How can subtask grading reveal agent progress on long tasks?

Long-Horizon-Terminal-Bench takes the standard Terminal-Bench setup — a reference solution or simulation engine per task — but decomposes each of its 46 tasks into fine-grained deterministic subtasks, each graded against the environment. This produces dense intermediate rewards and partial credit, which matters because the tasks are genuinely long: averaging 239 episodes, 9.8M tokens, and 88.9 minutes per run. On workflows that long, a binary final-outcome verdict discards almost all the information. The strongest model (Grok 4.5) passes only 28.3% at R≥0.95 while the mean across 17 models is just 6.4% — a near-floor result that outcome-only scoring would render as undifferentiated failure.

The design argument is that reward density is not a nicety but a measurement necessity once horizons stretch past one-shot problem solving. When agents make "meaningful but incomplete progress," a pass/fail benchmark cannot distinguish a model that got 80% of the way from one that stalled immediately — both read as zero. This grounds the broader claim that Should agent evaluation measure more than task success?: dense subtask grading is one concrete instrument for that shift. It also echoes the training-side finding that Can we reward reasoning steps without human annotation? — dense signals expose structure that outcome-only aggregation flattens, whether the goal is evaluation or optimization.