Conceptual Crossroads: Mapping Paradigm-Level Uncertainty in Language Models

TL;DR: This idea explores whether token-level uncertainty sometimes stems from competing conceptual frameworks rather than just alternate reasoning paths. We'll test this by creating prompts with embedded paradigm conflicts and measuring how hidden state dynamics differ between path-level and paradigm-level uncertainty.

Research Question: When language models exhibit uncertainty during reasoning, how can we distinguish between uncertainty arising from multiple reasoning paths versus uncertainty stemming from competing conceptual frameworks or paradigms?

Hypothesis: Paradigm-level uncertainty will manifest as distinct activation patterns compared to path-level uncertainty—specifically, paradigm conflicts will produce more persistent, globally distributed activation signatures across multiple layers, while path uncertainty will be more localized and transient.

Experiment Plan: Create a curated dataset of reasoning prompts that induce either (a) multiple valid reasoning paths to the same conclusion, or (b) genuine paradigm conflicts (e.g., quantum vs. classical physics explanations). Use the activation intervention techniques from Zur et al. to map uncertainty patterns across both types. Apply Sparse Autoencoders (Galichin et al.) to identify whether paradigm conflicts activate different reasoning features than path uncertainty. Compare the persistence and distribution of uncertainty signatures across model layers and time steps. Expected outcome: Paradigm conflicts will show higher activation of "uncertainty" and "reflection" features that persist across more tokens and layers than path uncertainty.

References: ['Zur, A., Geiger, A., Lubana, E., & Bigelow, E.J. (2025). Are language models aware of the road not taken? Token-level uncertainty and hidden state dynamics.', 'Galichin, A.V., et al. (2025). I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders. arXiv.org.']