self-play

The tweet introduces a research concept called Self-Play Pretraining with Zero Data, where two models start from random initialization and generate pretraining data through self-play on a universal Turing machine.

The article introduces Self-Play Pretraining with Zero Data, a research method where two LLMs start from random initialization and learn through a generator proposing programs for a universal Turing machine, demonstrating how inductive biases can enable general learning.

Skild AI demonstrated a robot trained to play football using self-play simulation equivalent to 140 years of practice in weeks, illustrating a new approach to skill acquisition through computer simulation.

Skild AI trained a Unitree G1 robot to play soccer by simulating 140 years of practice against its own past versions, demonstrating significant progress in AI and robotics.

This paper introduces CoVer, a co-training framework for code generation that addresses self-play RL failures by using information-gain rewards and diversity-pruned tests, achieving significant pass rate improvements on benchmarks.

DiagEvo improves language-model self-evolution by deriving training direction from internal failure history via hierarchical error-cause memory and double-confidence filtering, outperforming baselines that rely on external resources.

This paper introduces ShuttleArena, a physics-based self-play environment for badminton where agents learn interpretable tactical policies using PPO, showing competitive improvement in shot selection and recovery.

LURE introduces a zero-data self-play framework for LLM reasoning that uses a pursuit-evasion game to dynamically adjust task difficulty and provide dense rewards, outperforming baselines across multiple environments.

SPADE introduces a self-play reinforcement learning framework for language models that generates adaptive executable training environments to enhance reasoning and tool-use capabilities, demonstrating significant performance gains across multiple benchmarks.

Introduces RLSVR, a task-transformation paradigm that extends RLVR to open-ended tasks via self-verifiable rewards from self-play games, instantiated in SpyRL and Vision-Zero. It improves LLM performance on summarization, creative writing, and math reasoning.

This paper introduces SESA, a self-evolving skill-augmented search agent that co-evolves task generation and skill memory via tool-augmented search self-play. It improves accuracy across seven QA benchmarks over baselines while supporting memory-free deployment.

The Qwen team proposes the Skill Self-Play framework, which significantly improves model capabilities on tool-calling and reasoning tasks through the collaboration of Proposer, Solver, and a dynamic skill controller in self-play.

Users report that Kimi K3 (an AI model) can generate an entire Three Kingdoms deckbuilding roguelike game in a single shot, and then autonomously tune its own balance over ten thousand self-played games.

This paper introduces GPT-Red, an automated red-teaming agent trained via self-play at scale to discover novel prompt injection attacks against frontier LLMs, and uses it to adversarially train GPT-5.6, achieving the largest documented LLM safety training run.

This paper proposes RLSVR, a task-transformation paradigm that extends reinforcement learning with verifiable rewards to open-ended LLM tasks by creating self-verifiable proxy environments, instantiated via the SpyRL multi-agent self-play framework, showing gains on summarization, creative writing, and math reasoning.

OpenAI developed GPT-Red, an LLM trained via self-play to automatically red-team other models, finding vulnerabilities like a novel 'fake chain of thought' prompt injection. The approach improved the robustness of GPT-5.6.

OpenAI introduces GPT-Red, an automated red-teaming model trained via self-play reinforcement learning to improve model robustness against prompt injection attacks. Applied to GPT-5.6, it achieves 6x fewer failures on direct prompt injection benchmarks.

Introduces Hallucination Self-Play (HSP), a framework that bootstraps a detector using an evolved generator via reinforcement learning, enabling small LLMs to match advanced LLMs on faithfulness hallucination detection without external supervision.

This paper investigates design choices for training lightweight reinforcement learning agents in imperfect-information card games, using a fixed rule-based expert as a yardstick for Gin Rummy and Leduc Hold'em. It finds that trust region updates, careful reward design, curriculum learning, warm starting, and checkpoint retention improve performance, while several common techniques like reward shaping and large language model opponents do not help.

Proposes Anchored Self-Play (ASP), a method for scaling code repair supervision via generator–fixer self-play with an embedding-similarity reward and reference bug mixing, achieving +24% relative improvement in fix rates over standard self-play on a new benchmark BugSourceBench.