@DeFiMinty: Can AI systems keep generating harder training tasks for AI agents? New research from Tencent suggests they can. Recurs…
Summary
New research from Tencent introduces Recursive Synthetic Terminal Tasks (RST), a method that progressively generates harder, verifiable training tasks for AI agents. Starting from 639 tasks, it produced 37,484 verified tasks across 15 rounds, and reinforcement learning with these tasks improved Qwen3.5-27B from 22.7% to 32.0% on Terminal-Bench Hard.
View Cached Full Text
Cached at: 08/08/26, 07:06 AM
Can AI systems keep generating harder training tasks for AI agents?
New research from Tencent suggests they can. Recursive Synthetic Terminal Tasks (RST) starts with a terminal task that already works and adds new steps to make it harder.
It then updates the instructions and tests so they match the new workflow. Each revised task is validated in a fresh sandbox before it can be used to generate the next round of harder tasks.
Starting with 639 tasks, RST produced 37,484 verified tasks across 15 rounds. DeepSeek-V4-Pro could solve 90% of the first-round tasks within four attempts but only 2.5% of the final-round tasks.
The harder tasks also produced useful training data. Reinforcement learning with the generated tasks improved Qwen3.5-27B from 22.7% to 32.0% on Terminal-Bench Hard, an independently constructed benchmark.
Executable verification can turn a fixed set of tasks into progressively harder training data.
Yucheng Shi (@Yucheng__Shi): Building synthetic long-horizon terminal tasks is hard. Making them solvable and cheap is harder.
So we inverted the order: don’t write the task, build the solution.
We grow a reference solution and its runtime environment first, then fit a verifier and a task statement to it.
Similar Articles
Recursive Synthesis for Long-Horizon Terminal Tasks
This paper introduces Recursive Synthetic Terminal Tasks (RST), a recursive verified synthesis framework for generating long-horizon terminal-agent training data at scale, producing 37,484 tasks and improving Qwen3.5 models on terminal benchmarks.
Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading
Introduces Long-Horizon-Terminal-Bench, a benchmark of 46 long-horizon terminal tasks with dense reward-based grading, evaluating AI agents on planning, long-context, and debugging. Even the strongest model achieves only 15.2% pass@1, showing significant room for improvement.
@MaxForAI: Tian Yuandong @tydsh's startup team Recursive @Recursive_SI released a milestone: an automated AI research system. In this system, AI can complete the entire research loop of 'propose ideas → implement → run experiments → verify → select next experiment based on results'. Results show that with clear objectives...
The Recursive team released an automated AI research system that can autonomously complete the research loop, surpassing existing human community solutions on multiple benchmarks. For example, on NanoGPT Speedrun it compressed training time from 79.7 seconds to 77.5 seconds, and on SOL-ExecBench it improved the score to 0.754.
QUEST: Training Frontier Deep Research Agents with Fully Synthetic Tasks
QUEST is an open family of deep research agents trained with synthetic data and reinforcement learning, achieving strong performance across diverse long-horizon search tasks, approaching frontier closed-source agents.
@MangQiuyang: Open-ended coding training data may no longer be the bottleneck: AI can scale open-ended tasks—and even outperform huma…
FrontierSmith is a system that synthesizes open-ended coding problems at scale from closed-ended tasks. It generates, filters, and builds training environments; models trained on its data outperform those trained on human-curated open-ended data.