Search for: All records

Large Language Models (LLMs) have received much recent attention due to their human-level accuracy. While existing works mostly focus on either improving accuracy or testing accuracy robustness, the computation efficiency of LLMs, which is of paramount importance due to often vast generation demands and real-time requirements, has surprisingly received little attention. In this article, we make the first attempt to understand and test potential computation efficiency robustness in state-of-the-art LLMs. By analyzing the working mechanism and implementation of 20,543 public-accessible LLMs, we observe a fundamental property in LLMs that could be manipulated in an adversarial manner to reduce computation efficiency significantly. Our interesting observation is that the output length determines the computation efficiency of LLMs instead of the input, where the output length depends on two factors: an often sufficiently large yet pessimistic pre-configured threshold controlling the max number of iterations and a runtime-generated end of sentence (EOS) token. Our key motivation is to generate test inputs that could sufficiently delay the generation of EOS such that LLMs would have to go through enough iterations to satisfy the pre-configured threshold. We presentLLMEffiChecker, which can work under both white-box setting and black-box setting. In the white-box scenario,LLMEffiCheckerdevelops a gradient-guided technique that searches for a minimal and unnoticeable perturbation at character-level, token-level, and structure-level. In the black-box scenario,LLMEffiCheckeremploys a causal inference-based approach to find critical tokens and similarly applies three levels of imperceptible perturbation to them. Both the white-box and black-box settings effectively delay the appearance of EOS, compelling these inputs to reach the naturally unreachable threshold. To demonstrate the effectiveness ofLLMEffiChecker, we conduct a systematic evaluation on nine publicly available LLMs: Google T5, AllenAI WMT14, Helsinki-NLP translator, Facebook FairSeq, UNICAMP-DL translator, MarianMT, Google FLAN-T5, MBZUAI LaMini-GPT, and Salesforce CodeGen. Experimental results show thatLLMEffiCheckercan increase on average LLMs’ response latency and energy consumption by 325% to 3,244% and 344% to 3,616%, respectively, by perturbing just one character or token in the input sentence. Our case study shows that inputs generated byLLMEffiCheckersignificantly affect the battery power in real-world mobile devices (i.e., drain more than 30 times battery power than normal inputs). 

Natural language processing (NLP) has gained widespread adoption in the development of real-world applications. However, the black-box nature of neural networks in NLP applications poses a challenge when evaluating their performance, let alone ensuring it. Recent research has proposed testing techniques to enhance the trustworthiness of NLP-based applications. However, most existing works use a single, aggregated metric (i.e., accuracy) which is difficult for users to assess NLP model performance on fine-grained aspects, such as LCs. To address this limitation, we present ALiCT, an automated testing technique for validating NLP applications based on their LCs. ALiCT takes user-specified LCs as inputs and produces diverse test suite with test oracles for each of given LC. We evaluate ALiCT on two widely adopted NLP tasks, sentiment analysis and hate speech detection, in terms of diversity, effectiveness, and consistency. Using Self-BLEU and syntactic diversity metrics, our findings reveal that ALiCT generates test cases that are 190% and 2213% more diverse in semantics and syntax, respectively, compared to those produced by state-of-the-art techniques. In addition, ALiCT is capable of producing a larger number of NLP model failures in 22 out of 25 LCs over the two NLP applications. 

Deep learning-based code generation (DL-CG) applications have shown great potential for assisting developers in programming with human-competitive accuracy. However, lacking transparency in such applications due to the uninterpretable nature of deep learning models makes the automatically generated programs untrustworthy. In this paper, we develop DeciX, a first explanation method dedicated to DL-CG applications. DeciX is motivated by observing two unique properties of DL-CG applications: output-to-output dependencies and irrelevant value and semantic space. These properties violate the fundamental assumptions made in existing explainable DL techniques and thus cause applying existing techniques to DL-CG applications rather pessimistic and even incorrect. DeciX addresses these two limitations by constructing a causal inference dependency graph, containing a novel method leveraging causal inference that can accurately quantify the contribution of each dependency edge in the graph to the end prediction result. Proved by extensive experiments assessing popular, widely-used DL-CG applications and several baseline methods, DeciX is able to achieve significantly better performance compared to state-of-the-art in terms of several critical performance metrics, including correctness, succinctness, stability, and overhead. Furthermore, DeciX can be applied to practical scenarios since it does not require any knowledge of the DL-CG model under explanation. We have also conducted case studies that demonstrate the applicability of DeciX in practice. 

In recent times, a plethora of Large Code Generation Models (LCGMs) have been proposed, showcasing significant potential in assisting developers with complex programming tasks. Within the surge of LCGM proposals, a critical aspect of code generation research involves effectively benchmarking the programming capabilities of models. Benchmarking LCGMs necessitates the creation of a set of diverse programming problems, and each problem comprises the prompt (including the task description), canonical solution, and test inputs. The existing methods for constructing such a problem set can be categorized into two main types: manual methods and perturbation-based methods. However, %both these methods exhibit major limitations. %Firstly, manually-based methods require substantial human effort and are not easily scalable. Moreover, programming problem sets created manually struggle to maintain long-term data integrity due to the greedy training data collection mechanism in LCGMs. On the other hand, perturbation-based approaches primarily produce semantically homogeneous problems, resulting in generated programming problems with identical Canonical Solutions to the seed problem. These methods also tend to introduce typos to the prompt, easily detectable by IDEs, rendering them unrealistic. manual methods demand high effort and lack scalability, while also risking data integrity due to LCGMs' potentially contaminated data collection, and perturbation-based approaches mainly generate semantically homogeneous problems with the same canonical solutions and introduce typos that can be easily auto-corrected by IDE, making them ineffective and unrealistic. Addressing the aforementioned limitations presents several challenges: (1) How to automatically generate semantically diverse Canonical Solutions to enable comprehensive benchmarking on the models, (2) how to ensure long-term data integrity to prevent data contamination, and (3) how to generate natural and realistic programming problems. To tackle the first challenge, we draw key insights from viewing a program as a series of mappings from the input to the output domain. These mappings can be transformed, split, reordered, or merged to construct new programs. Based on this insight, we propose programming problem merging, where two existing programming problems are combined to create new ones. In addressing the second challenge, we incorporate randomness to our programming problem-generation process. Our tool can probabilistically guarantee no data repetition across two random trials. To tackle the third challenge, we propose the concept of a Lambda Programming Problem, comprising a concise one-sentence task description in natural language accompanied by a corresponding program implementation. Our tool ensures the program prompt is grammatically correct. Additionally, the tool leverages return value type analysis to verify the correctness of newly created Canonical Solutions. In our empirical evaluation, we utilize our tool on two widely-used datasets and compare it against nine baseline methods using eight code generation models. The results demonstrate the effectiveness of our tool in generating more challenging, diverse, and natural programming problems, comparing to the baselines.