Tag: #self-play5 results
Researchers propose Anchored Bipolicy Self-Play to address limitations in standard self-play red teaming for AI safety. It uses distinct role-specific LoRA adapters on a frozen base model, avoiding self-consistency collapse and maintaining adversarial pressure. Evaluations on Qwen2.5 models show superior safety and efficiency over baselines.
Two same-model LLM agents compete on coding problems; better execution winner forms DPO pairs for fine-tuning, repeating cycles. Pure execution reward (pass rate), local hardware friendly with specialist temps and memory consolidation. Early Colab A100 results: HumanEval Pass@1 from 0.671 to 0.683 (+1.2pp).
MEMO is a self-play framework optimizing inference-time context for multi-turn multi-agent LLM games, addressing variance and underperformance. It couples memory retention from trajectories with prompt evolution using TrueSkill and prioritized replay. Achieves 25% to 49.5% win rate gains for GPT-4o-mini across five text games.
A developer successfully created a superhuman-level agent for the game Generals.io using self-play reinforcement learning. The project leverages JAX for high-performance simulation and Vision Transformers to replace traditional CNNs for better scaling.
A new efficient MCTS implementation in Python/Numba called gumbel-mcts has been released on GitHub. Its PUCT version is 2-15x faster than baselines while matching policies exactly. Includes dense and sparse Gumbel MCTS, ideal for large action spaces like chess and low simulation budgets.