Position: Reframing Hallucination: Latent Space Geodesics as a Pathway for Generative Discovery

Zhihao Hao, Bob Zhang, LI Haisheng

TL;DR: We propose reinterpreting generative hallucinations as geodesic traversals in latent space essential for scientific discovery, offering a geometric framework to distinguish useful exploratory novelty from stochastic noise.

Abstract: Although evaluation practice for generative models has moved beyond purely retrieval-based metrics, many protocols still penalize deviations from known outputs, limiting their use in scientific discovery and creative reasoning. This position paper argues that uniformly suppressing such deviations can induce epistemic mode collapse, causing models to favor safe reproduction over exploratory hypothesis generation. We propose the Higher-Dimensional Cognitive Hypothesis (HDCH), which interprets some valuable hallucinations as high-dimensional latent-space traversals that appear erroneous when projected onto established knowledge. We distinguish Type I outputs, which are factually inconsistent or structurally incoherent, from Type II exploratory hypotheses, which are novel, structurally coherent, and worth further validation. Through controlled demonstrations, we illustrate that discovery-oriented generation benefits from calibrated instability, with exploratory yield peaking near a critical transition regime rather than increasing monotonically with randomness. We further advocate an Exploratory Signal-to-Noise Ratio (ESNR) framework that combines distributional divergence with external structural validation, shifting evaluation from static retrieval validation toward calibrated latent exploration.

Lay Summary: Generative AI systems are usually judged by how closely their answers match known facts or reference outputs. This is important for safety, but it can also discourage models from proposing new ideas in scientific discovery or creative reasoning. This position paper argues that some apparent hallucinations should be treated as exploratory hypotheses rather than immediately discarded. We propose a framework for separating harmful, incoherent errors from novel but structurally coherent outputs that may be worth further validation. The goal is not to make AI systems freely hallucinate, but to create safer evaluation methods that filter, sandbox, and verify promising exploratory outputs.

Primary Area: Research Priorities, Methodology, and Evaluation

Keywords: Generative AI, Hallucination, Latent Space, Scientific Discovery, Evaluation

Submission Number: 200