arXiv subjects

Mengzhuo Chen

Publications and source records attributed to Mengzhuo Chen.

10 recordsLinked to original sources

arXiv2026

Measuring the Creativity of Frontier LLMs in Automated Research

Frontier LLMs are increasingly capable of conducting automated research, yet their creativity in this setting has not been systematically evaluated. In this paper, we propose a set of metrics to evaluate creativity along the two dimensions of valueness and novelty. Valueness assesses whether each proposed idea is useful, while novelty is evaluated from three perspectives: whether the same idea has appeared before (Exact-Match P-Novelty), whether a previously unexplored variable or variable combination is explored (Variable-level P-Novelty), and whether the idea directly follows retrieved external knowledge or departs from it (H-Novelty). Our evaluation shows that the models achieve relatively similar scores on most creativity metrics, but differ substantially in Variable-level P-Novelty, which reflects the breadth of research-space exploration. Further correlation and idea-level performance analyses show that Variable-level P-Novelty is the creativity dimension most strongly associated with research performance.

cs.CL↗

arXiv2026

From Failed Trajectories to Reliable LLM Agents: Diagnosing and Repairing Harness Flaws

LLM agents increasingly rely on agent harness: the runtime infrastructure around the base model that defines execution environments, tool interfaces, context, lifecycle orchestration, observability, verification, and governance. Existing self-improving agents and automatic harness evolution methods mainly improve agents through runtime supervision, prompt optimization, workflow search, or harness modification based on final outcomes. However, they often fail to diagnose where the responsible evidence lies in failed trajectories and which harness implementation mechanism causes the unreliable behavior, resulting in broad, indirect, or poorly scoped changes. This paper proposes HarnessFix, a trace-grounded and diagnosis-driven framework for repairing agent harnesses. HarnessFix compiles raw execution traces and harness artifacts into a Harness-aware Trace Intermediate Representation (HTIR), which normalizes fragmented trajectory evidence, captures step-level data-flow and control-flow relations, and aligns runtime steps with the harness artifacts that shape their behavior. It then attributes failures to responsible steps and harness artifacts, and consolidates recurring diagnoses into repair-oriented flaw records. Finally, HarnessFix maps these records to scoped repair operators, generates patches under flaw-specific repair specifications, and accepts them through regression-aware validation. We evaluate HarnessFix on four popular benchmarks, and results show that it improves the performance over the initial harnesses by 6.3% to 18.4%, significantly outperforming human-designed and self-evolution baselines. HarnessFix highlights the value of treating failed trajectories not only as feedback signals, but also as structured evidence for diagnosing and repairing the harness mechanisms behind agent failures.

cs.SE↗

arXiv2026

Hybrid Open-Ended Tri-Evolution Makes Better Deep Researcher

Deep research and agent evolution serve as de-facto tasks for AI agents in real-world applications toward artificial general intelligence. The former enables autonomous retrieval and integration of information in open-ended environments to tackle open-ended research tasks, yet it is constrained by the static parametric deep research capabilities of agent systems. The latter allows agents to autonomously interact with the environment to gain experiences that evolve model capabilities. However, its effectiveness has been widely validated only on verifiable tasks with standard answers, leaving a gap with open-ended research tasks. To bridge these two critical tasks, we propose the Hybrid Open-Ended Tri-Evolution (HOTE) framework, which leverages hybrid-mode reinforcement learning to facilitate the collaborative evolution of a proposer, solver and judge based on web-scale knowledge, moving toward autonomous evolving agents in open-ended tasks and environments. Extensive experiments on three long-form deep research benchmarks demonstrate that the 8B model trained via HOTE surpasses the strongest static open 8-32B models as well as those trained by state-of-the-art deep research training methods with less time overhead, and further verify that the evolution of all three modules in HOTE is indispensable.

cs.AI↗

arXiv2026

UniReason-Med: A Shared Grounded Reasoning Interface for 2D-to-3D Transfer in Medical VQA

We study whether grounded reasoning supervision from abundant 2D medical images can improve 3D medical VQA when both input types are aligned through a common reasoning interface. We introduce UniReason-Med, a single-checkpoint framework that processes either a 2D image or a slice-serialized 3D volume at inference time, generating interleaved textual reasoning and localized visual evidence through shared box syntax, region-token injection, and a common grounded reasoning policy. To train this interface, we construct UniMed-CoT, a 220K instruction-tuning dataset with interleaved textual reasoning and grounded visual evidence, including 170K 2D and 50K 3D samples. Through supervised fine-tuning followed by outcome-level reinforcement learning, UniReason-Med learns to generate grounded reasoning traces without IoU/Dice-based localization rewards during RL. Data-mixture and component ablations show that joint 2D+3D grounded supervision substantially improves 3D reasoning over 3D-only training, while grounding and region-token injection consistently benefit both 2D and 3D tasks. These results suggest that a shared grounded reasoning interface can transfer reasoning structure from 2D images to slice-serialized volumetric medical understanding. The code and data are publicly available at https://github.com/IQuestLab/unireason-med.

cs.CV↗

arXiv2026

Seeing the Whole Elephant: A Benchmark for Failure Attribution in LLM-based Multi-Agent Systems

Failure attribution, i.e., identifying the responsible agent and decisive step of a failure, is particularly challenging in LLM-based multi-agent systems (MAS) due to their natural-language reasoning, nondeterministic outputs, and intricate interaction dynamics. A reliable benchmark is therefore essential to guide and evaluate attribution techniques. Yet existing benchmarks rely on partially observable traces that capture only agent outputs, omitting the inputs and context that developers actually use when debugging. We argue that failure attribution should be studied under full execution observability, aligning with real-world developer-facing scenarios where complete traces, rather than only outputs, are accessible for diagnosis. To this end, we introduce TraceElephant, a benchmark designed for failure attribution with full execution traces and reproducible environments. We then systematically evaluate failure attribution techniques across various configurations. Specifically, full traces improve attribution accuracy by up to 76\% over a partial-observation counterpart, confirming that missing inputs obscure many failure causes. TraceElephant provides a foundation for follow-up failure attribution research, promoting evaluation practices that reflect real-world debugging and supporting the development of more transparent MASs.

cs.MA↗

arXiv2024

Seeing is Believing: Vision-driven Non-crash Functional Bug Detection for Mobile Apps

Mobile app GUI (Graphical User Interface) pages now contain rich visual information, with the visual semantics of each page helping users understand the application logic. However, these complex visual and functional logic present new challenges to software testing. Existing automated GUI testing methods, constrained by the lack of reliable testing oracles, are limited to detecting crash bugs with obvious abnormal signals. Consequently, many non-crash functional bugs, ranging from unexpected behaviors to logical errors, often evade detection by current techniques. While these non-crash functional bugs can exhibit visual cues that serve as potential testing oracles, they often entail a sequence of screenshots, and detecting them necessitates an understanding of the operational logic among GUI page transitions, which is challenging traditional techniques. Considering the remarkable performance of Multimodal Large Language Models (MLLM) in visual and language understanding, this paper proposes Trident, a novel vision-driven, multi-agent collaborative automated GUI testing approach for detecting non-crash functional bugs. It comprises three agents: Explorer, Monitor, and Detector, to guide the exploration, oversee the testing progress, and spot issues. We also address several challenges, i.e., align visual and textual information for MLLM input, achieve functionality-oriented exploration, and infer test oracles for non-crash bugs, to enhance the performance of functionality bug detection. We evaluate Trident on 590 non-crash bugs and compare it with 12 baselines, it can achieve more than 14%-112% and 108%-147% boost in average recall and precision compared with the best baseline. The ablation study further proves the contribution of each module. Moreover, Trident identifies 43 new bugs on Google Play, of which 31 have been fixed.

cs.SE↗

arXiv2024

Unblind Text Inputs: Predicting Hint-text of Text Input in Mobile Apps via LLM

Mobile apps have become indispensable for accessing and participating in various environments, especially for low-vision users. Users with visual impairments can use screen readers to read the content of each screen and understand the content that needs to be operated. Screen readers need to read the hint-text attribute in the text input component to remind visually impaired users what to fill in. Unfortunately, based on our analysis of 4,501 Android apps with text inputs, over 0.76 of them are missing hint-text. These issues are mostly caused by developers' lack of awareness when considering visually impaired individuals. To overcome these challenges, we developed an LLM-based hint-text generation model called HintDroid, which analyzes the GUI information of input components and uses in-context learning to generate the hint-text. To ensure the quality of hint-text generation, we further designed a feedback-based inspection mechanism to further adjust hint-text. The automated experiments demonstrate the high BLEU and a user study further confirms its usefulness. HintDroid can not only help visually impaired individuals, but also help ordinary people understand the requirements of input components. HintDroid demo video: https://youtu.be/FWgfcctRbfI.

cs.HC↗

arXiv2023

Testing the Limits: Unusual Text Inputs Generation for Mobile App Crash Detection with Large Language Model

Mobile applications have become a ubiquitous part of our daily life, providing users with access to various services and utilities. Text input, as an important interaction channel between users and applications, plays an important role in core functionality such as search queries, authentication, messaging, etc. However, certain special text (e.g., -18 for Font Size) can cause the app to crash, and generating diversified unusual inputs for fully testing the app is highly demanded. Nevertheless, this is also challenging due to the combination of explosion dilemma, high context sensitivity, and complex constraint relations. This paper proposes InputBlaster which leverages the LLM to automatically generate unusual text inputs for mobile app crash detection. It formulates the unusual inputs generation problem as a task of producing a set of test generators, each of which can yield a batch of unusual text inputs under the same mutation rule. In detail, InputBlaster leverages LLM to produce the test generators together with the mutation rules serving as the reasoning chain, and utilizes the in-context learning schema to demonstrate the LLM with examples for boosting the performance. InputBlaster is evaluated on 36 text input widgets with cash bugs involving 31 popular Android apps, and results show that it achieves 78% bug detection rate, with 136% higher than the best baseline. Besides, we integrate it with the automated GUI testing tool and detect 37 unseen crashes in real-world apps from Google Play.

cs.SE↗

arXiv2023

Make LLM a Testing Expert: Bringing Human-like Interaction to Mobile GUI Testing via Functionality-aware Decisions

Automated Graphical User Interface (GUI) testing plays a crucial role in ensuring app quality, especially as mobile applications have become an integral part of our daily lives. Despite the growing popularity of learning-based techniques in automated GUI testing due to their ability to generate human-like interactions, they still suffer from several limitations, such as low testing coverage, inadequate generalization capabilities, and heavy reliance on training data. Inspired by the success of Large Language Models (LLMs) like ChatGPT in natural language understanding and question answering, we formulate the mobile GUI testing problem as a Q&A task. We propose GPTDroid, asking LLM to chat with the mobile apps by passing the GUI page information to LLM to elicit testing scripts, and executing them to keep passing the app feedback to LLM, iterating the whole process. Within this framework, we have also introduced a functionality-aware memory prompting mechanism that equips the LLM with the ability to retain testing knowledge of the whole process and conduct long-term, functionality-based reasoning to guide exploration. We evaluate it on 93 apps from Google Play and demonstrate that it outperforms the best baseline by 32% in activity coverage, and detects 31% more bugs at a faster rate. Moreover, GPTDroid identify 53 new bugs on Google Play, of which 35 have been confirmed and fixed.

cs.SE↗

arXiv2023

Chatting with GPT-3 for Zero-Shot Human-Like Mobile Automated GUI Testing

Mobile apps are indispensable for people's daily life, and automated GUI (Graphical User Interface) testing is widely used for app quality assurance. There is a growing interest in using learning-based techniques for automated GUI testing which aims at generating human-like actions and interactions. However, the limitations such as low testing coverage, weak generalization, and heavy reliance on training data, make an urgent need for a more effective approach to generate human-like actions to thoroughly test mobile apps. Inspired by the success of the Large Language Model (LLM), e.g., GPT-3 and ChatGPT, in natural language understanding and question answering, we formulate the mobile GUI testing problem as a Q&A task. We propose GPTDroid, asking LLM to chat with the mobile apps by passing the GUI page information to LLM to elicit testing scripts, and executing them to keep passing the app feedback to LLM, iterating the whole process. Within it, we extract the static context of the GUI page and the dynamic context of the iterative testing process, design prompts for inputting this information to LLM, and develop a neural matching network to decode the LLM's output into actionable steps to execute the app. We evaluate GPTDroid on 86 apps from Google Play, and its activity coverage is 71%, with 32% higher than the best baseline, and can detect 36% more bugs with faster speed than the best baseline. GPTDroid also detects 48 new bugs on the Google Play with 25 of them being confirmed/fixed. We further summarize the capabilities of GPTDroid behind the superior performance, including semantic text input, compound action, long meaningful test trace, and test case prioritization.

cs.SE↗