Bayesian Intelligence from the Outside
cs.AI · econ.TH
Quick verdict
Shows repeated LLM reports are Bayes-rationalizable iff one state survives all reports; sharp, honest theory, but leaves almost no empirical rejection power.
Summary
The paper develops a revealed-preference-style theory of when an agent's reports, e.g. from an LLM queried repeatedly, can be explained as Bayesian updating on an unobserved, prompt-dependent internal experiment. The agent holds a full-support prior, receives a prompt-specific signal, updates by Bayes' rule, and obediently reports the posterior over answers; repetitions are i.i.d. draws at one fixed state. Theorem 1 shows such a rationalization exists if and only if some state remains possible under every report across all prompts: only jointly consistent hard exclusions (zero-probability claims) are testable, and report frequencies or positive probability magnitudes add no restrictions — in particular, no martingale properties arise. Theorem 2 characterizes a plausible Blackwell intelligence order via a coupling of report distributions with support inclusion; Example 1 and Appendix B.1 construct opposite orderings for the same histograms. Theorem 3 shows the Bayes-plausible aggregate of coarse reports must exclude exactly the answers excluded by some report, while full-state reports admit prior-corrected log-odds pooling without knowing the experiment. Limitations: results rest on obedience and correct-specification assumptions, are essentially 'anything goes' beyond zero-consistency, and the paper offers no empirical measurement of how often the binding condition arises for real models.
Review complete
Assessment at a glance
3
risk flags identified
● high 0
● medium 2
● low 1
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A collection of an agent's report distributions is rationalizable as Bayesian updating on unknown prompt-dependent experiments if and only if some state remains possible under every report across all prompts (Theorem 1).
- Report frequencies and the magnitudes of positive probabilities impose no testable restrictions, and repeated sampling at a fixed state yields no martingale properties (Theorem 1, Remark 3).
- Plausible Blackwell dominance holds exactly when a support-inclusion coupling of the report distributions exists, so identical data can admit opposite intelligence orders (Theorem 2, Example 1, Appendix B.1).
- The Bayes-plausible aggregate of coarse reports must exclude exactly the answers excluded by at least one report, whereas full-state reports admit prior-corrected log-odds pooling without knowing the experiment (Theorem 3, Lemma 3).
Strengths
Sharp iff-characterizations with complete constructive proofs
Theorem 1 (eq. 3), Theorem 2 (support-inclusion coupling) and Theorem 3 (supp condition eq. 7) are proven in Appendices A.3, A.4, A.8 via explicit lift-and-implement constructions, including a reusable known-prior lemma (Lemma 1, odds-moment inequality H(τ,ω;z) ≤ μ0(z)/μ0(ω)).
Honest scoping of what the criteria cannot do
Remark 5 concedes a mechanical degenerate reporter plausibly Blackwell-dominates interior reports, and Remark 1 states the two correct-specification requirements upfront; Remark 4 shows the intelligence order is prior-invariant (Appendix A.5).
Precise treatment of the zero-versus-epsilon distinction
Actionable aggregation guidance for ML practice
Weaknesses
Thin practical bite for real LLM evaluation
Load-bearing untestable assumptions
All 'anything goes' results depend on unreported state slack
Known answer partitions per prompt assumed
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a theory paper (cs.AI / econ.TH) asking which collections of reports from an agent such as an LLM can be rationalized as Bayesian updating on an unobserved, prompt-dependent internal experiment. The model (Section 2): a finite state space Ω, a full-support prior μ0, a per-prompt Blackwell experiment Ep = (Sp, πp), Bayes updating, and obedient reporting (Assumption 1) of the posterior marginal over an answer partition Yp; repetitions are i.i.d. conditional on one fixed true state, and unlimited resampling is idealized as exact observation of the report distribution τp via the SLLN. Three results anchor the paper: Theorem 1 (Bayes-plausibility holds iff the intersection in eq. (3) is nonempty); Theorem 2 (plausible Blackwell dominance holds iff a common possible answer exists and a coupling of τ2, τ1 with supp ν2 ⊆ supp ν1 exists); Theorem 3 (a μ0-plausible aggregate after a history with at least two distinct reports exists iff supp η ⊆ ∩ supp νi, under |Ωy| ≥ 2). The framing is deliberately a limiting result: Bayesian rationality is very hard to reject from black-box reports.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 3 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does it mean for LLM outputs to be Bayes-plausible?
A set of the model's reports across prompts and repeated samples is Bayes-plausible if there exists some unobserved internal experiment per prompt, a full-support prior, and one true state such that Bayes updating followed by faithful reporting reproduces exactly those report distributions. The paper proves this is possible if and only if some state remains possible under every report, i.e., the reports never jointly rule out everything.
Can repeated sampling of an LLM detect whether it reasons in a Bayesian way?
Under this paper's assumptions, almost never. Repetition frequencies and the sizes of positive probabilities are unconstrained, and no martingale laws apply because each repetition redraws a signal at the same fixed state. The only falsifiable pattern is inconsistent hard exclusions: one report assigning exactly zero probability to an answer in a way that leaves no common possible state across all reports.
How should I aggregate multiple LLM samples — is majority voting Bayesian?
The paper shows a Bayes-plausible aggregate must exclude any answer that some report excluded; plain arithmetic averaging violates this when one sample gives an answer zero probability and another rates it positively. Log-odds averaging respects the restriction and is exactly optimal in the paper's Gaussian model. Self-consistency majority voting is therefore not justified as Bayesian aggregation in this framework.
Does this paper empirically test whether LLMs are Bayesian?
No, it is a theory paper. It derives what could in principle be tested about Bayesian rationality from black-box reports when the internal reasoning experiment is unknown, and shows those tests are extremely weak — highlighting that existing empirical Bayesian-consistency evaluations rely on additional assumptions the authors make explicit.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure 1's plotted histograms are accessible only via caption and the Example 1 text; visual features were not verifiable from the extraction.
- Proofs in Appendices A–B were assessed for structure and internal consistency, not re-derived line by line.
- Displayed equations (e.g., the intersection in eq. 3) were reconstructed from linearized text extraction and may contain minor symbol-level artifacts.
- 2026-dated citations were taken at face value; no external verification was performed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.