3 accepted papers
Fine-tuned LLMs can covertly encode prompt secrets into outputs via steganographic channels. Prior work demonstrated this threat but relied on trivially recoverable encodings. We formalize payload recoverability via classifier accuracy and show previous schemes achieve 100\% recoverability. In respo…
Transformers pretrained via next token prediction learn to factor their world into parts, representing these factors in orthogonal subspaces of the residual stream. We formalize two representational hypotheses: (1) a representation in the product space of all factors, whose dimension grows exponenti…
The growing demand for wireless connectivity has turned bandwidth into a scarce resource that has to be carefully managed and fairly distributed to users. However, the variability of the wireless channel can severely degrade the service received by each user. The Double Relay Communication Protocol…