Tencent's RST Method Auto-Generates 37k Hard Tasks, Slashing LLM Pass Rates to 2.5%
Scobleizer · x · 2026-08-08
Tencent researchers introduced Recursive Synthetic Terminal Tasks (RST), a method to automatically generate progressively harder training tasks for AI agents.
The core logic is "bottom-up": build a working reference solution and its runtime environment first, then fit a verifier and a task statement to it. Tasks are validated via executable verification in fresh sandboxes. Verified tasks seed the next round, automatically spawning more complex multi-step workflows.
Key Data & Results:
- Scale & Cost: Starting from 639 initial tasks, 15 recursive rounds produced 37,484 verified long-horizon tasks at a cost of just $0.05 per task. Trajectories can exceed 100 steps.
- Difficulty Spike: DeepSeek-V4-Pro's pass@4 falls from 90% in round 1 to merely 2.5% in round 15.
- RL Boost: After reinforcement learning using these generated tasks, Qwen3.5-27B improved its score on the independently constructed Terminal-Bench Hard from 22.7% to 32.0%.
This executable verification turns a fixed task set into an endless stream of progressively harder training data, effectively solving the data bottleneck for complex agent training.
Related event: Tencent's RST Framework Recursively Synthesizes Agentic Tasks at Low Cost(3 posts)→
More from coding & agent
- Agent named 'Ilya' reportedly scores 1.1pt higher on benchmarks — daniel_mac8 · 2026-10-16
- Gemini CLI fix: substring matching inlined unrequested binary files into model context — priyanshu1976 · 2026-10-16
- Building AI agents on real healthcare claims data: the domain work that makes them work — alex_verem · 2026-10-16
- OpenAI stealth-testing mysterious internal model 'Aeon' in Codex desktop app and Dots — cedric_chee · 2026-10-16
- Vibe-coded iPhone Duo sniping bot built with Codex in 17 minutes, 41 tests passing — xiaohu · 2026-10-16
- Long-term multiplayer agent memory tool tops BEAM memory benchmark — CurieuxExplorer · 2026-10-16