arXiv subjects

Zihan Wu

Publications and source records attributed to Zihan Wu.

At least 19 recordsLinked to original sources

arXiv2026

When to Compile a Computer-Use Agent? Measuring Payback and Making Compilation Decisions for Token Efficiency

Compiling GUI procedures that agents execute repeatedly into programs can reduce their token costs. However, measuring payback and deciding when to compile have two challenges. First, compilation costs are uncertain because attempts can require repair and still fail to produce a usable program. Second, future reuse is unknown because tasks may stop arriving or GUI drift may stop the program from working. To address these challenges, we propose PACE (Payback-Aware Compilation from Experience), a system with a measurement protocol and an online compilation algorithm. The measurement protocol records successful and failed compilation costs, and compares agent and program execution costs on matched task inputs to estimate per-use savings and payback counts. Using these measurements, the online algorithm compares estimated future savings with compilation costs, including failed attempts, based on past task arrivals and compilation outcomes. It checks execution and compilation charges against a cumulative budget determined by observed task arrivals before allowing either action. Under stated action-cost assumptions, total cost after each arrival is at most $1+ε$ times the cost of running every task with the agent. For successful compilation attempts, estimated payback counts excluding source agent runs are 2-16 uses. In simulations using recorded task arrivals, PACE reduces token costs by 17.3% compared with ReAct, 24.9% with the AutoRPA adaptation, and 17.3% with the ToolPro adaptation on average ($ε=0.25$).

cs.AI↗

arXiv2026

On the Capacity of DNA Labeling in the Single-Label Setting

DNA labeling has attracted increasing attention in biomedical applications, including molecular imaging, diagnostics, and genomic analysis. In a DNA labeling process, a set of DNA sequence patterns, referred to as labels, is designed according to the requirements of a specific application. For each DNA sequence, the labeling process generates an output sequence that records the positions of the labels. DNA sequences with different labeling outputs can therefore be distinguished through the labeling process. To quantify this capability, the labeling capacity is defined as the exponential growth rate of the maximum number of DNA sequences that can be distinguished through the labeling process as the sequence length tends to infinity [2]. To date, the labeling capacities of several cases in the single-label setting have been determined. In this paper, we formulate the labeling process as a deterministic channel and show that its zero-error capacity is equal to the labeling capacity. For a single label, the corresponding channel can be represented by a star graph. Thus, characterizing the labeling capacity of a single label is equivalent to determining the zero-error capacity of the corresponding star graph. We derive the zero-error capacities of all star graphs, thereby providing a complete characterization of the labeling capacities for all single-label cases. Furthermore, we develop a general method for constructing capacity-achieving codes. These results apply to labeling problems over arbitrary finite alphabets and are not restricted to the DNA alphabet. Finally, for a fixed label length, we exactly characterize the range of achievable labeling capacities and identify all single-label structures that attain the minimum and maximum capacities.

cs.IT↗

arXiv2026

VidAct: Learning Manipulation from In-the-Wild Videos with Object-Centric 3D Awareness

Video demonstrations offer a scalable alternative to costly robot data for learning manipulation, yet existing reconstruction-based approaches often rely on constrained camera viewpoints or human-to-robot retargeting, while the reconstructed trajectories are difficult to adapt to new objects configurations without distorting the trajectory shape. Another key limitation is that the resulting policies often lack precise object-level 3D geometry awareness, limiting object grounding and object shape awareness critical for precise manipulation. To bridge these gaps, we propose VidAct, an efficient video-to-robot framework that learns object-centric, 3D-aware manipulation policies from a single monocular video per task and enables zero-shot real-world deployment. VidAct consists of three key components. First, VidAct reconstructs object meshes and motion from arbitrary demo videos and canonicalizes the motion in the static object frame, avoiding embodiment-specific retargeting and accommodating diverse camera viewpoints. Second, VidAct employ residual trajectory transfer for adapting the reconstructed motion to novel object configurations while preserving its motion shape. Finally, as the key policy-learning component, VidAct predicts simulation-provided privileged complete-object point clouds at each frame as an auxiliary task while retaining RGB-only deployment, providing dense object-centric supervision over both object pose and 3D geometry. Experiments on human, robot, generated, and internet videos demonstrate broad video applicability and zero-shot deployment. Per-frame complete-object 3D supervision improves policy generalization and sim-to-real success, while residual trajectory transfer enables reliable trajectory adaptation with better shape preservation.

cs.RO↗

arXiv2026

LoLBench: Evaluating Coding Agents with Long-Horizon Proposals on Large Software Systems

Modern coding agents can deliver increasingly large repository-level changes, and recent benchmarks reflect this by emphasizing long-horizon tasks with large reference implementations. Many benchmarks evaluate coding agents' implementation capability to produce correct code edits from detailed specifications. However, practical modular development tasks also require the perception capability of grounding user intent and high-level design to derive a specification. We introduce LoLBench to evaluate both capabilities through the entire proposal-to-implementation process on large software systems. It is a multilingual benchmark of 100 tasks across 29 software systems in five domains. Each task provides a human-written enhancement proposal with user intent and high-level design. On average, proposals contain about 5,000 words, software systems contain 2.4 million source lines of code (LoC), and implementation pull requests (PRs) change approximately 5,500 LoC. Across 28 agents we evaluated, the best agent resolves only 14% of tasks and achieves a 52.7% Fail-to-Pass (F2P) pass rate. Failure analysis identifies incomplete code localization as a major bottleneck, while providing reference-derived file trees alongside API specifications improves resolved rates by 16--22 percentage points (2.4--17$\times$), reaching at most 34%. These results show that both perception and implementation remain central challenges for coding agents in practical modular development on large software systems. LoLBench is available at https://huggingface.co/datasets/lolbench26/LoLBench.

cs.SE↗

arXiv2026

A Plan-Tracing Interface for AI-Supported Algorithm Planning

Planning an algorithm in natural language allows learners to get formative feedback on their approach before coding. But these descriptions can be ambiguous, making it challenging for learners to translate them into code and for LLMs to provide feedback. To address these challenges, we introduce plan tracing, in which learners manually simulate how the described algorithm would execute on a concrete input. We develop an interface that enables plan tracing and allows learners to receive AI feedback on their plans before writing code. We report on an exploratory between-subjects study with 20 participants who solved an algorithm design task, with or without the plan tracing interface. We observed how plan tracing shaped students' plans, the feedback they received, and their experiences using the interface. Students who performed plan tracing wrote plans with fewer code-like steps but more goal-driven descriptions. We did not detect a difference in the quality of the LLM feedback between conditions. Students used plan tracing to debug and verify their strategy, describing it as tedious but worthwhile when they were uncertain about their solution. We reflect on the design and use of our tool, identifying what worked, what didn't, and why, and offer recommendations for instructors and tool designers.

cs.HC↗

arXiv2026

A Deterministic $(2+\varepsilon)$-Approximation for Weighted Feedback Vertex Set in Tournaments

We study the weighted feedback vertex set problem in tournaments. For every fixed integer $k\geq 2$, we give a deterministic $(2+1/k)$-approximation algorithm with running time $n^{2^{O(k)}}$, apart from polynomial dependence on the encoding length of the weights. Consequently, for every fixed $\varepsilon>0$, weighted feedback vertex set in tournaments has a deterministic $(2+\varepsilon)$-approximation running in time $n^{2^{O(1/\varepsilon)}}$. The algorithm combines two ingredients. When the triangle graph of the tournament has bounded clique number, a chain decomposition of its transitive complement yields an exact dynamic program for a maximum-weight transitive subtournament. When the clique number is large, a structural theorem for triangle graphs supplies a constant-size strongly good cost vector. A local-ratio reduction with this cost vector gives the claimed guarantee. As a by-product, the dynamic program solves weighted feedback vertex set exactly in $\mathcal B_7$-free tournaments in time $O(n^7)$, where $\mathcal B_7$ is the family of seven-vertex tournaments with feedback vertex set number at least three.

cs.DS↗

arXiv2026

A Unified Issue Resolution Benchmark for Requirement Clarification, Planning, and Code Generation for Coding Agents

Large language model-powered coding agents are increasingly used to modify existing code repositories, for example, by adding features or fixing bugs. Yet existing repository-level benchmarks typically evaluate only whether the final patch passes tests. Satisfying a user request requires a long chain of interdependent reasoning and decisions: an agent must recover explicit and implicit requirements, formulate a repository-grounded implementation plan, and translate it into correct code. A pass/fail outcome cannot characterize how an unsuccessful trajectory diverges from the requirements and implementation process needed for a correct patch. To address this gap, we introduce SWE-RPG, a repository-level benchmark that combines executable patch evaluation with validated ground-truth references (GTs) for (1) Requirement Clarification and (2) Implementation Planning. These intermediate GTs support retrospective, GT-aligned diagnosis of complete coding-agent trajectories across clarification, planning, code generation, and artifact submission. SWE-RPG comprises 163 tasks from 31 Python and Java repositories, including 113 bug fixes and 50 feature additions. We evaluate 3 coding agents, including Claude Code, Codex, and OpenCode, with 6 large language model backends, including Claude-Sonnet-5 and GPT-5.6-Terra. Results show that the evaluated popular coding agents still struggle to implement user requests in existing repositories, achieving an average resolved rate of only 31.5% on SWE-RPG. Intermediate-GT diagnosis further identifies implicit requirement recovery as the main bottleneck, accounting for 24.5%--46.0% of agent runs. This result suggests implicit-requirement recovery as a key candidate direction for improving coding agents. The benchmark data and evaluation code are available at https://github.com/Xin-Zhou-smu/SWE-RPG-Bench.

cs.SE↗

arXiv2026

Exploring the Design Space of LLM-Based Programming Support in CS Education: A Scoping Review through the Lens of Assistance Governance

As large language models (LLMs) become integrated into programming education, learner-facing systems increasingly differ in how that assistance is bounded, enacted, and controlled. These governance decisions are often described implicitly, making it difficult to compare systems in educationally meaningful ways. To address this gap, we conduct a scoping review and qualitative synthesis of 90 peer-reviewed LLM-based programming support systems in CS education. We analyze assistance governance through three dimensions, which we refer to collectively as PEA: Policy, capturing what forms of help are allowed or restricted; Enforcement, capturing how those boundaries are operationalized through interaction and system behavior; and Authority, capturing who can configure, adapt, or override them during use. Our findings show that systems often share similar pedagogical goals, but implement those goals through varied enforcement mechanisms. At the same time, authority remains highly centralized in system logic, with fewer systems giving learners or instructors runtime control. This work contributes PEA as a three-dimensional analytic lens, a governance codebook empirically refined within these dimensions, and a map of underexplored configurations in the current design space of LLM-based programming support. By making these explicit and comparable, PEA offers a vocabulary for analyzing existing systems and designing future tools that are pedagogically bounded, configurable, and accountable.

cs.HC↗

arXiv2026

What Resolve Rate Hides: Trajectory Structure Diagnostics for Coding Agents

Coding agents are ranked almost entirely by resolve rate: whether their final patch passes the target tests. Yet two agents can reach the same outcome through very different processes, and a single pass/fail label says nothing about why a run failed or why an accepted run spent extra steps, time, or tokens. This process evidence lives in the trajectory, which records a run's searches, reads, edits, tool calls, validation, and reversions. However, raw traces are heterogeneous and hard to compare across runs. We present TraceProbe, a trajectory-diagnostic framework that recovers what resolve rate hides. TraceProbe normalizes each raw run into a canonical nine-type action taxonomy with deterministic effect labels, then applies two rule-based modules: Insight names single-trajectory anti-patterns adapted from established debugging practice (e.g., search loops, verification skips), while Converge aligns pairs of runs and classifies where their behavior diverges under controlled references. Applying TraceProbe to 2,500 trajectories from five production settings on SWE-Bench Verified, we find that (i) file choice is too coarse to separate success from failure, whereas function selection and completion behavior localize it; (ii) Insight anti-patterns act mainly as corpus-level difficulty clues, with search loops the most stable; and (iii) even resolved runs differ in how quickly they reach relevant code and how much failed work they incur. Trajectory structure thus adds auditable diagnostic context to outcomes by localizing inspection targets, suggesting failure hypotheses, and prioritizing runs for review.

cs.SE↗

arXiv2026

P$^2$RAG: Efficient Privacy-Preserving RAG Service Supporting Arbitrary Top-$k$ Retrieval

Retrieval-Augmented Generation (RAG) enables large language models to use external knowledge, but outsourcing the RAG service raises privacy concerns for both data owners and users. Privacy-preserving RAG systems address these concerns by performing secure top-$k$ retrieval, which is typically implemented using secure sorting to identify relevant documents. However, existing systems face challenges supporting arbitrary $k$ due to their inability to change $k$, new security issues, and in particular, efficiency degradation with large $k$. This is a significant limitation because applications such as finance, law, and healthcare require a $k$ that is large enough to cause huge overhead for existing systems. Also, modern long-context models generally achieve higher accuracy with larger retrieval sets. We propose P$^2$RAG, an efficient privacy-preserving RAG service that supports arbitrary top-$k$ retrieval. Unlike existing systems, P$^2$RAG avoids sorting candidate documents. Instead, it uses an interactive bisection method to determine the set of top-$k$ documents. For security, P$^2$RAG uses secret sharing on two semi-honest non-colluding servers to protect the data owner's database and the user's prompt. It enforces restrictions and verification to defend against malicious users and tightly bounds the information leakage of the database. The experiments show that P$^2$RAG is 3--300$\times$ faster than the state-of-the-art PRAG for $k = 16$--$1024$.

cs.CR↗

arXiv2026

MulVul: Retrieval-augmented Multi-Agent Code Vulnerability Detection via Cross-Model Prompt Evolution

Large Language Models (LLMs) struggle to automate real-world vulnerability detection due to two key limitations: the heterogeneity of vulnerability patterns undermines the effectiveness of a single unified model, and manual prompt engineering for massive weakness categories is unscalable. To address these challenges, we propose \textbf{MulVul}, a retrieval-augmented multi-agent framework designed for precise and broad-coverage vulnerability detection. MulVul adopts a coarse-to-fine strategy: a \emph{Router} agent first predicts the top-$k$ coarse categories and then forwards the input to specialized \emph{Detector} agents, which identify the exact vulnerability types. Both agents are equipped with retrieval tools to actively source evidence from vulnerability knowledge bases to mitigate hallucinations. Crucially, to automate the generation of specialized prompts, we design \emph{Cross-Model Prompt Evolution}, a prompt optimization mechanism where a generator LLM iteratively refines candidate prompts while a distinct executor LLM validates their effectiveness. This decoupling mitigates the self-correction bias inherent in single-model optimization. Evaluated on 130 CWE types, MulVul achieves 34.79\% Macro-F1, outperforming the best baseline by 41.5\%. Ablation studies validate cross-model prompt evolution, which boosts performance by 51.6\% over manual prompts by effectively handling diverse vulnerability patterns.

cs.SE↗

arXiv2025

ExpStar: Towards Automatic Commentary Generation for Multi-discipline Scientific Experiments

Experiment commentary is crucial in describing the experimental procedures, delving into underlying scientific principles, and incorporating content-related safety guidelines. In practice, human teachers rely heavily on subject-specific expertise and invest significant time preparing such commentary. To address this challenge, we introduce the task of automatic commentary generation across multi-discipline scientific experiments. While recent progress in large multimodal models (LMMs) has demonstrated promising capabilities in video understanding and reasoning, their ability to generate fine-grained and insightful experiment commentary remains largely underexplored. In this paper, we make the following contributions: (i) We construct \textit{ExpInstruct}, the first dataset tailored for experiment commentary generation, featuring over 7\textit{K} step-level commentaries across 21 scientific subjects from 3 core disciplines (\ie, science, healthcare and engineering). Each sample includes procedural descriptions along with potential scientific principles (\eg, chemical equations and physical laws) and safety guidelines. (ii) We propose ExpStar, an automatic experiment commentary generation model that leverages a retrieval-augmented mechanism to adaptively access, evaluate, and utilize external knowledge. (iii) Extensive experiments show that our ExpStar substantially outperforms 14 leading LMMs, which highlights the superiority of our dataset and model. We believe that ExpStar holds great potential for advancing AI-assisted scientific experiment instruction.

cs.CV↗

arXiv2025

Scalable Co-Clustering for Large-Scale Data through Dynamic Partitioning and Hierarchical Merging

Co-clustering simultaneously clusters rows and columns, revealing more fine-grained groups. However, existing co-clustering methods suffer from poor scalability and cannot handle large-scale data. This paper presents a novel and scalable co-clustering method designed to uncover intricate patterns in high-dimensional, large-scale datasets. Specifically, we first propose a large matrix partitioning algorithm that partitions a large matrix into smaller submatrices, enabling parallel co-clustering. This method employs a probabilistic model to optimize the configuration of submatrices, balancing the computational efficiency and depth of analysis. Additionally, we propose a hierarchical co-cluster merging algorithm that efficiently identifies and merges co-clusters from these submatrices, enhancing the robustness and reliability of the process. Extensive evaluations validate the effectiveness and efficiency of our method. Experimental results demonstrate a significant reduction in computation time, with an approximate 83% decrease for dense matrices and up to 30% for sparse matrices.

cs.DC↗

arXiv2025

Learner and Instructor Needs in AI-Supported Programming Learning Tools: Design Implications for Features and Adaptive Control

AI-supported tools can help learners overcome challenges in programming education by providing adaptive assistance. However, existing research often focuses on individual tools rather than deriving broader design recommendations. A key challenge in designing these systems is balancing learner control with system-driven guidance. To explore user preferences for AI-supported programming learning tools, we conducted a participatory design study with 15 undergraduate novice programmers and 10 instructors to gather insights on their desired help features and control preferences, as well as a follow-up survey with 172 introductory programming students. Our qualitative findings show that learners prefer help that is encouraging, incorporates visual aids, and includes peer-related insights, whereas instructors prioritize scaffolding that reflects learners' progress and reinforces best practices. Both groups favor shared control, though learners generally prefer more autonomy, while instructors lean toward greater system guidance to prevent cognitive overload. Additionally, our interviews revealed individual differences in control preferences. Based on our findings, we propose design guidelines for AI-supported programming tools, particularly regarding user-centered help features and adaptive control mechanisms. Our work contributes to the human-centered design of AI-supported learning environments by informing the development of systems that effectively balance autonomy and guidance, enhancing AI-supported educational tools for programming and beyond.

cs.HC↗

arXiv2025

Dynamic Influence Tracker: Measuring Time-Varying Sample Influence During Training

Existing methods for measuring training sample influence on models only provide static, overall measurements, overlooking how sample influence changes during training. We propose Dynamic Influence Tracker (DIT), which captures the time-varying sample influence across arbitrary time windows during training. DIT offers three key insights: 1) Samples show different time-varying influence patterns, with some samples important in the early training stage while others become important later. 2) Sample influences show a weak correlation between early and late stages, demonstrating that the model undergoes distinct learning phases with shifting priorities. 3) Analyzing influence during the convergence period provides more efficient and accurate detection of corrupted samples than full-training analysis. Supported by theoretical guarantees without assuming loss convexity or model convergence, DIT significantly outperforms existing methods, achieving up to 0.99 correlation with ground truth and above 98\% accuracy in detecting corrupted samples in complex architectures.

stat.ML↗

arXiv2025

LiveVal: Time-aware Data Valuation via Adaptive Reference Points

Time-aware data valuation enhances training efficiency and model robustness, as early detection of harmful samples could prevent months of wasted computation. However, existing methods rely on model retraining or convergence assumptions or fail to capture long-term training dynamics. We propose LiveVal, an efficient time-aware data valuation method with three key designs: 1) seamless integration with SGD training for efficient data contribution monitoring; 2) reference-based valuation with normalization for reliable benchmark establishment; and 3) adaptive reference point selection for real-time updating with optimized memory usage. We establish theoretical guarantees for LiveVal's stability and prove that its valuations are bounded and directionally aligned with optimization progress. Extensive experiments demonstrate that LiveVal provides efficient data valuation across different modalities and model scales, achieving 180 speedup over traditional methods while maintaining robust detection performance.

cs.LG↗

arXiv2025

Classic4Children: Adapting Chinese Literary Classics for Children with Large Language Model

Chinese literary classics hold significant cultural and educational value, offering deep insights into morality, history, and human nature. These works often include classical Chinese and complex narratives, making them difficult for children to read. To bridge this gap, we introduce a child-friendly literary adaptation (CLA) task to adapt the Chinese literary classic into engaging and accessible text for children. However, recent large language models (LLMs) overlook children's reading preferences (\ie, vivid character portrayals, concise narrative structures, and appropriate readability), which poses challenges in CLA. In this paper, we propose a method called InstructChild, which augments the LLM with these preferences for adaptation. Specifically, we first obtain the characters' personalities and narrative structure as additional information for fine-grained instruction tuning. Then, we devise a readability metric as the reward to align the LLM with the children's reading level. Finally, a lookahead decoding strategy is applied to improve the readability of the generated text during inference. To support the evaluation of CLA task, we construct the Classic4Children dataset, which comprises both the original and child-friendly versions of the Four Great Classical Novels of Chinese literature. Experimental results show that our InstructChild significantly improves automatic and human evaluation performance.

cs.CL↗

arXiv2025

Personalized Parsons Puzzles as Scaffolding Enhance Practice Engagement Over Just Showing LLM-Powered Solutions

As generative AI products could generate code and assist students with programming learning seamlessly, integrating AI into programming education contexts has driven much attention. However, one emerging concern is that students might get answers without learning from the LLM-generated content. In this work, we deployed the LLM-powered personalized Parsons puzzles as scaffolding to write-code practice in a Python learning classroom (PC condition) and conducted an 80-minute randomized between-subjects study. Both conditions received the same practice problems. The only difference was that when requesting help, the control condition showed students a complete solution (CC condition), simulating the most traditional LLM output. Results indicated that students who received personalized Parsons puzzles as scaffolding engaged in practicing significantly longer than those who received complete solutions when struggling.

cs.HC↗