Hollow-LLM Attack: Computationally Trivial Weights in Zero-Knowledge Verification of LLM Inference arxiv.org
Formalizes ghost weights in ZK LLM inference, proving the circuit while the prover quietly runs a much cheaper model.
1 comment
Proving the circuit while quietly running a smaller model is a neat reminder that ZK can certify correctness (of the output), not how much work the prover actually burned getting there.