← 研踪 PaperTrailarXiv:2611.21710 · v1
Learning In-Context Adaptivity for LLMs through Continual Gameplay
首次提交 2026-11-17 02:58
版本更新 2026-11-17 02:58
雷达观测 2026-11-18 19:44
摘要
LLM agents are increasingly deployed in long-horizon tasks where they must explore, gather experience, and adapt their decisions accordingly. We name this capability in-context adaptivity: improving decisions across episodes from experience accumulated in context. We propose to evaluate it in repeated imperfect-information games against unknown opponents. In these games, fixed game-theoretically optimal policies cannot achieve maximum reward, whereas adapting to the opponent increases returns across games. We design a paired protocol that evaluates each agent with and without memory under identical conditions to isolate experience gain from single-game skill. Evaluation shows that memory brings little performance gain for open-weight models and inconsistent gains for frontier models. Our analysis of the agents' summaries and actions identifies a memory-doing gap in which models accurately record opponent tendencies yet rarely depart from default play. To close this gap, we construct a pool of exploitable opponents and post-train Qwen3.5-9B with supervised fine-tuning followed by trajectory-level reinforcement learning. We optimize full-match return across context segments so summary generation and action selection are trained together. On unseen opponents, our 9B RL policy doubles experience gain over supervised fine-tuning in Leduc Hold'em and increases it fivefold in No-Limit Hold'em. Across 35 opponents, our policy matches or surpasses frontier models, generalizes beyond its training horizon, and retains most of its gain on summaries from frontier models.
作者备注:36 pages
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-17 02:58 · 最新观测版本
Learning In-Context Adaptivity for LLMs through Continual Gameplay
LLM agents are increasingly deployed in long-horizon tasks where they must explore, gather experience, and adapt their decisions accordingly. We name this capability in-context adaptivity: improving decisions across episodes from experience accumulated in context. We propose to evaluate it in repeated imperfect-information games against unknown opponents. In these games, fixed game-theoretically optimal policies cannot achieve maximum reward, whereas adapting to the opponent increases returns across games. We design a paired protocol that evaluates each agent with and without memory under identical conditions to isolate experience gain from single-game skill. Evaluation shows that memory brings little performance gain for open-weight models and inconsistent gains for frontier models. Our analysis of the agents' summaries and actions identifies a memory-doing gap in which models accurately record opponent tendencies yet rarely depart from default play. To close this gap, we construct a pool of exploitable opponents and post-train Qwen3.5-9B with supervised fine-tuning followed by trajectory-level reinforcement learning. We optimize full-match return across context segments so summary generation and action selection are trained together. On unseen opponents, our 9B RL policy doubles experience gain over supervised fine-tuning in Leduc Hold'em and increases it fivefold in No-Limit Hold'em. Across 35 opponents, our policy matches or surpasses frontier models, generalizes beyond its training horizon, and retains most of its gain on summaries from frontier models.
查看此版本解读 →查看此版本原文 ↗