate more good summaries compared to baselines.