Long-Horizon Terminal-Bench
What it measures and how
Compare 90-minute, 2-hour and 3-hour budgets separately. Judging protocols before and after fixes cannot be ranked together. The best results from different Agents are not common model scores.
How this evidence is used
Candidate under observation: awaiting new batches explicitly identifying the fixed protocol. Existing old scores are not automatically collected into public boards or ranked. This does not claim every old result exploited a flaw.
Limitations and attribution
The organization disclosed judging-information leakage. Current evidence does not establish that existing scores use the same isolated judging protocol. Task-source overlap with other professional benchmarks needs review.
Data license: Apache 2.0 · Official IntelligenceLab/LHTB-leaderboard results dataset; retain attribution and change notes.
Results are published by Tencent HY / Lehigh et al.. Raw scores and the e/acc consensus index use different scales and cannot be added directly.