We are excited to announce CLIN 🤖: The first continually learning language agent that excels in both task adaptation and generalization to unseen tasks and environments in a pure zero-shot setup. Aristo Team at Ai2 Ai2 Website: Let's dive in 🧵 (1/n)

(2/n) CLIN uses a persistent, dynamic memory for this novel nonparametric learning paradigm across 18 tasks in ScienceWorld benchmark, beating SOTA reflective agent Reflexion by 23 points.

(3/n) CLIN creates (Trial1) or adapts (Trial2+) a memory of causal abstractions to help in future trials by reflecting on the last trial and current memory. It does this using a suitably prompted LLM to generate the updated memory.

(4/n) We find insights centered around causal abstraction are particularly helpful for generalization than "generic insights," similar findings w @AndrewLampinen

(5/n) In CLIN, a controller takes the current task, retrievals from memory, and the trial so far to generate the next goal. The executor converts this to a valid action. The simulator returns an observation. Memory is updated at the end of each trial by the memory generator.

(6/n) CLIN shows rapid adaptation: (a) CLIN readily adapts to both short and long tasks. CLIN becomes more efficient in later trials by solving the tasks with a lower number of (average) steps. (b) CLIN outperforms state-of-the-art Reflexion by 23 absolute points.

(7/n) Positive Transfer: CLIN beats previous Reinforcement Learning agents and Language agents with positive transfer learning. Both in generalization over unseen environments (a) and tasks (b), CLIN takes fewer steps than the base model and achieves better performance.

(8/n) We expand on the direction of "gradient-free learning," showing goal-directed exploration and continual updates to insights based on past experience significantly boost performance without any startup cost or requirement. @DrJimFan

(9/n) While Voyager shows the importance of procedural memory (skills), we show a "semantic" memory has better portability to out-of-distribution task/environment samples.

(10/n) Language agents have the unique advantages of expressing, memorizing, and executing in natural language leading to incredible flexibility for adaptation and generalization. We are expanding this exciting line of work @ai2_aristo

(n/n) This was only possible with the huge team effort: @bhavana_dalvi, @peterjansen_ai, Oyvind, Niket, @liharryzhang, Chris Callison-Burch, and Peter Clark! CLIN is open source: Paper:

@ai2_aristo @allen_ai Congrats!!