Recursive Synthesis Generates 37k Agent Tasks at $0.05 Each
burny_tech · x · 2026-08-14
The paper Recursive Synthesis for Long-Horizon Terminal Tasks introduces RST, a recursive verified synthesis framework designed to overcome the bottleneck of expensive training data for terminal agents.
- Core Mechanism: Starting from a small set of verified seed tasks, RST extends reference solutions, realigns verifiers and instructions, validates results in fresh sandboxes, and reuses accepted tasks as new seeds.
- Scale & Cost: Across 15 recursive rounds, it generated 37,484 tasks from 639 seeds at roughly $0.05 per task. Task difficulty scales substantially (median reference solution grows from 67 to 374 lines).
- Training Utility: Fine-tuning Qwen3.5 on these tasks yields up to 10-point improvements on terminal-agent benchmarks. Agentic PPO further boosts performance, delivering relative gains of 20% to 41% over the base model.
- Potential: After 15 rounds, the recursion shows no ceiling, maintaining stable synthesis yield and validation rates.
More from coding & agent
- Over 800 Fake AI Skills and MCP Servers Found Delivering Malware — HaktanSuren · 2026-08-14
- Stop Waiting: Local LLM Users Trapped in the 'Next Model' Excuse Loop — ForsookComparison · 2026-08-14
- From Reactive to Proactive: A Design Guide for Ambient Agents — blaizedsouza · 2026-08-14
- Open-source Kubernetes Dashboard Kite Integrates AI Agents and Multi-Cluster Management — tom_doerr · 2026-08-14
- AI developer: Without rapidly building quality evals, you'll never catch up; RSI is nothing without hillclimbing — marktenenholtz · 2026-08-14
- Povio Worklog MCP: Automating Worklog Generation via Natural Language — modelcontextprotocol · 2026-08-14