Inverted Task Synthesis: 37K Long-Horizon Tasks at $0.05 Each, DeepSeek Pass Rate Drops to 2.5%

teortaxesTex · x · 2026-08-07

Building solvable and cheap synthetic long-horizon terminal tasks is challenging. Researchers propose an inverted approach: instead of writing the task first, they build and run the reference solution and its runtime environment, then fit a verifier and task statement to it. Tasks are solvable by construction with built-in reward signals, and verified tasks seed the next generation.

Over 15 recursive rounds, this method generated 37,484 verified long-horizon tasks at a cost of just $0.05 per task. In evaluations, DeepSeek-V4-Pro's pass@4 plummeted from 90% in Round 1 to 2.5% in Round 15. Trajectories reach 100+ steps, becoming brutally harder without breaking the generation factory.

Training with PPO on this dataset improved Qwen3.5-27B's performance on TB2 from 41.2 to 49.4. The paper, data, trajectories, and models are fully open-sourced.

Related event: Tencent's RST Framework Synthesizes Long-Horizon Tasks at Low Cost(2 posts)→

Original post →

More from coding & agent

coding & agent channel →