Unelicitable Backdoors in Language Models via Cryptographic Transformer Circuits.
Andis DragunsAndrew GritsevskiySumeet Ramesh MotwaniCharlie Rogers-SmithJeffrey LadishChristian Schröder de WittPublished in: CoRR (2024)
Keyphrases
- language model
- language modeling
- n gram
- document retrieval
- probabilistic model
- language modelling
- speech recognition
- query expansion
- test collection
- information retrieval
- statistical language models
- query terms
- retrieval model
- context sensitive
- fault diagnosis
- relevance model
- vector space model
- ad hoc information retrieval
- word error rate
- language model for information retrieval
- okapi bm
- language models for information retrieval
- cross lingual
- smoothing methods
- term dependencies
- query specific
- document ranking
- translation model
- pseudo relevance feedback
- text mining
- bayesian networks