SPADE: Self-Play in Adaptive Synthetic Executable Environments

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Bo Liu, Simon Yu, Yiding Jiang, Ao Qu, Andrew Zhao, Zichen Liu, Junsu Kim, Zijian Zhou, Seungone Kim, Tongzheng Ren, Mickel Liu, Hanfei Yu, Zhaorun Chen, Weiyan Shi, Paul Pu Liang, Luke Zettlemoyer, Yejin Choi, Natasha Jaques
• Venue: arXiv: cs.CL
• Published: 2026-08-19
• Collected: 2026-08-20
• URL: http://arxiv.org/abs/2608.19197v1
• PDF: https://arxiv.org/pdf/2608.19197v1
• DOI: UNKNOWN
• Registry group: academic_apis_and_preprints
• Evidence role: canonical_academic_index
• Language: en
• Country / region: UNKNOWN

Plain-Language Brief

The paper introduces SPADE (Self-Play in Adaptive Synthetic Executable Environments), a framework that allows a language model to design and learn in its own self-generated, diverse, adaptive environments. This approach improves the performance of the language model on various math, science, code, and reasoning benchmarks.

What the Document Says

• SPADE is a self-play reinforcement learning framework that allows a single large language model (LLM) to play two roles: an Environment Designer and a Reasoning Agent.
• The Environment Designer writes complete, long-horizon training environments as executable code using an OpenAI Gym-style reset()/step() interface.
• The Reasoning Agent learns to act in these self-designed environments.
• Each environment is a stateful, multi-turn environment consisting of state transitions, reward functions, and verification code.
• The Reasoning Agent's regret is estimated using the gap between its reward with and without privileged hints.
• Through extensive experimentation, several components were found to be critical for success: grounding the Environment Designer on documents sampled from a large pretraining corpus and giving it an accumulated environment memory.
• SPADE improves over the strongest fixed-environment baseline by +5.3 on average across eight held-out math, science, code, and reasoning benchmarks.
• SPADE lifts the tool-use setting by +5.7 on BFCL-v4 multi-turn and +13.9 on ACEBench-Agent, and on the games setting, the margin over the strongest baseline grows with model scale.

Key Concepts Mentioned

• SPADE
• Self-Play in Adaptive Synthetic Executable Environments
• reinforcement learning
• large language model
• Environment Designer
• Reasoning Agent
• regret estimation
• privileged hints
• stateful multi-turn environment

Claims Made by the Document

• SPADE improves over the strongest fixed-environment baseline by +5.3 on average across eight held-out math, science, code, and reasoning benchmarks.
• SPADE lifts the tool-use setting by +5.7 on BFCL-v4 multi-turn and +13.9 on ACEBench-Agent.
• On the games setting, the margin over the strongest baseline grows with model scale.

Evidence Used

• Improves over the strongest fixed-environment baseline by +5.3...
• lifts the tool-use setting by +5.7...
• On the games setting, the margin over the strongest baseline grows with model scale.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Continuous self-improvement requires an ever-expanding pool of self-generated, diverse, adaptive goals. For language agents, existing training environment pools (hand-curated, statically synthesized, or frozen-verifier) keep the goal distribution fixed as the learner scales. We introduce SPADE (Self-Play in Adaptive Synthetic Executable Environments), a self-play RL framework in which a single LLM plays two roles: an Environment Designer that writes complete, long-horizon training environments as executable code with an OpenAI Gym-style reset()/step() interface, and a Reasoning Agent that learns to act in them. Each is a stateful, multi-turn environment (state transitions, reward functions, and verification code), so one interface spans reasoning problems and multi-step agentic tool use. The Reasoning Agent's regret is estimated using the gap between its reward with and without privileged hints; in optimizing this regret signal the Environment Designer learns to target environments at the edge of the agent's capabilities while keeping them feasible. Through extensive experimentation, we find several components critical to success: grounding the Environment Designer on documents sampled from a large pretraining corpus, and giving it an accumulated environment memory. Scaling to 30B-parameter models, SPADE improves over the strongest fixed-environment baseline by +5.3 on average across eight held-out math, science, code, and reasoning benchmarks, and lifts the tool-use setting by +5.7 on BFCL-v4 multi-turn and +13.9 on ACEBench-Agent; on the games setting, the margin over the strongest baseline grows with model scale. By making environment design itself a learnable component, SPADE takes a concrete step toward open-ended self-improvement.