ASK+: Improving Small Language Model Guidance in Partially Observable Environments

ASK+ is a new technique that supplies small language models with trajectory-aware context (partially revealed map, visited positions, action history) and structured chain-of-thought reasoning, turning them from passive redundancy checks into informative consultants for reinforcement learning agents.

On DoorKey, ASK+ achieves 93% success vs 89% for both PPO and vanilla ASK. On FourRooms, success climbs from 53% to 70%.

Yes, teams in the region can leverage this technique to improve decision-making systems in incomplete information environments, such as robotics and autonomous vehicles, without needing large models.