Skip to content
#small language model Book Open access

Bethe Consensus Decoding: Joint Inference on Empirical Pairwise Factor Graphs for Structured LLM Outputs

Nov 2026 · Proceedings of the 35th ACM International Conference on Information and Knowledge Management · 0 citations · 9 references

TL;DR

BCD is introduced, a training-free post-hoc aggregator that reconciles K samples drawn from a frozen LLM by minimising a Bethe free energy on a sparse pairwise factor graph and shows that on tree-structured factor graphs in the population limit, BCD strictly dominates IMV in the induced tree distribution.

Abstract

Large language models increasingly serve as zero-shot extractors for structured outputs whose fields are mutually dependent: a salary range is correlated with experience level; a destination is correlated with the airline. The dominant unsupervised aggregator, independent majority voting (IMV), decodes each field marginally and therefore commits combinatorially inconsistent outputs whenever the per-field modes do not co-occur. We introduce Bethe Consensus Decoding (BCD), a training-free post-hoc aggregator that reconciles K samples drawn from a frozen LLM by minimising a Bethe free energy on a sparse pairwise factor graph. The graph is constructed from empirical mutual information using a cardinality-aware χ2 threshold, sum-product loopy belief propagation with damping yields pseudo-marginal beliefs, and a sequential greedy clamping procedure with a pairwise-belief ratio correction extracts a globally coherent output. We show that on tree-structured factor graphs in the population limit, BCD strictly dominates IMV in the induced tree distribution. On five real datasets ranging from slot filling to 15-field resume parsing, BCD with Llama-3-70B-Instruct improves average exact-match accuracy by 4.6 points over IMV and by 2.2 points over the strongest unsupervised structured baseline. We additionally evaluate BCD on Llama-3-8B, GPT-4o-mini, and Gemini-1.5-Flash and observe strong absolute performance on each (up to 44.96 EM on GPT-4o-mini), though backbone-matched IMV baselines are out of scope for this study. BCD is competitive with a supervised pairwise CRF that uses the same graph topology: the margin is small (≤ 1.3 pp) on the three data-rich benchmarks (D1, D2, D4) and substantially larger (3.7--4.7 pp, statistically significant under Holm--Bonferroni correction) on the two data-scarce benchmarks (D3, D5). Aggregation runs in milliseconds on CPU and reduces to IMV exactly when the empirical edge set is empty.

Read PDF

Similar papers

Preprint Oct 2026

Quantum-Inspired Pairwise Ranking Aggregation via Active LLM Feedback

Ranking aggregation from multiple retrievers is a powerful yet under-explored technique in modern retrieval and RAG systems. We introduce LLM-IRA, the first extension of the quantum-inspired interactive ranking aggregation framework to document reranking. By leveraging large language models (LLMs) solely for pairwise j...

Quan Mai, Susan Gauch, Gibson Nkhata · 0 citations
#artificial intelligence Preprint Sep 2026

Greedy Decoding Is Not Precision-Invariant: Cross-Precision Output Divergence in LLM Inference

An empirical error-propagation analysis is developed and finds that 22 layers of accumulated body error do not distinguish flipping from non-flipping steps; the outcome depends primarily on the top-two logit margin at the LM head relative to the directional perturbation between the top-two candidates.

Gao-Yuan Du, A. Khan, Rex Zhou et al. · 1 citation
#machine learning Preprint Sep 2026

GraphVQ: Structure-Aware Autoregressive Decoding over Context-Quantized Graph Tokens

Graph foundation models need a discrete token representation, but casting a graph as a generatable token sequence faces a structural obstacle: edges spanning beyond the serialization window cannot be emitted in one pass--so one-pass autoregressive generators systematically under-produce cycles--and a single global cond...

Yu-Xi Yao, Zi-Jun Zhao · 0 citations
Book Open access Nov 2026

Cavity-Calibrated Prediction Sets

Cavity-Calibrated Prediction Sets (CCPS) are introduced, which prove distribution-free marginal and joint coverage guarantees and establish a correlation-decay condition under which cavity scores converge to true leave-one-out scores.

Ze-Hua Cheng, Wei Dai, Jia-Hao Sun · 0 citations
#artificial intelligence Preprint Sep 2026

Composable Decoding on the Probability Simplex: Theory and Implementation

CompoSimplex is introduced, a library with configurable support rules, regularisation primitives, and simplex solvers for constructing and evaluating compositional decoders, and it is shown that compositions can realise trade-offs between single-sample quality, multi-sample quality, and diversity that are not attained...

Xiao-Tong Ji, Ahmed Khaled Khamis, Rasul Tutunov et al. · 0 citations
Open access Sep 2026

Efficient Attention Pruning for Resource-Constrained Super-Resolution Using Statistical Thresholds and Quantum-Guided Sparsity

Non-local attention improves single-image super-resolution (SISR) by letting every query position aggregate information from all other positions, but the softmax normalisation at its core assigns strictly positive weight to every candidate regardless of relevance, diluting genuinely informative long-range correlations...

L. N. K. Reddy, K. P. Priya · 0 citations

Related blog posts

Google DeepMind Blog Nov 18, 2026

Genie 4 is now available

We are rolling out a new Project Genie experience to eligible Google AI Ultra subscribers, opening access to our new Genie 4 world model.

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.