New AI Learns to Read Opponents Like a Poker Pro
This could make AI assistants smarter at adapting to you over time.
Researchers evaluated in-context adaptivity in repeated imperfect-information games against unknown opponents. Memory brought little performance gain for open-weight models and inconsistent gains for frontier models, revealing a memory-doing gap: models accurately record opponent tendencies yet rarely depart from default play.
They post-trained Qwen3.5-9B with supervised fine-tuning followed by trajectory-level reinforcement learning. On unseen opponents, the 9B RL policy doubled experience gain over supervised fine-tuning in Leduc Hold'em and increased it fivefold in No-Limit Hold'em. Across 35 opponents, the policy matched or surpassed frontier models, generalized beyond its training horizon, and retained most of its gain on summaries from frontier models.
- Most AI models don't improve from experience even when they remember past interactions, a problem called the 'memory-doing gap'.
- A smaller AI trained to use memory adaptively doubled or quintupled its performance against new opponents in poker-like games.
- This could lead to AI assistants that personalize to you over time, but real-world applications are still untested.
Why It Matters
AI that learns from experience could make assistants more personalized, efficient, and useful in everyday tasks.