How Accurately Do Large Language Models Understand Code?

Large Language Models (LLMs) are increasingly used in post-development tasks such as code repair and testing. A key factor in these tasks' success is the model's deep understanding of code. However, the extent to which LLMs truly understand code remains largely unevaluated. Quantifying code comprehension is challenging due to its abstract nature and the lack of a standardized metric. Previously, this was assessed through developer surveys, which are not feasible for evaluating LLMs. Existing LLM benchmarks focus primarily on code generation, fundamentally different from code comprehension. Additionally, fixed benchmarks quickly become obsolete as they become part of the training data. This paper presents the first large-scale empirical investigation into LLMs' ability to understand code. Inspired by mutation testing, we use an LLM's fault-finding ability as a proxy for its deep code understanding. This approach is based on the insight that a model capable of identifying subtle functional discrepancies must understand the code well. We inject faults in real-world programs and ask the LLM to localize them, ensuring the specifications suffice for fault localization. Next, we apply semantic-preserving code mutations (SPMs) to the faulty programs and test whether the LLMs still locate the faults, verifying their confidence in code understanding. We evaluate nine popular LLMs on 600,010 debugging tasks from 670 Java and 637 Python programs. We find that LLMs lose the ability to debug the same bug in 78% of faulty programs when SPMs are applied, indicating a shallow understanding of code and reliance on features irrelevant to semantics. We also find that LLMs understand code earlier in the program better than later. This suggests that LLMs' code comprehension remains tied to lexical and syntactic features due to tokenization designed for natural languages, which overlooks code semantics.

Advancing Semantic Caching for LLMs with Domain-Specific Embeddings and Synthetic Data

This report investigates enhancing semantic caching effectiveness by employing specialized, fine-tuned embedding models. Semantic caching relies on embedding similarity rather than exact key matching, presenting unique challenges in balancing precision, query latency, and computational efficiency. We propose leveraging smaller, domain-specific embedding models, fine-tuned with targeted real-world and synthetically generated datasets. Our empirical evaluations demonstrate that compact embedding models fine-tuned for just one epoch on specialized datasets significantly surpass both state-of-the-art open-source and proprietary alternatives in precision and recall. Moreover, we introduce a novel synthetic data generation pipeline for the semantic cache that mitigates the challenge of limited domain-specific annotated data, further boosting embedding performance. Our approach effectively balances computational overhead and accuracy, establishing a viable and efficient strategy for practical semantic caching implementations.

Human-in-the-Loop Synthetic Text Data Inspection with Provenance Tracking

Data augmentation techniques apply transformations to existing texts to generate additional data. The transformations may produce low-quality texts, where the meaning of the text is changed and the text may even be mangled beyond human comprehension. Analyzing the synthetically generated texts and their corresponding labels is slow and demanding. To winnow out texts with incorrect labels, we develop INSPECTOR, a human-in-the-loop data inspection technique. INSPECTOR combines the strengths of provenance tracking techniques with assistive labeling. INSPECTOR allows users to group related texts by their transformation provenance, i.e., the transformations applied to the original text, or feature provenance, the linguistic features of the original text. For assistive labeling, INSPECTOR computes metrics that approximate data quality, and allows users to compare the corresponding label of each text against the predictions of a large language model. In a user study, INSPECTOR increases the number of texts with correct labels identified by 3X on a sentiment analysis task and by 4X on a hate speech detection task. The participants found grouping the synthetically generated texts by their common transformation to be the most useful technique. Surprisingly, grouping texts by common linguistic features was perceived to be unhelpful. Contrary to prior work, our study finds that no single technique obviates the need for human inspection effort. This validates the design of INSPECTOR which combines both analysis of data provenance and assistive labeling to reduce human inspection effort.

Privacy-Aware Semantic Cache for Large Language Models

Large Language Models (LLMs) like ChatGPT, Google Bard, Claude, and Llama 2 have revolutionized natural language processing and search engine dynamics. However, these models incur exceptionally high computational costs. For instance, GPT-3 consists of 175 billion parameters and inference on these models also demands billions of floating-point operations. Caching is a natural solution to reduce LLM inference costs on repeated queries. However, existing caching methods are incapable of finding semantic similarities among LLM queries, leading to unacceptable false hit-and-miss rates. This paper introduces MeanCache, a semantic cache for LLMs that identifies semantically similar queries to determine cache hit or miss. Using MeanCache, the response to a user's semantically similar query can be retrieved from a local cache rather than re-querying the LLM, thus reducing costs, service provider load, and environmental impact. MeanCache leverages Federated Learning (FL) to collaboratively train a query similarity model in a distributed manner across numerous users without violating privacy. By placing a local cache in each user's device and using FL, MeanCache reduces the latency and costs and enhances model performance, resulting in lower cache false hit rates. Our experiments, benchmarked against the GPTCache, reveal that MeanCache attains an approximately 17% higher F-score and a 20% increase in precision during semantic cache hit-and-miss decisions. Furthermore, MeanCache reduces the storage requirement by 83% and accelerates semantic cache hit-and-miss decisions by 11%, while still surpassing GPTCache.

ProvFL: Client-Driven Interpretability of Global Model Predictions in Federated Learning

Federated Learning (FL) trains a collaborative machine learning model by aggregating multiple privately trained clients' models over several training rounds. Such a long, continuous action of model aggregations poses significant challenges in reasoning about the origin and composition of such a global model. Regardless of the quality of the global model or if it has a fault, understanding the model's origin is equally important for debugging, interpretability, and explainability in federated learning. FL application developers often question: (1) what clients contributed towards a global model and (2) if a global model predicts a label, which clients are responsible for it? We introduce, neuron provenance, a fine-grained lineage capturing mechanism that tracks the flow of information between the individual participating clients in FL and the final global model. We operationalize this concept in ProvFL that functions on two key principles. First, recognizing that monitoring every neuron of every client's model statically is ineffective and noisy due to the uninterpretable nature of individual neurons, ProvFL dynamically isolates influential and sensitive neurons in the global model, significantly reducing the search space. Second, as multiple clients' models are fused in each round to form a global model, tracking each client's contribution becomes challenging. ProvFL leverages the invertible nature of fusion algorithms to precisely isolate each client's contribution derived from selected neurons. When asked to localize the clients responsible for the given behavior (i.e., prediction) of the global model, ProvFL successfully localizes them with an average provenance accuracy of 97%. Additionally, ProvFL outperforms the state-of-the-art FL fault localization approach by an average margin of 50%.

FedDefender: Backdoor Attack Defense in Federated Learning

Federated Learning (FL) is a privacy-preserving distributed machine learning technique that enables individual clients (e.g., user participants, edge devices, or organizations) to train a model on their local data in a secure environment and then share the trained model with an aggregator to build a global model collaboratively. In this work, we propose FedDefender, a defense mechanism against targeted poisoning attacks in FL by leveraging differential testing. Our proposed method fingerprints the neuron activations of clients' models on the same input and uses differential testing to identify a potentially malicious client containing a backdoor. We evaluate FedDefender using MNIST and FashionMNIST datasets with 20 and 30 clients, and our results demonstrate that FedDefender effectively mitigates such attacks, reducing the attack success rate (ASR) to 10\% without deteriorating the global model performance.

FedDebug: Systematic Debugging for Federated Learning Applications

In Federated Learning (FL), clients train a model locally and share it with a central aggregator to build a global model. Impermissibility to access client's data and collaborative training makes FL appealing for applications with data-privacy concerns such as medical imaging. However, these FL characteristics pose unprecedented challenges for debugging. When a global model's performance deteriorates, finding the round and the clients responsible is a major pain point. Developers resort to trial-and-error debugging with subsets of clients, hoping to increase the accuracy or let future FL rounds retune the model, which are time-consuming and costly. We design a systematic fault localization framework, FedDebug, that advances the FL debugging on two novel fronts. First, FedDebug enables interactive debugging of realtime collaborative training in FL by leveraging record and replay techniques to construct a simulation that mirrors live FL. FedDebug's {\em breakpoint} can help inspect an FL state (round, client, and global model) and seamlessly move between rounds and clients' models, enabling a fine-grained step-by-step inspection. Second, FedDebug automatically identifies the client responsible for lowering global model's performance without any testing data and labels--both are essential for existing debugging techniques. FedDebug's strengths come from adapting differential testing in conjunction with neurons activations to determine the precise client deviating from normal behavior. FedDebug achieves 100\% to find a single client and 90.3\% accuracy to find multiple faulty clients. FedDebug's interactive debugging incurs 1.2\% overhead during training, while it localizes a faulty client in only 2.1\% of a round's training time. With FedDebug, we bring effective debugging practices to federated learning, improving the quality and productivity of FL application developers.

Sibylvariant Transformations for Robust Text Classification

The vast majority of text transformation techniques in NLP are inherently limited in their ability to expand input space coverage due to an implicit constraint to preserve the original class label. In this work, we propose the notion of sibylvariance (SIB) to describe the broader set of transforms that relax the label-preserving constraint, knowably vary the expected class, and lead to significantly more diverse input distributions. We offer a unified framework to organize all data transformations, including two types of SIB: (1) Transmutations convert one discrete kind into another, (2) Mixture Mutations blend two or more classes together. To explore the role of sibylvariance within NLP, we implemented 41 text transformations, including several novel techniques like Concept2Sentence and SentMix. Sibylvariance also enables a unique form of adaptive training that generates new input mixtures for the most confused class pairs, challenging the learner to differentiate with greater nuance. Our experiments on six benchmark datasets strongly support the efficacy of sibylvariance for generalization performance, defect detection, and adversarial robustness.