Continuous optimization recovers input text from GPT-2 last-layer hidden states at up to 97.5% exact match on short prompts, with failures concentrated on high-frequency function words.
In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Recovering Input Text from Hidden States: Study of Gradient-Based Inversion of Decoder-Only Language Models
Continuous optimization recovers input text from GPT-2 last-layer hidden states at up to 97.5% exact match on short prompts, with failures concentrated on high-frequency function words.