Future Lens: Anticipating Subsequent Tokens from a Single Hidden State
This addresses the interpretability of transformer models for researchers, though it is incremental as it builds on existing causal intervention methods.
The paper investigates whether a single hidden state in a transformer model can predict future tokens, finding that at some layers, it achieves over 48% accuracy in approximating subsequent token predictions.
We conjecture that hidden state vectors corresponding to individual input tokens encode information sufficient to accurately predict several tokens ahead. More concretely, in this paper we ask: Given a hidden (internal) representation of a single token at position $t$ in an input, can we reliably anticipate the tokens that will appear at positions $\geq t + 2$? To test this, we measure linear approximation and causal intervention methods in GPT-J-6B to evaluate the degree to which individual hidden states in the network contain signal rich enough to predict future hidden states and, ultimately, token outputs. We find that, at some layers, we can approximate a model's output with more than 48% accuracy with respect to its prediction of subsequent tokens through a single hidden state. Finally we present a "Future Lens" visualization that uses these methods to create a new view of transformer states.