Chenguang Zhu

Ming Zhong, Aston Zhang, Xuewei Wang et al. · meta-ai

The development and evaluation of Large Language Models (LLMs) have largely focused on individual capabilities. However, this overlooks the intersection of multiple abilities across different types of expertise that are often required for real-world tasks, which we term cross capabilities. To systematically explore this concept, we first define seven core individual capabilities and then pair them to form seven common cross capabilities, each supported by a manually constructed taxonomy. Building on these definitions, we introduce CrossEval, a benchmark comprising 1,400 human-annotated prompts, with 100 prompts for each individual and cross capability. To ensure reliable evaluation, we involve expert annotators to assess 4,200 model responses, gathering 8,400 human ratings with detailed explanations to serve as reference examples. Our findings reveal that, in both static evaluations and attempts to enhance specific abilities, current LLMs consistently exhibit the "Law of the Weakest Link," where cross-capability performance is significantly constrained by the weakest component. Specifically, across 58 cross-capability scores from 17 models, 38 scores are lower than all individual capabilities, while 20 fall between strong and weak, but closer to the weaker ability. These results highlight the under-performance of LLMs in cross-capability tasks, making the identification and improvement of the weakest capabilities a critical priority for future research to optimize performance in complex, multi-dimensional scenarios.

Yulong Chen, Yang Liu, Jianhao Yan et al. · cambridge, tencent-ai

The impressive performance of Large Language Models (LLMs) has consistently surpassed numerous human-designed benchmarks, presenting new challenges in assessing the shortcomings of LLMs. Designing tasks and finding LLMs' limitations are becoming increasingly important. In this paper, we investigate the question of whether an LLM can discover its own limitations from the errors it makes. To this end, we propose a Self-Challenge evaluation framework with human-in-the-loop. Starting from seed instances that GPT-4 fails to answer, we prompt GPT-4 to summarize error patterns that can be used to generate new instances and incorporate human feedback on them to refine these patterns for generating more challenging data, iteratively. We end up with 8 diverse patterns, such as text manipulation and questions with assumptions. We then build a benchmark, SC-G4, consisting of 1,835 instances generated by GPT-4 using these patterns, with human-annotated gold responses. The SC-G4 serves as a challenging benchmark that allows for a detailed assessment of LLMs' abilities. Our results show that only 44.96\% of instances in SC-G4 can be answered correctly by GPT-4. Interestingly, our pilot study indicates that these error patterns also challenge other LLMs, such as Claude-3 and Llama-3, and cannot be fully resolved through fine-tuning. Our work takes the first step to demonstrate that LLMs can autonomously identify their inherent flaws and provide insights for future dynamic and automatic evaluation.

Yue Yu, Zhengxing Chen, Aston Zhang et al. · allen-ai

Reward modeling is crucial for aligning large language models (LLMs) with human preferences, especially in reinforcement learning from human feedback (RLHF). However, current reward models mainly produce scalar scores and struggle to incorporate critiques in a natural language format. We hypothesize that predicting both critiques and the scalar reward would improve reward modeling ability. Motivated by this, we propose Critic-RM, a framework that improves reward models using self-generated critiques without extra supervision. Critic-RM employs a two-stage process: generating and filtering high-quality critiques, followed by joint fine-tuning on reward prediction and critique generation. Experiments across benchmarks show that Critic-RM improves reward modeling accuracy by 3.7%-7.3% compared to standard reward models and LLM judges, demonstrating strong performance and data efficiency. Additional studies further validate the effectiveness of generated critiques in rectifying flawed reasoning steps with 2.5%-3.2% gains in improving reasoning accuracy.

Shan Jiang, Chenguang Zhu, Sarfraz Khurshid

Software testing remains the most widely used methodology for validating quality of code. However, effectiveness of testing critically depends on the quality of test suites used. Test cases in a test suite consist of two fundamental parts: (1) input values for the code under test, and (2) correct checks for the outputs it produces. These checks are commonly written as assertions, and termed test oracles. The last couple of decades have seen much progress in automated test input generation, e.g., using fuzzing and symbolic execution. However, automating test oracles remains a relatively less explored problem area. Indeed, a test oracle by its nature requires knowledge of expected behavior, which may only be known to the developer and may not not exist in a formal language that supports automated reasoning. Our focus in this paper is automation of test oracles for clients of widely used Java libraries, e.g., java.lang and java.util packages. Our key insight is that Javadocs that provide a rich source of information can enable automated generation of test oracles. Javadocs of the core Java libraries are fairly detailed documents that contain natural language descriptions of not only how the libraries behave but also how the clients must (not) use them. We use large language models as an enabling technology to embody our insight into a framework for test oracle automation, and evaluate it experimentally. Our experiments demonstrate that LLMs can generate oracles for checking normal and exceptional behaviors from Javadocs, with 98.8% of these oracles being compilable and 96.4% accurately reflecting intended properties. Even for the few incorrect oracles, errors are minor and can be easily corrected with the help of additional comment information generated by the LLMs.

Shan Jiang, Chenguang Zhu, Sarfraz Khurshid

JavaScript obfuscators are widely deployed to protect intellectual property and resist reverse engineering, yet their correctness has been largely overlooked compared to performance and resilience. Existing evaluations typically measure resistance to deobfuscation, leaving the critical question of whether obfuscators preserve program semantics unanswered. Incorrect transformations can silently alter functionality, compromise reliability, and erode security-undermining the very purpose of obfuscation. To address this gap, we present OBsmith, a novel framework to systematically test JavaScript obfuscators using large language models (LLMs). OBsmith leverages LLMs to generate program sketches abstract templates capturing diverse language constructs, idioms, and corner cases-which are instantiated into executable programs and subjected to obfuscation under different configurations. Besides LLM-powered sketching, OBsmith also employs a second source: automatic extraction of sketches from real programs. This extraction path enables more focused testing of project specific features and lets developers inject domain knowledge into the resulting test cases. OBsmith uncovers 11 previously unknown correctness bugs. Under an equal program budget, five general purpose state-of-the-art JavaScript fuzzers (FuzzJIT, Jsfunfuzz, Superion, DIE, Fuzzilli) failed to detect these issues, highlighting OBsmith's complementary focus on obfuscation induced misbehavior. An ablation shows that all components except our generic MRs contribute to at least one bug class; the negative MR result suggests the need for obfuscator-specific metamorphic relations. Our results also seed discussion on how to balance obfuscation presets and performance cost. We envision OBsmith as an important step towards automated testing and quality assurance of obfuscators and other semantic-preserving toolchains.

Ripon K. Saha, Akira Ura, Sonal Mahajan et al.

Automatic machine learning, or AutoML, holds the promise of truly democratizing the use of machine learning (ML), by substantially automating the work of data scientists. However, the huge combinatorial search space of candidate pipelines means that current AutoML techniques, generate sub-optimal pipelines, or none at all, especially on large, complex datasets. In this work we propose an AutoML technique SapientML, that can learn from a corpus of existing datasets and their human-written pipelines, and efficiently generate a high-quality pipeline for a predictive task on a new dataset. To combat the search space explosion of AutoML, SapientML employs a novel divide-and-conquer strategy realized as a three-stage program synthesis approach, that reasons on successively smaller search spaces. The first stage uses a machine-learned model to predict a set of plausible ML components to constitute a pipeline. In the second stage, this is then refined into a small pool of viable concrete pipelines using syntactic constraints derived from the corpus and the machine-learned model. Dynamically evaluating these few pipelines, in the third stage, provides the best solution. We instantiate SapientML as part of a fully automated tool-chain that creates a cleaned, labeled learning corpus by mining Kaggle, learns from it, and uses the learned models to then synthesize pipelines for new predictive tasks. We have created a training corpus of 1094 pipelines spanning 170 datasets, and evaluated SapientML on a set of 41 benchmark datasets, including 10 new, large, real-world datasets from Kaggle, and against 3 state-of-the-art AutoML tools and 2 baselines. Our evaluation shows that SapientML produces the best or comparable accuracy on 27 of the benchmarks while the second best tool fails to even produce a pipeline on 9 of the instances.