Long-Horizon Terminal-Bench
Tencent HY / Lehigh and collaborators / Coding · Sustained terminal use for complex engineering and tool tasks.
What it measures and how
90-minute, two-hour, and three-hour budgets are separate. Judging protocols before and after fixes cannot be mixed. Different agents' best results cannot represent common model conditions.
How this evidence is used
Candidate under review: awaiting a batch explicitly using the corrected protocol. Existing results are not automatically collected or scored publicly. This does not imply every old result exploited a vulnerability.
Evaluation limits & attribution
The evaluator disclosed judging-information leakage. Existing results cannot currently be verified under one isolated judging protocol. Task-source overlap with other specialist evaluations also needs review.
Data license: Apache 2.0 · Official IntelligenceLab/LHTB-leaderboard result dataset. Retain attribution and change notices.
Results are published by Tencent HY / Lehigh and collaborators. Original scores and the NexGateHub consensus index use different scales and cannot be added directly.