Transformers are Bayesian Networks

Authors: Greg Coppola Year: 2026 Tags: transformers, belief-propagation, bayesian-networks, formal-verification, mechanistic-interpretability, hallucination

TL;DR

Every sigmoid transformer with any weights implements one round of weighted loopy belief propagation (BP) per layer on an implicit factor graph defined by those weights; with explicitly constructed weights it implements exact BP, and on tree-structured grounded knowledge bases this yields provably correct Bayesian posteriors. All five main claims are formally verified in Lean 4, and each is corroborated by a supporting experiment.

First pass — the five C's

Category. Theory paper with constructive proofs (formally machine-checked in Lean 4) plus small-scale experimental confirmation.

Context. Probabilistic graphical models / transformer mechanistic interpretability. Builds directly on: Pearl's sum-product algorithm [12] for exact BP on trees; Coppola's own QBBN bipartite factor-graph language [2, 3] for grounded probabilistic logic; Elhage et al. [4] routing/residual-stream view of attention; Joshi [8] framing of full self-attention as a GNN on a complete graph; Turing-Good log-odds weight-of-evidence algebra (credited to wartime Bletchley Park work).

Correctness. Load-bearing assumptions: (1) sigmoid (not ReLU/GELU) activation is required for the exact internal equivalence — the ReLU empirical experiment is treated as compatible but not identical; (2) tree structure of the factor graph is required for the no-hallucination corollary; (3) independence of incoming messages is exact on trees but violated on loopy graphs; (4) the implicit factor graph G(W) is constructed to match the transformer's computation by definition, making Theorem 6.1 partly tautological — the authors acknowledge this in §6.3 but the rebuttal is not fully convincing. Assumptions (1)–(3) are stated clearly; (4) is understated.

Contributions. - Formal proof (Lean 4) that any sigmoid transformer with arbitrary weights implements weighted loopy BP on a well-defined implicit factor graph G(W). - Constructive explicit weight matrices for exact BP on any pairwise factor graph, with a binarization argument showing two attention heads per layer always suffices. - Uniqueness theorem: a sigmoid transformer that produces exact Bayesian posteriors for all inputs is forced to have BP weights (w₀ = w₁ = 1, b = 0 in FFN; projectDim/crossProject in attention). - Finite concept space theorem linking verifiability to a finite grounded symbol set, framing hallucination as a structural consequence of absent grounding rather than a fixable scaling failure.

Clarity. Mathematically organized and largely readable; the author acknowledges that prose was generated in collaboration with Claude (Anthropic), which introduces occasional puffery (e.g., the Leibniz section); the loopy-BP experimental section is cut off mid-sentence in the manuscript, suggesting an incomplete draft.

Second pass — content

Main thrust: The sigmoid FFN computes exactly the Turing-Good log-odds update (σ(logit(m₀) + logit(m₁))), and attention computes exactly the BP gather step, so every sigmoid transformer forward pass is one round of weighted loopy BP; with explicit BP weights and a tree-structured knowledge base, this yields exact Bayesian marginals and provably zero hallucination.

Supporting evidence: - Bayes-learner experiment: 2-layer, 2-head transformer encoder (~5,000 params), trained on 18,000 randomly generated two-variable factor graphs (factor entries ∈ [0.05, 1.0] uniform), validated on 2,000 held-out graphs; val MAE = 0.000752; posteriors matched to three decimal places; reported as 99.3% improvement over baseline (baseline not defined). - Table 1 shows max per-graph errors of 0.0003–0.0023 on four held-out examples. - Turing machine simulation: 5 structurally distinct machines (binary incrementer, decrementer, bitwise complement, left shift, right shift) all reached 100% validation accuracy in 4 epochs with identical hyperparameters, 2-layer 2-head encoder, dmodel = 32, ~4K params. - Scaling theorem: any k-ary factor graph binarizes exactly via AND associativity and log-odds additivity, requiring only d·⌈log₂ k⌉ layers (depth scales with reasoning complexity; head count fixed at 2). - Finite concept space theorem: a factor graph with n nodes induces exactly 2n² distinct routing keys; formally verified in Lean 4.

Figures & tables: Table 1 (transformer vs. BP posteriors, 4 examples) — no error bars, no confidence intervals, no distributional summary beyond val MAE. Table 2 (8-dim token encoding) and Table 3 (three softmaxes) are expository. No figures are described in the provided text. Statistical significance is never reported. The loopy-BP experimental results table/figure is missing (text cuts off).

Follow-up references: - Pearl [12] — classical exactness of BP on trees, the cornerstone of the no-hallucination corollary. - Coppola [2, 3] — the QBBN grounding language and AND/OR bipartite factor-graph structure this paper formally implements. - Elhage et al. [4] — routing/residual-stream view of attention that the BP construction makes formally precise. - Yedidia, Freeman & Weiss [21] (cited as Bethe free energy reference) — loopy BP fixed points, relevant for the practical convergence claims.

Third pass — critique

Implicit assumptions: - The G(W) construction (§6.2) is defined to match the transformer's computation, making Theorem 6.1 largely definitional — the real content is that sigmoid FFN = Ψ_or exactly and attention = BP gather exactly, which is the actual mathematical substance but is not cleanly separated from the definitional scaffolding. - The uniqueness theorem (§6.5) requires that the transformer produces exact posteriors "for all inputs" — a condition no trained model satisfies; its applicability to real trained transformers is asserted but not demonstrated. - The no-hallucination result requires full grounding of the knowledge base in advance, a preprocessing step that is not automatic and whose difficulty is not discussed. - The independence assumption embedded in updateBelief is exact only when messages arrive from disjoint subtrees; the paper relies on this for the tree result and notes only briefly that it is violated on loopy graphs. - All analysis targets encoder-style transformers (Vaswani et al.); causal masking, autoregressive generation, and decoder-only architectures (GPT-family) are not addressed despite being the dominant production form.

Missing context or citations: - No engagement with prior BP-attention correspondence literature beyond brief mentions of Joshi [8] and one unnamed reference in §11.3; claims of novelty relative to this body of work are not substantiated. - The "99.3% improvement over baseline" is stated without defining the baseline. - No comparison to neural theorem provers, probabilistic logic networks, or neurosymbolic systems that also aim at grounded probabilistic inference. - The Bethe free energy convergence conditions for loopy BP ([10, 15] cited but not given full citations in available text) are invoked to justify practical convergence without quantifying when the claimed "sparse, long loops" condition holds in realistic knowledge bases.

Possible experimental / analytical issues: - The main experiment uses only the minimal two-variable factor graph (v₀—f₁—v₂), where exact BP converges in one round matching a single forward pass; no multi-hop reasoning chains are tested, leaving the depth-equals-hops claim unconfirmed experimentally. - No ablation comparing sigmoid vs. ReLU transformers on identical tasks to quantify the gap between "exact" and "compatible" internal representations claimed in §10.4. - Turing machine experiments test only single-step prediction; multi-step sequential reasoning is not evaluated. - Table 1 shows four cherry-picked examples; distributional statistics (error histogram, worst-case error, etc.) are absent. - The loopy-BP empirical section (§10.5) is textually incomplete in the manuscript — results for five graph structures cannot be evaluated. - The prose was generated via iterative collaboration with Claude (acknowledged); informal arguments in philosophical sections (§9.5) may carry more confidence than the mathematics warrants. - No reproducibility artifacts (code, model checkpoints) are described in the available text beyond repository names.

Ideas for future work: - Apply the BP weight construction to decoder-only, causally masked transformers and determine whether an analogous equivalence holds or what breaks. - Test the depth-equals-reasoning-hops prediction on multi-hop QA benchmarks (e.g., HotpotQA, ProntoQA) by varying transformer depth against inference chain length. - Train sigmoid and ReLU transformers side by side on identical BP tasks and probe internal representations (e.g., logit-space activations) to empirically test whether ReLU models discover BP-equivalent internals or merely BP-compatible outputs. - Characterize conditions under which the implicit factor graph G(W) of a large pretrained LLM is approximately tree-structured, as this would determine the empirical reach of the no-hallucination corollary.