Skip to content

Why do we not pass the list of admissible actions into the prompt? #38

Description

@Deepakgthomas

Hello, Thank you for releasing the code. I have been trying to reproduce the ALFWorld experiments and have a design question.

At every step, ALFWorld's env exposes the full list of legal actions via info['admissible_commands']. The current prompt setup relies on the two few-shot trajectories to implicitly teach the action vocabulary (go to cabinet 1, take potato 1 from countertop 2, etc.), and the model is expected to produce syntactically valid actions from that demonstration alone.

I was curious why the admissible actions aren't explicitly included in the prompt at each step (e.g. appended as Valid actions: [...]). In my experiments with more recent models, I noticed that Large instruction-tuned LLMs (Llama 70B/90B) produce valid actions inconsistently across repeated runs.

I did go over the past issues of this repository where it was said that recent GPT models perform much worse in this type of task. I was thinking that if we passed the actions to the prompt, there might be an improvement. Moreover, I noticed that the Reflexion paper also followed the same approach which makes me wonder if this was a deliberate design choice.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions