Long-Horizon-Terminal-Bench — testing agents on complex multi-step tasks
New benchmark with 46 long-horizon terminal tasks designed to measure agent progress beyond final outcomes. Covers experiment reproduction, software engineering, multimodal analysis, games, and scientific computing.
Each task decomposes into fine-grained milestones with dense reward signals, replacing the sparse binary pass/fail of existing benchmarks. Reveals gaps in how well agents handle intermediate steps and partial solutions.