BayesBench: Evaluating LLM Belief Trajectories Under Multi-Turn Evidence Accumulation

arXiv cs.AI Papers

Summary

BayesBench evaluates how closely large language models' belief updates match Bayesian reasoning in multi-turn evidence accumulation tasks, finding that while scaling improves latent inference, models struggle to use that understanding for downstream predictions.

arXiv:2606.30850v1 Announce Type: new Abstract: Large language models (LLMs) are typically deployed in multi-turn conversations, where each turn provides new evidence that should reduce epistemic uncertainty about their environment. Acting rationally then requires inferring the unobserved quantities that govern it and updating beliefs about them as evidence accumulates. Yet most evaluations only score the model's final-turn answer in a single-turn format, leaving this process unexamined. We ask how closely LLMs' belief updates match those of a rational Bayesian reasoner in multi-turn settings, and introduce BayesBench, a suite of simulation environments that probe this across three progressively complex tasks: (i) Bayesian estimation, where the model infers an unknown parameter from sequential evidence; (ii) Bayesian prediction, where the model turns inferred beliefs about a latent variable into outcome forecasts; and (iii) latent-framed Bayesian prediction, where observations are filtered through a user-persona framing, requiring joint inference over the latent state and the persona. Across seven LLMs (3B--70B), scaling improves latent inference and evidence accumulation, with updates occasionally matching the Bayesian posterior. However, these gains do not reliably carry over to downstream prediction, exposing a gap between inferring latent structure and using it to rationally update beliefs about the target outcome.
Original Article
View Cached Full Text

Cached at: 07/01/26, 05:36 AM

# BayesBench: Evaluating LLM Belief Trajectories Under Multi-Turn Evidence Accumulation
Source: [https://arxiv.org/abs/2606.30850](https://arxiv.org/abs/2606.30850)
[View PDF](https://arxiv.org/pdf/2606.30850)

> Abstract:Large language models \(LLMs\) are typically deployed in multi\-turn conversations, where each turn provides new evidence that should reduce epistemic uncertainty about their environment\. Acting rationally then requires inferring the unobserved quantities that govern it and updating beliefs about them as evidence accumulates\. Yet most evaluations only score the model's final\-turn answer in a single\-turn format, leaving this process unexamined\. We ask how closely LLMs' belief updates match those of a rational Bayesian reasoner in multi\-turn settings, and introduce BayesBench, a suite of simulation environments that probe this across three progressively complex tasks: \(i\) Bayesian estimation, where the model infers an unknown parameter from sequential evidence; \(ii\) Bayesian prediction, where the model turns inferred beliefs about a latent variable into outcome forecasts; and \(iii\) latent\-framed Bayesian prediction, where observations are filtered through a user\-persona framing, requiring joint inference over the latent state and the persona\. Across seven LLMs \(3B\-\-70B\), scaling improves latent inference and evidence accumulation, with updates occasionally matching the Bayesian posterior\. However, these gains do not reliably carry over to downstream prediction, exposing a gap between inferring latent structure and using it to rationally update beliefs about the target outcome\.

## Submission history

From: Ankur Samanta \[[view email](https://arxiv.org/show-email/c97c227a/2606.30850)\] **\[v1\]**Mon, 29 Jun 2026 19:30:32 UTC \(852 KB\)

Similar Articles

A Model of Multi-turn Human Persuadability Using Probabilistic Belief Tracing

arXiv cs.CL

This paper introduces PersuasionTrace, a framework for studying multi-turn persuasion in human-LLM interaction, using a Bayesian-network simulated target that models belief updates. The framework reveals that LLMs are persuasive across topics and modalities, and that the Bayesian target better matches human belief dynamics than vanilla LLM simulators.

DLawBench: Evaluating LLMs Through Multi-Turn Legal Consultation

arXiv cs.CL

DLawBench is a new benchmark for evaluating large language models in multi-turn legal consultation, covering Chinese and US law with four client types. Experiments show significant room for improvement, with the best model achieving only 0.562 on legal reasoning.

Probabilistic Attribution For Large Language Models

arXiv cs.CL

This paper proposes a model-agnostic probabilistic token attribution measure for LLMs using Bayes' rule to invert next-token log probabilities, capturing the model's internal representation of token sequences and improving interpretability through entropy analysis.