Feedback-Adapters-for-Thinking-Steps-in-Frozen-Models

Feedback Adapters for Thinking Steps in Frozen Models

Problem:

We have generally knowledgeable models that are capable of some reasoning, but fail tragically in others and they use more or less a fixed compute budget at processing time, which can be seen as a constraint on their depth of reasoning.

Solution:

The Feedback Adapter: a layerwise thinking step that allows for at inference time variable compute budgets in the hopes that it leads to increased depth of reasoning in particularly language models.

General Algorithm:

For each layer:
    For i in range(layer's feedback times)
        First generate it's outputs from the inputs
        Then the Feedback adapter takes those outputs and creates a vector that can be fed back into this layer as the inputs.
    Generate the outputs of this layer from the final inputs generated by the feedback adapter.

Additional Context

In theory this doesn't even need to be applied to a frozen model and could be trained in tandem with a model at pretraining time.

I think emergent utility would come from training a feedback-adapter over a set of pre-trained models, probably with identical architectures but different weights, to create a meta-feedback adapter that is useable ubiquitously as a meta-reasoning function.