Novelty Blindness in Large Language Models: The Missing System 0 and the Structural Inability to Prospect for Discovery
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Large language models (LLMs) do not miss knowledge—they miss importance gradients. We identify a structural deficiency we term novelty blindness: the inability to autonomously assess the epistemic value of information and to prospect for discovery-class insights. Human cognition employs at least three processing layers: System 1 (fast pattern matching), System 2 (deliberate reasoning), and what we here call System 0—a missing class of inference-time operation that would operationalize prediction error and compression gain into cognitive control flow. We formalize this as an absent control function C(t) = f(E_p(i), ΔK(i)) that never enters LLM inference despite the architecture’s capacity to approximate its components. A three-phase empirical pilot across four major LLM systems (Gemini, GPT-4o, Claude, Perplexity) demonstrates: (a) identical baseline scores of 1/8 across all models, providing evidence consistent with architecture-general novelty blindness; (b) strong recovery to 7–10/12 under direct elicitation, confirming the capability-policy distinction; and (c) further improvement to 8–14/14 under a structured domain jitter protocol grounded in stochastic resonance dynamics. We identify a universal containment snap-back in which models actively steer back to the retrieval context even when explicitly invited to prospect—a behavior likely reinforced by RLHF alignment alongside base-model instruction-following priors. We formalize a two-axis taxonomy (salience failure × propagation failure), propose a benchmark methodology anchored to embedding-space metrics, and advance competing scaling hypotheses with explicit falsification criteria.
Fields of Study & Subjects
Topic Modeling
Natural Language Processing Techniques
Artificial Intelligence Applications
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 28 March 2026
- Language
- EN