Selected Publications (view all )

PlayCoder: Making LLM-Generated GUI Code Playable

Zhiyuan Peng, Wei Tao^#, Xin Yin, Chenhao Ying, Yuan Luo, Yiwen Guo (^# co-first author)

ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering 2026 FSE 2026 CCF-A

A GUI agent plays the generated game the way a user would, so playability is judged by interacting with the live program instead of by inspecting static code.

[Paper] [Code] [Cite]

RepoGenesis: Benchmarking End-to-End Microservice Generation from Readme to Repository

Zhiyuan Peng, Xin Yin^#, Pu Zhao, Fangkai Yang, Lu Wang, Ran Jia, Xu Chen, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang (^# co-first author)

Annual Meeting of the Association for Computational Linguistics 2026 ACL 2026 Main CCF-A

An agent reads a README and delivers a runnable multi-service repository, verified by building, launching, and calling the service inside an isolated sandbox.

[Paper] [Code] [Cite]

PrefGen: A Preference-Driven Methodology for Secure Yet Gas-Efficient Smart Contract Generation

Zhiyuan Peng, Xin Yin^#, Zijie Zhou, Chenhao Ying, Chao Ni, Yuan Luo (^# co-first author)

IEEE/ACM Automated Software Engineering Conference 2025 ASE 2025 CCF-A

A preference-driven generate-and-feedback loop that steers the model toward code satisfying two competing objectives at once: security and gas efficiency.

[Paper] [Code] [Cite]

SolEval: Benchmarking Large Language Models for Repository-level Solidity Code Generation

Zhiyuan Peng, Xin Yin^#, Rui Qian, Peiqin Lin, Yongkang Liu, Chenhao Ying, Yuan Luo (^# co-first author)

Conference on Empirical Methods in Natural Language Processing 2025 EMNLP 2025 Main TH-CPL-A

Repository-level generation with compilation and on-chain test execution in the loop, forcing the model to resolve cross-contract dependencies before its output can run.

[Paper] [Code] [Cite]

All publications

Open Source Projects

YimMenu Contributor

I built the Auto Drive agent: a perception-decision-control loop that navigates along roads to a map waypoint, roams autonomously when no goal is set, detects control conflicts, and yields on manual takeover, with HUD telemetry for its internal state. Validated across roughly 2,000 km of long-horizon in-game driving.

SoulBanner Author

A multi-persona .skill monorepo that distills stable expression styles and judgment frameworks into callable, composable agent skill modules, shipped with a router entry skill over 6 categories and a unified authoring template so the community can extend it via PRs. The catalog is consumed directly by OpenHire's Skill Catalog, so the skills I publish are the skills my agents run.

RepoGenesis Author

A benchmark that turns a README into a deployable multi-language web microservice repository, and then proves it by building, launching, and calling the service inside a Docker sandbox. Covers 11 frameworks and 18 application domains. Accepted by ACL 2026 Main (top 15% of accepted papers).

GOF2_remake Author

A playable WebGL space combat/trading game built from scratch by agents over 136 commits of long-horizon iteration: six frontier systems plus the PTD Home base, nine flyable hulls, a 13-chapter story chain, and an authoritative economy backend over REST + SSE. Art and audio assets are generated, registered in manifest.json, and scheduled on demand, while Playwright desktop and mobile e2e suites turn "does it boot and stay playable" into an automatic regression on every iteration.

LoveEmulator Author

An AI-driven life simulation game that splits procedural content generation into four controllable layers: the engine samples a situation skeleton, an LLM writes the narrative and player choices, a validation layer checks JSON structure, safety boundaries, and logical consistency, and the engine resolves outcomes deterministically from intent tags, relationship state, and an RNG seed. Generated content keeps LLM variety while staying reproducible and numerically bounded.

OpenHire Author

A digital-employee orchestration platform. One control agent decomposes a fuzzy goal and fans it out to multiple role-based worker agents, each running in its own Docker container with an isolated workspace and permission scope. A pluggable worker adapter admits any image-packaged agent (openclaw, hermes, and nanobot are supported today), and the platform is reachable via CLI, an OpenAI-compatible API, or IM channels such as Feishu, with an Admin console for sessions, transcripts, and runtime state. A governance loop closes the system: successful runs are distilled into reusable skills, case packages, and consolidated long-term memory, so the next run starts stronger.

Agent Stack

• Orchestration: goal decomposition by a control agent, role-based routing, parallel multi-agent collaboration and cross-checking
• Execution: containerized worker lifecycle, per-agent isolated workspaces, permission scoping, sandboxed build-run-verify loops
• Tools & Interfaces: tool use / function calling, OpenAI-compatible API, CLI and IM (Feishu) channels, GUI interaction
• Memory & Skills: skill extraction and reuse, reusable case packages, long-term memory consolidation with diffs and safe rollback
• Observability: session transcripts, trace review, workspace and runtime inspection