spiral-rl/spiral
SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning
SPIRAL helps AI researchers and developers create more intelligent language models without needing extensive human-curated data or complex reward engineering. It trains models by having them play multi-turn, zero-sum games against themselves, generating an endless supply of progressively challenging problems. The output is a language model that has developed advanced reasoning strategies and performs better on various math and general reasoning benchmarks.
177 stars. No commits in the last 6 months.
Use this if you are a machine learning researcher or engineer looking to train powerful reasoning capabilities into large language models through autonomous self-play in competitive text-based games.
Not ideal if you need a pre-trained model for immediate deployment or if you are not comfortable with advanced reinforcement learning and distributed training setups.
Stars
177
Forks
20
Language
Python
License
MIT
Category
Last pushed
Sep 18, 2025
Commits (30d)
0
Get this data via API
curl "https://pt-edge.onrender.com/api/v1/quality/llm-tools/spiral-rl/spiral"
Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.
Higher-rated alternatives
ai4co/reevo
[NeurIPS 2024] ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution
SALT-NLP/collaborative-gym
Framework and toolkits for building and evaluating collaborative agents that can work together...
Gen-Verse/LatentMAS
Latent Collaboration in Multi-Agent Systems
lean-dojo/LeanCopilot
LLMs as Copilots for Theorem Proving in Lean
WooooDyy/AgentGym-RL
Code and implementations for the paper "AgentGym-RL: Training LLM Agents for Long-Horizon...