TerminalBench-Science selection: 70 tasks chosen from 920 proposals

BenBlaiszik · x · 2026-08-28

TerminalBench-Science set a high bar for task inclusion, selecting only 70 tasks for v0.1 from 920 proposals. Each task underwent domain review, technical review, and a final difficulty calibration. The author estimates over 80 hours of effort went into each task to ensure they truly challenge top-tier models.

Related event: Stanford-led Team Launches Terminal-Bench-Science, a Research-Agent Benchmark; Best Model Solves Only 30%(11 posts)→

Original post →

More from Research

• DSPy.Flex Optimizes Python Modules with Fewer LLM Calls — lateinteraction · 2026-08-28
• MIT Framework PottsMPNN Designs Proteins Beyond Natural Structures — nordicinst · 2026-08-28
• Podcast: Discussing Python Tooling Handbook and the Value of Blogs in the Age of Agents — tdhopper · 2026-08-28
• MIT's PottsMPNN Framework Advances Protein Design Beyond Natural Sequences — MIT News AI · 2026-08-28
• Debate on the Term "Double Blind" in Model Evaluation Context — BlancheMinerva · 2026-08-28
• Debugging exact DDP checkpoint resumption: CUDA atomics, Triton autotune and bucket layout pitfalls — ReinforcedKnowledge · 2026-08-28

Research channel →