litlas
ログインアカウント作成

Future Lens: Anticipating Subsequent Tokens from a Single Hidden State

著者: Koyena Pal, Jiuding Sun, Andrew Yuan, Byron C. Wallace, David Bau - Proceedings of the 27th Conference on Computational Natural Language Learning (CoNLL) 2023 被引用: 56

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.

Serial Order: A ParallelDistributed Processing…Serial Order: A Parallel Distributed Processing ApproachQualitative SpatialReasoning over Question…Qualitative Spatial Reasoning over Questions (Short Paper)The Pile: An 800GBDataset of Diverse Text…The Pile: An 800GB Dataset of Diverse Text for Language ModelingPrefix-Tuning:Optimizing Continuous…Prefix-Tuning: Optimizing Continuous Prompts for GenerationLocating and EditingFactual Associations in…Locating and Editing Factual Associations in GPTConfident AdaptiveLanguage ModelingConfident Adaptive Language ModelingEmergent Abilities ofLarge Language ModelsEmergent Abilities of Large Language ModelsEliciting LatentPredictions from…Eliciting Latent Predictions from Transformers with the Tuned LensFinding Neurons in aHaystack: Case Studies…Finding Neurons in a Haystack: Case Studies with Sparse ProbingMass-Editing Memory in aTransformerMass-Editing Memory in a TransformerJump to Conclusions:Short-Cutting…Jump to Conclusions: Short-Cutting Transformers With Linear TransformationsLinearity of RelationDecoding in Transformer…Linearity of Relation Decoding in Transformer Language ModelsDo language models planahead for future tokens?Do language models plan ahead for future tokens?Unlocking the Future:Exploring Look-Ahead…Unlocking the Future: Exploring Look-Ahead Planning Mechanistic Interpretability in Large Language ModelsExtracting Paragraphsfrom LLM Token…Extracting Paragraphs from LLM Token ActivationsA Practical Review ofMechanistic…A Practical Review of Mechanistic Interpretability for Transformer-Based Language ModelsSpeculative Streaming:Fast LLM Inference…Speculative Streaming: Fast LLM Inference without Auxiliary ModelsLet's Think Dot by Dot:Hidden Computation in…Let's Think Dot by Dot: Hidden Computation in Transformer Language ModelsNeuron-Level KnowledgeAttribution in Large…Neuron-Level Knowledge Attribution in Large Language ModelsHow Do LLMs Use TheirDepth?How Do LLMs Use Their Depth?Emergent ResponsePlanning in LLMsEmergent Response Planning in LLMsThe Dual-Route Model ofInductionThe Dual-Route Model of InductionFrom Tokens to Words: Onthe Inner Lexicon of…From Tokens to Words: On the Inner Lexicon of LLMsUnboxing the Black Box:Mechanistic…Unboxing the Black Box: Mechanistic Interpretability for Algorithmic Understanding of Neural NetworksFuture Lens:Anticipating Subsequent…Future Lens: Anticipating Subsequent Tokens from a Single Hidden State過去の参考文献中心の論文この論文を引用する論文古い新しい

ノードをクリックするとフォーカスを固定、空白をクリックすると本論文に戻ります。ホバーで一時的にプレビューできます。各ノードのページはタイトルから開けます。