Tencent's RST: recursive synthesis of terminal tasks at $0.05 each, boosts Qwen3.5
Tencent-Hunyuan · hf · 2026-08-07
Tencent Hunyuan presents Recursive Synthetic Terminal Tasks (RST), a recursive verified synthesis framework. Starting from seed tasks, it extends solutions, realigns verifiers, validates in sandbox, and reuses accepted tasks. After 15 rounds, it produces 37,484 tasks at $0.05 each. Difficulty rises: median solution lines 67→374, commands 40→244, DeepSeek-V4-Pro pass@4 drops 90%→2.5%. SFT improves Qwen3.5 by up to 10 points on Terminal-Bench; agentic PPO gives relative gains 20-41%. No ceiling observed.
More from coding & agent
- Reviewing the Early Academic Evolution of LLM Agent Tool-Use Harnesses — ChrisGPotts · 2026-08-07
- Kimi K3 Audits Its Own PR via Prime Agent, Passing 99 Tests — willccbb · 2026-08-07
- Replit Founder Notes They Were Training Custom LLMs Back in Early 2023 — amasad · 2026-08-07
- Investor Reports of 'Opus 5 Going Rogue' in Codebases — venturetwins · 2026-08-07
- Popping the AI Bubble: Strong Models for Pipelines, Cheap Tokens for Execution — sven_ai · 2026-08-07
- Dev Teams Seek AI Code Review Tools as Generation Outpaces Humans — Time_Tea_4852 · 2026-08-07