AutoEnvScaling Turns Agents Into Environment Builders

// 1h agoRESEARCH PAPER

AutoEnvScaling Turns Agents Into Environment Builders

AutoEnvScaling treats RL environment design as a terminal task: a proposer agent reads solver failures, builds and validates new Harbor environments, and uses fresh solver rollouts to guide the next iteration. The paper reports up to 25× higher valid-environment rates, 75% lower cost per valid environment, and measurable Terminal-Bench gains.

// ANALYSIS

AutoEnvScaling makes the data flywheel executable instead of manually curated, bringing recursive self-improvement closer to a practical training workflow.

–Terminal access lets proposers write, run, debug, and validate complete environments rather than generate brittle task descriptions.
–Solver feedback is central: environments built from actual failures outperform static pools and prompt-only generation.
–GPT-5.6-sol-generated environments improved Qwen3.5-35B-A3B from 34.6 to 38.9 on Terminal-Bench 2.1.
–Qwen3.6-35B-A3B improved from 41.1 to 53.3 on Terminal-Bench 2.1 when acting as both proposer and solver.
–Terminal-Bench 4.0 gains remained modest, showing that benchmark generalization and robust RSI are still open problems.

// TAGS

autoenvscalingagenttrainingsynthetic-datatraining-infraevaluationcoding-agent

DISCOVERED

1h ago

2026-10-15

PUBLISHED

1h ago

2026-10-15

RELEVANCE

9/ 10

AUTHOR

arnaudmercier

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

BENCHMARK51m ago

Gemini 4 Argon Matches Astra on ARC-AGI-3

ARC Prize says Google’s Gemini 4 Argon shows a performance profile on ARC-AGI-3 broadly similar to OpenAI’s GPT-6 Astra, putting Google back among frontier contenders. The interactive benchmark tests exploration, memory, goal discovery, and adaptive planning rather than static question answering.

RESEARCH1h ago

AI-Native Air Interface Exposes Physical Backdoor

A Cryptology ePrint paper shows that poisoning just 0.5% of training data can backdoor AI-native 5G-Advanced functions, with a physical trigger requiring only a 2-bit reconfigurable intelligent surface. The attack achieved near-total control over CSI feedback and beam prediction in experiments.

LAUNCH1h ago

ServiceNow unveils AI Passport for agents

ServiceNow introduced AI Passport, a cryptographically signed authority layer that verifies what first- and third-party AI agents and MCP servers can access or execute at runtime. Expected in Q4 2026, it complements expanded discovery, observability, cost governance, and policy enforcement through ServiceNow’s Super Intelligence Control Tower.