Search papers, labs, and topics across Lattice.
Tencent WorkBuddy Bench is a novel multi-domain benchmark designed to evaluate coding agents across four distinct work domains: Code, Web, Office, and Security. The benchmark employs a contamination-resistant task construction methodology, where tasks are derived from real-world scenarios and rewritten to prevent reverse-engineering, ensuring robustness against dataset leakage. The open release of the benchmark, including task directories and evaluation protocols, allows for reproducibility and transparency in assessing coding agent performance across various contexts.
Real-world coding tasks can be evaluated without risk of contamination, thanks to a unique reverse-engineering approach that ensures task prompts are untraceable.
We introduce Tencent WorkBuddy Bench, a multi-domain evaluation suite for coding agents; this report documents its construction methodology, scoring protocol, and a cross-model leaderboard. At its core is a unified evaluation framework for constructing and running distribution-informed coding-agent tasks across four work domains - Code, Web, Office, and Security. Rather than adapting public issue text, every task is reverse-engineered from a real commit, pull request, or business scenario and rewritten as a short, colloquial, role-played request, so that a task's prompt is not recoverable by web-searching the underlying issue, pull request, or commit thread. Because the dataset is released openly - task directories, environment images, evaluation harness, tests, and reference solutions - contamination resistance rests on this construction together with dataset versioning rather than on secrecy. The four subsets - repository-level engineering, front-end development, office and business workflows, and red-/blue-team security - probe complementary facets of real work, each with its own verification style. All are packaged in a uniform task-directory format and run, under a uniform and reproducible protocol, on two agent harnesses (CodeBuddy Code and Claude Code); the full open release makes the benchmark reproducible end to end and directly auditable, since any third party can re-run each task and inspect its content. Because each subset uses a different scoring instrument, scores are not comparable across subsets and the suite reports no suite-wide average. We report a cross-model leaderboard across several model families.
