From Drift to Coherence: Stabilizing Beliefs in LLMs
Researchers have identified and partially resolved a fundamental coherence problem in how large language models form beliefs during inference. Prior work showed LLMs violate the martingale property, a key assumption underlying implicit Bayesian reasoning, but only in synthetic settings. This study extends the analysis to realistic multiple-choice QA tasks and introduces prompted predictive resampling, a technique where models generate repeated answers to the same question. The method exposes early belief drift but reveals that models self-correct through resampling, eventually stabilizing their predictions. This finding matters for practitioners relying on LLM confidence calibration and for theorists building formal models of LLM reasoning under uncertainty.
Modelwire context
ExplainerThe practical implication buried in this paper is that confidence scores from a single LLM forward pass may be systematically unreliable not because of calibration failures in the traditional sense, but because the model's internal belief state hasn't yet converged. Resampling is less a fix than a diagnostic that reveals how much inference-time variance practitioners have been quietly ignoring.
The coherence problem described here sits in productive tension with the SoftMoE paper covered the same day, which addresses how expert routing affects compute allocation during inference. Both papers, from different angles, are probing what actually happens inside a model during a single forward pass and whether the outputs can be trusted at face value. The dimensionality and modularity work ('Dimensionality Controls When Modularity Helps') adds a third angle: architectural choices that look principled may not behave as expected under realistic conditions. Together these papers suggest a quiet but consistent research pressure on the assumption that LLM outputs are stable, well-formed probability estimates rather than snapshots of an unsettled process.
Watch whether prompted predictive resampling gets adopted as a standard confidence-elicitation baseline in calibration benchmarks over the next two conference cycles. If it does, that would confirm the martingale framing has moved from theoretical curiosity to practical evaluation infrastructure.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsLarge Language Models · Prompted Predictive Resampling
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.