LHTB 2.0 ships 77 tasks with 3-hour horizons and dense ladders after 1,145 runs

The benchmark extends the original 46 tasks by correcting defects and adding 31 reconstructed scenarios generated from parameterized reference implementations. Each extended task supplies a working artifact, six public behavioral traces, three held-out evaluation cases, and dense progress ladders that score partial completion every 30 minutes. Evaluation horizon lengthens from 90 minutes to three hours while preserving the 13 domains and eight task types.

Binary outcome scores on the 30 reconstruction tasks average 0.221 while dense ladder scores reach 0.653, demonstrating that final-state verification discards measurable intermediate progress. GPT-6-astra leads with 0.69 dense reward and 37 tasks at R ≥ 0.95; GPT-5.6-sol follows at 0.67 and 32 tasks. These gaps appear only when process rewards track sustained execution rather than terminal checks.

The pipeline forces models to generalize across generated variants instead of memorizing fixed outputs, directly addressing the memorization failure mode observed in prior terminal benchmarks. By logging verifier states at fixed intervals, LHTB 2.0 isolates whether agents maintain trajectory or stall, providing the first public dataset that separates short-term success from multi-hour reliability.

Future agent training runs can now optimize against the dense ladders rather than sparse binary signals, enabling measurable iteration on long-horizon robustness before deployment.