analiza aiqbench
This research investigates belief drift in large language models during autoregressive generation, demonstrating how predictive distributions can self-stabilize through answer resampling techniques. The authors propose specific prompting strategies and fine-tuning losses to accelerate convergence toward coherent Bayesian-like inference without compromising accuracy on standard benchmarks. These findings advance the theoretical understanding of LLM reliability by addressing fundamental violations of martingale properties in practical usage regimes.
From Drift to Coherence: Stabilizing Beliefs in LLMs
arXiv:2606.17832v1 Announce Type: new Abstract: Large language models (LLMs) are often hypothesized to perform implicit Bayesian inference, yet a key coherence condition, the martingale property of predictive beliefs, has been shown to fail in controlled synthetic in-context learning settings. We revisit this question in a more typical usage regime: generic multiple-choice question answering. Exploiting the discrete answer space, we compute exact predictive distributions and study belief dynamics induced by autoregressive answer resampling. We introduce prompted predictive resampling (PPR), where an LLM generates a sequence of answers to the same question. Empirically, PPR reveals early-stage belief drift, indicating martingale violations. However, after sufficient resampling steps, the belief process self-stabilizes and converges to a coherent predictive distribution. Based on this observation, we further propose (i) a seed-answer prompting strategy to accelerate stabilization, and (ii) a self-consistency loss that amortizes early-stage drift into the model via fine-tuning. Experiments on multiple-choice QA benchmarks show that our methods substantially reduce belief drift and improve predictive coherence without sacrificing accuracy.
Abstract:Large language models (LLMs) are often hypothesized to perform implicit Bayesian inference, yet a key coherence condition, the martingale property of predictive beliefs, has been shown to fail in controlled synthetic in-context learning settings. We revisit this question in a more typical usage regime: generic multiple-choice question answering. Exploiting the discrete answer space, we compute exact predictive distributions and study belief dynamics induced by autoregressive answer resampling. We introduce prompted predictive resampling (PPR), where an LLM generates a sequence of answers to the same question. Empirically, PPR reveals early-stage belief drift, indicating martingale violations. However, after sufficient resampling steps, the belief process self-stabilizes and converges to a coherent predictive distribution. Based on this observation, we further propose (i) a seed-answer prompting strategy to accelerate stabilization, and (ii) a self-consistency loss that amortizes early-stage drift into the model via fine-tuning. Experiments on multiple-choice QA benchmarks show that our methods substantially reduce belief drift and improve predictive coherence without sacrificing accuracy.
Machine Learning (cs.LG)
arXiv:2606.17832 [cs.LG]
(or arXiv:2606.17832v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2606.17832 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 253 cuvinte
Subiecte conexe
Correlated Coherent Errors in Stabilizer Codes: A General Cumulant Framework and Interference-Based Error Suppression
arXiv · 27 iul. 2026
Testing Catability and Coherent Superposition of $2\mathcal{D}$ Graphene Quantum system
arXiv · 11 iun. 2026
A Bayesian Mirror Architecture for Emergent Consciousness: Circular Hierarchies, Self-Manifolds, and Hybrid Event-Self Binding
arXiv · 8 oct. 2026
Cumulant Framework Analyzes Quantum Noise Beyond Standard Models
Quantum Zeitgeist · 1 aug. 2026