• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
AI
Announcement

AutoEnvScaling uses a coding agent to build training environments for terminal tasks

A post describing the project reports Qwen3.6-35B-A3B rose from 41.1 to 53.3 on Terminal-Bench 2.1.

Natasha JaquesNJ
Weiyan ShiWS
Simon Yu ✈️ EMNLP2026SY
4 Sources, 4h ago, first seen 4h ago

TLDR

A post describes AutoEnvScaling as a follow-up to SPADE that automates post-training environments and curriculum for terminal tasks. Its coding agent reads a reasoning agent’s rollouts, then builds, runs and fixes environments in a sandbox. The author says each checkpoint builds tasks to train the next and reports Qwen3.6-35B-A3B going from 41.1 to 53.3 on Terminal-Bench 2.1.

Combined views

6.4K

4 Sources, first seen 4h ago

101 likes4 comments68 saves46 reposts

Combined views

6.4K

4 Sources, first seen 4h ago

101 likes4 comments68 saves46 reposts
Image: @simon_ycl on X

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Featured Source

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

4 Sources

Simon Yu ✈️ EMNLP2026@simon_yclRecursive self-improvement needs agents that can build their own training environments. But researchers still part of the loop, making it hard to scale. We introduce AutoEnvScaling, which turns Data Flywheel into a terminal task: a proposer agent builds complete Harbor tasks and sandbox. It builds just like human researchers. This enables self-play RL for terminal agents: one model improves as both builder and solver, without human-in-the-loop.4h
Bo Liu (Benjamin Liu)@Benjamin_eecsa minimal RSI would be: a coding agent writes the repo that trains its successor, the pretrain, mid-train and post-train recipe plus a new harness. As a follow-up to SPADE, AutoEnvScaling automates one piece of that repo: post-training environments and curriculum for the terminal domain. The environment designer is now a coding agent: it reads the reasoning agent's rollouts, then builds, runs and fixes each environment in a sandbox. Building a terminal task and solving one are both coding, so one model plays both roles and gets better at both. Each checkpoint builds the tasks that train the next. Qwen3.6-35B-A3B goes from 41.1 to 53.3 on Terminal-Bench 2.1 :))4h
Weiyan Shi@shi_weiyanRT @simon_ycl: Recursive self-improvement needs agents that can build their own training environments. But researchers still part of the lo…1h
Natasha Jaques@natashajaquesOur new paper pushes multi-agent recursive self-improvement even further. This time the proposer agent builds whole Harbor sandbox environments to challenge the solver agent. It learns from the solver’s successes and failures to update its own memories, skills, and tools, and build increasingly challenging---yet valid---tasks. We show powerful benefits both when using a proprietary model as the proposer, and in self-play, when both the solver and proposer are the same open-weight model. In this RSI setting, Qwen3.6-35B-A3B doubles its score on Terminal Bench 4, and performs so well on HiL-Bench it actually exceeds Claude Opus-5. More evidence of the power of multi-agent co-adaptation to unlock continuous improvement and emergent capabilities.27m
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI

    4 Sources

    Simon Yu ✈️ EMNLP2026@simon_yclRecursive self-improvement needs agents that can build their own training environments. But researchers still part of the loop, making it hard to scale. We introduce AutoEnvScaling, which turns Data Flywheel into a terminal task: a proposer agent builds complete Harbor tasks and sandbox. It builds just like human researchers. This enables self-play RL for terminal agents: one model improves as both builder and solver, without human-in-the-loop.4h
    Bo Liu (Benjamin Liu)@Benjamin_eecsa minimal RSI would be: a coding agent writes the repo that trains its successor, the pretrain, mid-train and post-train recipe plus a new harness. As a follow-up to SPADE, AutoEnvScaling automates one piece of that repo: post-training environments and curriculum for the terminal domain. The environment designer is now a coding agent: it reads the reasoning agent's rollouts, then builds, runs and fixes each environment in a sandbox. Building a terminal task and solving one are both coding, so one model plays both roles and gets better at both. Each checkpoint builds the tasks that train the next. Qwen3.6-35B-A3B goes from 41.1 to 53.3 on Terminal-Bench 2.1 :))4h
    Weiyan Shi@shi_weiyanRT @simon_ycl: Recursive self-improvement needs agents that can build their own training environments. But researchers still part of the lo…1h
    Natasha Jaques@natashajaquesOur new paper pushes multi-agent recursive self-improvement even further. This time the proposer agent builds whole Harbor sandbox environments to challenge the solver agent. It learns from the solver’s successes and failures to update its own memories, skills, and tools, and build increasingly challenging---yet valid---tasks. We show powerful benefits both when using a proprietary model as the proposer, and in self-play, when both the solver and proposer are the same open-weight model. In this RSI setting, Qwen3.6-35B-A3B doubles its score on Terminal Bench 4, and performs so well on HiL-Bench it actually exceeds Claude Opus-5. More evidence of the power of multi-agent co-adaptation to unlock continuous improvement and emergent capabilities.27m
    Today's Rank

    #10

    Today's Rank

    #10