Environment Evolution for Terminal Agents
Summary
This paper proposes environment evolution to incrementally increase task difficulty for terminal agents off-policy, enhancing continuous learning signals and improving performance on benchmarks like Terminal-Bench 2.1.
View Cached Full Text
Cached at: 09/04/26, 03:57 AM
Paper page - Environment Evolution for Terminal Agents
Source: https://huggingface.co/papers/2609.04128 Authors:
,
,
,
,
,
,
,
,
,
,
Abstract
Environment evolution incrementally raises task difficulty off-policy to sustain continuous learning signals for terminal agents, improving benchmark performance through multi-agent harnesses.
Scaling interactive and verifiable environments is critical for training terminal agents. As frontier models become more capable, environments synthesized from scratch become less challenging and thus provide limited learning signals. Recentco-evolutionmethods iteratively synthesize environments near the model’s learnable frontier based on weaknesses exposed during rollouts. However, their dependence onon-policy rolloutslimits generalization and the continuous provision of learning signals as the model becomes stronger. In this paper, we propose environment evolution, which incrementally increases environment difficultyoff-policyand schedules the evolved environments generation by generation during training to provide continuous learning signals. We derive three evolution directions that influence environment difficulty from themulti-turn learning objectiveand then implement evolution along these directions through aloop-engineered multi-agent harness. Quantitative rollout experiments with Hy4 preview, Claude Opus 5, and GPT-5.6 Sol show that environment evolution consistently produces more difficult environments. We validate its effectiveness on Qwen3.6-27B and Qwen3.6-35B-A3B through simplelong-horizon RL training, improving their performance by 14.4 and 18.0 percentage points on Terminal-Bench 2.1, respectively.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2609\.04128
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2609.04128 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2609.04128 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2609.04128 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Terminal-World: Scaling Terminal-Agent Environments via Agent Skills
Terminal-World introduces a fully automated pipeline that uses agent skills to synthesize high-quality training data for terminal agents, enabling models to outperform baselines with only 1.2% of the training data. The method co-derives task instructions, environments, and teacher trajectories from skill primitives.
ECHO: Terminal Agents Learn World Models for Free
ECHO introduces a hybrid objective that combines policy-gradient loss with environment observation prediction to provide dense supervision from terminal feedback, doubling performance on TerminalBench-2.0 for Qwen3 models.
LiteCoder-Terminal: Scaling Long-Horizon Terminal Environments for Learning Language Agents
LiteCoder-Terminal-Gen introduces a zero-dependency synthetic pipeline that generates executable terminal training environments, producing SFT and RL datasets that enable language agents to achieve significant performance gains on Terminal Bench benchmarks.
Testing Agents on Long-Horizon Terminal Work (GitHub Repo)
Long-Horizon Terminal-Bench (LHTB) is a 46-task benchmark for evaluating LLM agents on sustained terminal work over hundreds of steps, revealing that even the best models solve only ~28% of tasks.
Do Agent Optimizers Compound? A Continual-Learning Evaluation on Terminal-Bench 2.0
This paper introduces a two-phase continual-learning evaluation on Terminal-Bench 2.0 to test whether gains from agent-optimization methods compound when applied recursively. It finds that only RELAI-VCL, which incorporates regression control, achieves compounded improvement.