Good Ol' LSTM Thinking Steps
#2
by
TimeLordRaps
- opened
Do you remember how LSTMs can be fed back through themselves for thinking steps, how would I go about adapting this model for test-time aware thinking steps, where the model is meta-aware of its own necessary reasoning expectations, ie makes pre-estimates for quantity of thinking steps, would be useful to augment xLSTM LM with reasoning strength estimation, or just create a general adapter for reasoning estimation statistics generally. Happy researching, love the model.