analiza aiqbench
This arXiv paper is primarily about aggregating answers from multiple LLM agents without labels, which is squarely AI and machine learning. Its core contribution is a variational inference method (BP-TD) that discounts correlated agents sharing a base model, so the aggregated answer is less biased by a bloc of near-duplicate models. The results are mixed: the method beats majority vote on MMLU only in some pool configurations, and with even 20 labelled questions, picking the best single agent outperforms every aggregator tested, so readers should treat the label-free gains as conditional rather than general.
Where Is the Truth When AI Agents Disagree? Dependence-Aware Belief Propagation for Unsupervised Truth Discovery (Extended Version)
arXiv:2611.04805v1 Announce Type: cross Abstract: LLM agents often give conflicting answers to the same question, and many deployed agents are variants of one base model, so their errors are correlated. We study truth discovery among such agents: inferring the true answer and each agent's reliability without labels. Classical truth-discovery and crowdsourcing estimators reward an agreeing bloc and can fall below majority vote. We propose BP-TD, which adds per-agent decorrelation weights to mean-field variational inference on a one-coin Dawid-Skene model, so that agents sharing a base model count together as about one vote. Its first iteration is exactly a weighted vote; later iterations estimate agent reliability. With 13 local agents on MMLU, BP-TD beats majority vote without any labels (0.716 vs. 0.624 on 160 questions; 0.707 vs. 0.674 on 10,000) and ranks agents by their true accuracy, which classical estimators invert. The gain depends on the pool: over 2,852 sub-pools of four benchmarks, decorrelation helps when the largest correlated bloc is weaker than the other agents and hurts when it is the strongest, and we found no label-free statistic that tells the two cases apart. Labels change the picture quickly: with 20 labelled questions, selecting the single best agent is more accurate than every aggregator we tested on MMLU, ARC and TruthfulQA.
Abstract:LLM agents often give conflicting answers to the same question, and many deployed agents are variants of one base model, so their errors are correlated. We study truth discovery among such agents: inferring the true answer and each agent's reliability without labels. Classical truth-discovery and crowdsourcing estimators reward an agreeing bloc and can fall below majority vote. We propose BP-TD, which adds per-agent decorrelation weights to mean-field variational inference on a one-coin Dawid-Skene model, so that agents sharing a base model count together as about one vote. Its first iteration is exactly a weighted vote; later iterations estimate agent reliability. With 13 local agents on MMLU, BP-TD beats majority vote without any labels (0.716 vs. 0.624 on 160 questions; 0.707 vs. 0.674 on 10,000) and ranks agents by their true accuracy, which classical estimators invert. The gain depends on the pool: over 2,852 sub-pools of four benchmarks, decorrelation helps when the largest correlated bloc is weaker than the other agents and hurts when it is the strongest, and we found no label-free statistic that tells the two cases apart. Labels change the picture quickly: with 20 labelled questions, selecting the single best agent is more accurate than every aggregator we tested on MMLU, ARC and TruthfulQA.
Extended version of a short paper accepted at the 2026 IEEE International Conference on Big Data (IEEE BigData 2026). 11 pages, 6 figures, 3 tables. Code: this https URL ; code and data archive: doi:https://doi.org/10.5281/zenodo.21294747
Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Databases (cs.DB); Machine Learning (stat.ML)
I.2.7; H.2.8; G.3
arXiv:2611.04805 [cs.AI]
(or arXiv:2611.04805v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2611.04805 arXiv-issued DOI via DataCite (pending registration)
Text integral sub licența deschisă a publicației · 271 cuvinte