Yifeng He

Xiangyi Li, Wenbo Chen, Yimin Liu et al. · berkeley

Agent Skills are structured packages of procedural knowledge that augment LLM agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark of 86 tasks across 11 domains paired with curated Skills and deterministic verifiers. Each task is evaluated under three conditions: no Skills, curated Skills, and self-generated Skills. We test 7 agent-model configurations over 7,308 trajectories. Curated Skills raise average pass rate by 16.2 percentage points(pp), but effects vary widely by domain (+4.5pp for Software Engineering to +51.9pp for Healthcare) and 16 of 84 tasks show negative deltas. Self-generated Skills provide no benefit on average, showing that models cannot reliably author the procedural knowledge they benefit from consuming. Focused Skills with 2--3 modules outperform comprehensive documentation, and smaller models with Skills can match larger models without them.

Yifeng He, Ethan Wang, Jicheng Wang et al.

Test-time scaling has emerged as a promising approach for improving code generation by exploring large solution spaces at inference time. However, existing methods often rely on public test cases that are unavailable in practice, or require extensive LLM inference for candidate selection, leading to significant token consumption and time overhead. We present DiffCodeGen, a novel test-time scaling method for code generation based on coverage-guided differential analysis. DiffCodeGen generates diverse code candidates using various sampling and prompting strategies, then applies coverage-guided fuzzing to synthesize inputs without requiring any existing tests or large language models. By executing all candidates on these inputs, DiffCodeGen captures their dynamic behavior and clusters candidates based on behavioral similarity. DiffCodeGen selects the medoid of the largest cluster as the final output. Unlike prior test-time scaling methods that invoke additional LLM inference for candidate selection, DiffCodeGen performs selection without any extra model calls, incurring little to no additional token consumption. DiffCodeGen is fully asynchronous, naturally suited to the current trend of agentic coding, and is thus efficient and highly scalable. We evaluate DiffCodeGen across 4 large language models, demonstrating consistent improvements over baselines. Compared to state-of-the-art test-time scaling methods, DiffCodeGen achieves competitive or superior performance while using only a fraction of time and tokens. DiffCodeGen is model-agnostic and can be combined with reasoning models to further boost performance.

Jicheng Wang, Yifeng He, Zili Wang et al.

Tool-augmented LLM agents call APIs whose intermediate outputs, such as presigned URLs, session tokens, and OAuth state parameters, are observation contracts: artifacts whose later use is constrained by the external system that produced them. We show that observation contract compliance (preserving the temporal validity and byte-level integrity) is an emergent, regression-prone capability: it is neither guaranteed by general tool-use ability nor consistently improved by larger or newer models. To measure this, we introduce ContractBench, a benchmark of 33 dual-axis tasks that probe two orthogonal failure modes no existing benchmark evaluates: validity failures (using an artifact after expiry) and integrity failures (corrupting an artifact's bytes through the observation-to-action pipeline). Our evaluation is deterministic and programmatic, with a virtual clock controlling time and SHA-256 hashes verifying byte integrity. We assign each outcome a failure label drawn from real-world API specifications. We evaluate 38 models and report four findings: (i) no evaluated model clears 80%, with Claude-Opus-4.6 leading at 77.8%, revealing that current frontier models still fail to comply with observation contracts; (ii) a sharp within-family capability cliff in Qwen 3.5 between 4B (0%) and 9B (56.6%), smoothing to 70.7% at 397B-A17B: what emerges across the cliff is mid-trajectory restraint, not tool-call competence; (iii) non-monotonic scaling across the GPT-5 family: agentic post-training can erode compliance through sycophancy-driven regression; (iv) our failure taxonomy works as an actionable in-context reward signal, yielding +7.1 pp on 42 paired GPT-5.1 failures.

Shuang Chen, Yifeng He, Barry Lennox et al.

Long-term monitoring and exploration of extreme environments, such as underwater storage facilities, is costly, labor-intensive, and hazardous. Automating this process with low-cost, collaborative robots can greatly improve efficiency. These robots capture images from different positions, which must be processed simultaneously to create a spatio-temporal model of the facility. In this paper, we propose a novel approach that integrates data simulation, a multi-modal deep learning network for coordinate prediction, and image reassembly to address the challenges posed by environmental disturbances causing drift and rotation in the robots' positions and orientations. Our approach enhances the precision of alignment in noisy environments by integrating visual information from snapshots, global positional context from masks, and noisy coordinates. We validate our method through extensive experiments using synthetic data that simulate real-world robotic operations in underwater settings. The results demonstrate very high coordinate prediction accuracy and plausible image assembly, indicating the real-world applicability of our approach. The assembled images provide clear and coherent views of the underwater environment for effective monitoring and inspection, showcasing the potential for broader use in extreme settings, further contributing to improved safety, efficiency, and cost reduction in hazardous field monitoring. Code is available on https://github.com/ChrisChen1023/Micro-Robot-Swarm.

Jianyu Zhao, Yuyang Rong, Yiwen Guo et al.

Semantic understanding of programs has attracted great attention in the community. Inspired by recent successes of large language models (LLMs) in natural language understanding, tremendous progress has been made by treating programming language as another sort of natural language and training LLMs on corpora of program code. However, programs are essentially different from texts after all, in a sense that they are normally heavily structured and syntax-strict. In particular, programs and their basic units (i.e., functions and subroutines) are designed to demonstrate a variety of behaviors and/or provide possible outputs, given different inputs. The relationship between inputs and possible outputs/behaviors represents the functions/subroutines and profiles the program as a whole. Therefore, we propose to incorporate such a relationship into learning, for achieving a deeper semantic understanding of programs. To obtain inputs that are representative enough to trigger the execution of most part of the code, we resort to fuzz testing and propose fuzz tuning to boost the performance of program understanding and code representation learning, given a pre-trained LLM. The effectiveness of the proposed method is verified on two program understanding tasks including code clone detection and code classification, and it outperforms current state-of-the-arts by large margins. Code is available at https://github.com/rabbitjy/FuzzTuning.

Yifeng He, Jiabo Huang, Yuyang Rong et al.

The remarkable capability of large language models (LLMs) in generating high-quality code has drawn increasing attention in the software testing community. However, existing code LLMs often demonstrate unsatisfactory capabilities in generating accurate and complete tests since they were trained on code snippets collected without differentiating between code for testing purposes and other code. In this paper, we present a large-scale dataset UniTSyn, which is capable of enhancing the prowess of LLMs for Unit Test Synthesis. Associating tests with the tested functions is crucial for LLMs to infer the expected behavior and the logic paths to be verified. By leveraging Language Server Protocol, UniTSyn achieves the challenging goal of collecting focal-test pairs without per-project execution setups or per-language heuristics that tend to be fragile and difficult to scale. It contains 2.7 million focal-test pairs across five mainstream programming languages, making it possible to be utilized for enhancing the test generation ability of LLMs. The details of UniTSyn can be found in Table 1. Our experiments demonstrate that, by building an autoregressive model based on UniTSyn, we can achieve significant benefits in learning and understanding unit test representations, resulting in improved generation accuracy and code coverage across all evaluated programming languages. Code and data will be publicly available.

Yifeng He, Ziye Tang, Hao Chen

Information cocoons on social media limit users' exposure to posts with diverse viewpoints. Modern platforms use stance detection as an important signal in recommendation and ranking pipelines, which can route posts primarily to like-minded audiences and reduce cross-cutting exposure. This restricts the reach of dissenting opinions and hinders constructive discourse. We take the creator's perspective and investigate how content can be revised to reach beyond existing affinity clusters. We present ContentFuzz, a confidence-guided fuzzing framework that rewrites posts while preserving their human-interpreted intent and induces different machine-inferred stance labels. ContentFuzz aims to route posts beyond their original cocoons. Our method guides a large language model (LLM) to generate meaning-preserving rewrites using confidence feedback from stance detection models. Evaluated on four representative stance detection models across three datasets in two languages, ContentFuzz effectively changes machine-classified stance labels, while maintaining semantic integrity with respect to the original content.

Yifeng He, Luning Yang, Christopher Castro Gaw Gonzalo et al.

Large Language Models (LLMs) are increasingly integrated into the software engineering ecosystem. Their test-time compute (TTC) reasoning capabilities show significant potential for understanding program logic and semantics beyond mere token recognition. However, current benchmarks for code reasoning lack a formal, program-centric deductive framework to ensure sound evaluation, and are incapable of assessing whether models genuinely reason about program semantics or merely exploit superficial associations between natural language and code tokens. To bridge this gap, we introduce TF-Bench, a benchmark designed to evaluate LLM reasoning based on type inference in System F, a task we refer to as program semantics reasoning. By employing verified transformations to remove semantically irrelevant natural language, we construct TF-Bench_pure, a purely semantics-driven variant of TF-Bench. Our analysis reveals substantial limitations in state-of-the-art LLMs, with the best-performing LLM (Claude-3.7-sonnet) achieving only 55.85% accuracy on TF-Bench_pure. Additionally, we propose two novel metrics to assess robustness and the effectiveness of test-time reasoning, underscoring critical limitations in current LLM capabilities and highlighting essential directions for future research.

Yuchun Shu, Bo Hu, Yifeng He et al.

Accurately finding the wrong words in the automatic speech recognition (ASR) hypothesis and recovering them well-founded is the goal of speech error correction. In this paper, we propose a non-autoregressive speech error correction method. A Confidence Module measures the uncertainty of each word of the N-best ASR hypotheses as the reference to find the wrong word position. Besides, the acoustic feature from the ASR encoder is also used to provide the correct pronunciation references. N-best candidates from ASR are aligned using the edit path, to confirm each other and recover some missing character errors. Furthermore, the cross-attention mechanism fuses the information between error correction references and the ASR hypothesis. The experimental results show that both the acoustic and confidence references help with error correction. The proposed system reduces the error rate by 21% compared with the ASR model.

Yifeng He, Ethan Wang, Yuyang Rong et al.

AI agents have been boosted by large language models. AI agents can function as intelligent assistants and complete tasks on behalf of their users with access to tools and the ability to execute commands in their environments. Through studying and experiencing the workflow of typical AI agents, we have raised several concerns regarding their security. These potential vulnerabilities are not addressed by the frameworks used to build the agents, nor by research aimed at improving the agents. In this paper, we identify and describe these vulnerabilities in detail from a system security perspective, emphasizing their causes and severe effects. Furthermore, we introduce defense mechanisms corresponding to each vulnerability with design and experiments to evaluate their viability. Altogether, this paper contextualizes the security issues in the current development of AI agents and delineates methods to make AI agents safer and more reliable.

Yifeng He, Jicheng Wang, Yuyang Rong et al.

Testing is essential to modern software engineering for building reliable software. Given the high costs of manually creating test cases, automated test case generation, particularly methods utilizing large language models, has become increasingly popular. These neural approaches generate semantically meaningful tests that are more maintainable compared with traditional automatic testing methods like fuzzing. However, the diversity and volume of unit tests in current datasets are limited, especially for newer but important languages. In this paper, we present a novel data augmentation technique, FuzzAug, that introduces the benefits of fuzzing to large language models by introducing valid testing semantics and providing diverse coverage-guided inputs. Doubling the size of training datasets, FuzzAug improves the performance from the baselines significantly. This technique demonstrates the potential of introducing prior knowledge from dynamic software analysis to improve neural test generation, offering significant enhancements in neural test generation.

Hongxiang Zhang, Yuyang Rong, Yifeng He et al.

Greybox fuzzing has achieved success in revealing bugs and vulnerabilities in programs. However, randomized mutation strategies have limited the fuzzer's performance on structured data. Specialized fuzzers can handle complex structured data, but require additional efforts in grammar and suffer from low throughput. In this paper, we explore the potential of utilizing the Large Language Model to enhance greybox fuzzing for structured data. We utilize the pre-trained knowledge of LLM about data conversion and format to generate new valid inputs. We further fine-tuned it with paired mutation seeds to learn structured format and mutation strategies effectively. Our LLM-based fuzzer, LLAMAFUZZ, integrates the power of LLM to understand and mutate structured data to fuzzing. We conduct experiments on the standard bug-based benchmark Magma and a wide variety of real-world programs. LLAMAFUZZ outperforms our top competitor by 41 bugs on average. We also identified 47 unique bugs across all trials. Moreover, LLAMAFUZZ demonstrated consistent performance on both bug trigger and bug reached. Compared to AFL++, LLAMAFUZZ achieved 27.19% more branches in real-world program sets on average. We also demonstrate a case study to explain how LLMs enhance the fuzzing process in terms of code coverage.

Jiabo Huang, Jianyu Zhao, Yuyang Rong et al.

Large language models (LLMs) for natural language processing have been grafted onto programming language modeling for advancing code intelligence. Although it can be represented in the text format, code is syntactically more rigorous in order to be properly compiled or interpreted to perform a desired set of behaviors given any inputs. In this case, existing works benefit from syntactic representations to learn from code less ambiguously in the forms of abstract syntax tree, control-flow graph, etc. However, programs with the same purpose can be implemented in various ways showing different syntactic representations while the ones with similar implementations can have distinct behaviors. Though trivially demonstrated during executions, such semantics about functionality are challenging to be learned directly from code, especially in an unsupervised manner. Hence, in this paper, we propose FuzzPretrain to explore the dynamic information of programs revealed by their test cases and embed it into the feature representations of code as complements. The test cases are obtained with the assistance of a customized fuzzer and are only required during pre-training. FuzzPretrain yielded more than 6%/9% mAP improvements on code search over its counterparts trained with only source code or AST, respectively. Our extensive experimental results show the benefits of learning discriminative code representations with program executions.

Thanh Hong-Phuoc, Yifeng He, Ling Guan

All current popular hand-crafted key-point detectors such as Harris corner, MSER, SIFT, SURF... rely on some specific pre-designed structures for the detection of corners, blobs, or junctions in an image. In this paper, a novel sparse coding based key-point detector which requires no particular pre-designed structures is presented. The key-point detector is based on measuring the complexity level of each block in an image to decide where a key-point should be. The complexity level of a block is defined as the total number of non-zero components of a sparse representation of that block. Generally, a block constructed with more components is more complex and has greater potential to be a good key-point. Experimental results on Webcam and EF datasets [1, 2] show that the proposed detector achieves significantly high repeatability compared to hand-crafted features, and even outperforms the matching scores of the state-of-the-art learning based detector.