Hallucination Is Generative Memory With Its Verifier Turned Down: One Constraint Axis Links Dreaming Sleep and LLM Confabulation
The paper argues that hallucination in large language models and dreaming in the brain are both generative reconstructions from distributed memory under incomplete constraints, organized into a four‑stage account of encoding, latent representation, generative reconstruction, and reality verification. It shows that the difference between reliable output and hallucination is governed by the strength of constraints and verification, and presents a small test on 300 PopQA questions where adding a verifier reduces hallucination rates in two language models. The study provides a framework and testa…
The paper argues that hallucination in large language models and dreaming in the brain are both generative reconstructions from distributed memory under incomplete constraints, organized into a four‑stage account of encoding, latent representation, generative reconstruction, and reality verification. It shows that the difference between reliable output and hallucination is governed by the strength of constraints and verification, and presents a small test on 300 PopQA questions where adding a verifier reduces hallucination rates in two language models. The study provides a framework and testa…
The framework offers Swiss AI researchers a new approach to align generative models with reality verification mechanisms, potentially informing Swiss AI policy and development of safer AI systems.
Original source record. Open the original record to verify the underlying announcement.