acceptodds
Under review as a conference paper at ICLR 2027

Posterior-Level Aggregation for Multi-LLM Multiple-Choice Decision Making: A Blackwell Perspective

Abstract

The rapid development of large language models (LLMs) has motivated research on answer aggregation in multi-LLM systems, where multiple models contribute predictions toward a shared decision. Existing aggregation approaches, such as voting and debate, are often developed without a common formal account of the information retained by their outputs. In this paper, we analyse multi-LLM answer aggregation (MLAA) using Blackwell’s informativeness framework. Within the Blackwell information-structure abstraction, we show that voting and debate induce information structures that are no more informative than the joint private information of all agents. More generally, this result establishes joint private information as a protocol-independent upper bound, while the corresponding ideal decision is based on the posterior conditioned on the joint private information. Motivated by this analysis, we develop a practical framework that estimates option-level distributions from sampled responses and combines them through generalised Bayesian aggregation. The framework is instantiated using product-of-experts and logarithmic opinion-pooling rules. Experiments on six multiple-choice QA benchmarks show that these posterior-level aggregation methods are competitive with or achieve higher accuracy than other multi-LLM voting and debate methods.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multi-LLM Multiple-Choice Decision Making

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_YPt3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper uses Blackwell’s informativeness framework to motivate posterior aggregation in multi-LLM systems. It estimates calibrated option probabilities from generated responses and combines them using PoE or learned LOP. Experiments on six multiple-choice benchmarks report competitive accuracy against voting and debate baselines.

Critical strengths

The modeling is clear and well motivated, and the theoretical framework is presented coherently.

The experiments cover a broad range of baselines, with useful ablations and evaluations across different model configurations.

Critical weaknesses

The connection between the Blackwell perspective and the proposed aggregation rules seems indirect, since the tractable Bayesian factorization is justified under conditional independence instead of any Blackwell structure family. Also, the results concerning voting and debate follow almost directly from the definition of garbling, which is interesting but somewhat not a substantial theoretical contribution.

The learned weights α\alpha provide an intuitive adjustment, but dependence among agents’ evidence may vary across questions. If one set of weights is shared across all questions, it cannot explicitly adapt the correction to this variation. The paper should clarify whether the weights depend on the question and better justify their interpretation as a correction for dependence.

The evaluation seems using only several hundred test questions per benchmark. It is unclear whether this reflects an experimental design choice or computational constraints. If computation is limiting, please clarify whether the proposed pipeline or the debate baselines account for most of the cost. Voting itself has negligible aggregation overhead once responses are available. Given the modest gains over some baselines, the tradeoff between accuracy and computational cost warrants closer examination.

Questions for authors

For relatively easy questions, are the agents’ choices nearly deterministic, and do the estimated posteriors reflect this through concentrated probability mass? Are the improvements primarily on difficult questions, or do easy questions also benefit? Is there any statistics related to performance gains and posterior concentration by question difficulty?

6 comments
AuthorsOpenReview ↗

Clarification of Connection

Clarification of Connection. The connection is through the target of approximation: Blackwell’s ordering establishes the joint private information as a protocol-independent information upper bound, and Bayesian decision theory identifies the corresponding pooled posterior as sufficient for optimal decisions. The result identifies a shared information upper bound, and what we did are derived from the theoretic results. Our method directly targets this posterior by estimating and combining per-agent option distributions. Conditional independence serves a separate purpose: it makes the ideal pooled posterior tractable through a product factorization; it is not required to establish the upper bound. With dependent LLMs and imperfect posterior estimates, PoE/LOP provide practical approximations to this target. Thus, although the garbling results follow directly from classical theory, they establish the reference point that guides our method.

AuthorsOpenReview ↗

Question-dependent weights

Question-dependent weights.

Instead of one scalar αm\alpha_m per agent, we learn a generator that outputs a weight vector for each question (new App. M, Tab. 11, p. 27). The generator does not read the question itself; its input is a vector of statistics of the agents' calibrated posteriors on that question. For agent m∈Mm \in M let Pm(⋅∣x)P_m(\cdot\mid x) be its calibrated posterior from Eq. 9, cm(x)=max⁡yPm(y∣x)c_m(x)=\max_y P_m(y\mid x) its confidence and hm(x)=−∑yPm(y∣x)log⁡Pm(y∣x)/log⁡Jh_m(x)=-\sum_y P_m(y\mid x)\log P_m(y\mid x)/\log J its normalised entropy. The feature vector is $ \phi(x)=\big[c_1,\dots,c_M,;h_1,\dots,h_M\big]\in\mathbb R^{10}\quad(M=5). $ The generator is a one-hidden-layer MLP gθ:R10→RMg_\theta:\mathbb R^{10}\to\mathbb R^{M}. The question-dependent weights are:

-- MLP-α\alpha: αm(x)=softplus(gθ(ϕ(x))m)\alpha_m(x)={softplus}(g_\theta(\phi(x))_m) is trained from scratch.

-- MLP-Finetune: αm(x)=αmLOPexp⁡(gθ(ϕ(x))m)\alpha_m(x)=\alpha_m^{\text{LOP}}\exp(g_\theta(\phi(x))_m) predicts a multiplicative correction to the global LOP weights and is initialised exactly at LOP, so any change from the global rule is learned from data.

So in both cases α(x)∈(0,∞)M\alpha(x)\in(0,\infty)^M, and the pooled posterior is Eq. 8 with αm\alpha_m replaced by αm(x)\alpha_m(x). None of the features uses the label or the question text, so α(x)\alpha(x) is available at test time from the same five responses per agent that PoE and LOP already use. Only θ\theta is trained; αLOP\alpha^{\text{LOP}} and the calibration head are fixed. The networks are trained per case with the negative log-likelihood on the small set in App. D.3. Neither gives a consistent gain: only 1 points over LOP on Professional Medicine, while worse than PoE on HellaSwag and CSQA, and the variant initialised at LOP moves in the same directions, so the labelled split pulls it away from weights that generalised better. The difficulty analysis (App. P) explains why: posterior concentration and agreement separate contested from uncontested items, but they do not distinguish a unanimous correct answer from a unanimous wrong one, which is the case a per-question dependence correction would need to recognise.

AuthorsOpenReview ↗

Test-set size and compute

Test-set size and compute. The sizes follow Choi et al. (2025a), so that the debate baselines match their published setting; Formal Logic and Professional Medicine use the full MMLU test splits. Cost is dominated by debate (App. S, Tab. 17): on MedMCQA, five-round debate uses 14.6k--23.1k tokens per item. PoE and LOP use the same five samples per agent as self-consistency (about 4.9k tokens) plus under one second of NLI scoring, so self-consistency is the cost-matched voting baseline; single-sample majority voting is cheaper (about 1k tokens). The average gains are modest partly because most items are easy: 67--76% are answered correctly by at least four agents and solved by nearly every rule. On contested items the differences are large (next answer).

Reviewer_4tQaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper analyzes multi-LLM answer aggregation (MLAA) through Blackwell’s informativeness framework. It establishes that existing protocols like voting and debate are garblings of agents' joint private information, acting as a protocol-independent upper bound. Motivated by this, the authors propose a practical framework using generalized Bayesian aggregation (such as product-of-experts and logarithmic opinion pooling) applied to per-agent posterior distributions estimated from sampled responses via semantic scoring and permutation-equivariant calibration. Experiments across six multiple-choice QA benchmarks show that these posterior-level aggregation methods are competitive with or outperform existing multi-LLM voting and debate approaches.

Critical strengths
  • Grounding multi-LLM answer aggregation in Blackwell’s informativeness framework provides a principled decision-theoretic perspective on the limitations of voting and debate.
  • The proposed posterior estimation and generalized Bayesian aggregation pipeline is thorough and empirically competitive across a diverse set of benchmarks.
Critical weaknesses

The framework assumes conditional independence or relies on heuristic weighting (LOP) to manage agent dependence, which diverges from the reality of LLMs trained on overlapping corpora.

Questions for authors
  • How sensitive is the performance of the learned weighting (αm\alpha_m) to the choice and size of the validation split used for training?
  • How does it translate to non-MCQ domain (long-text generation)?
3 comments
AuthorsOpenReview ↗

Conditional independence and heuristic weighting

Conditional independence and heuristic weighting. We agree that this is the central limitation and we measure it directly (Tab. 13, Page 29). The learned αm\alpha_m are not meant to correct dependence exactly; they are a low-dimensional tempering of each source’s contribution, fit by maximum likelihood on the labelled split.

AuthorsOpenReview ↗

Sensitivity of to the validation split

Sensitivity of αm\alpha_m to the validation split. We ran this during the response period (new App. N, Tab. 12, p. 28). We refit αm\alpha_m on 50%, 100% and 200% of the labelled split, redrawing the 50% subset for each seed. Because the labelled split is pooled across datasets, the 200% condition adds 100 items from the training/validation splits of HellaSwag, CSQA, HH-RLHF and MedMCQA; the MMLU Formal Logic and Professional Medicine validation splits (14 and 31 items) are subsampled with the rest in the 50% condition and used in full in the 100% and 200% conditions. Across these sizes, LOP accuracy changes by at most 0.8 points on any dataset, within the seed variance of the main table, and the ordering of αm\alpha_m across agents is the same at every size. With only M=5M=5 free parameters, αm\alpha_m is much less sensitive to the split than the calibration head, which we hold fixed here to isolate the effect of αm\alpha_m.

AuthorsOpenReview ↗

Non-MCQ domains

Non-MCQ domains. As mentioned in Sec.7, one promising extension is to model open-ended responses via a Dirichlet Process (DP) mixture over the embedding space. Loosely speaking, it is analogous to performing classification on an infinite number of categories. The DP property naturally groups semantically coherent open-ended answers, yielding clusters inferred from data rather than pre-specified. Each agent's posterior over discovered clusters is then estimated from the Monte Carlo sampling and our method applies directly via Eq. (7). We will leave this for future work.

Reviewer_fk1SOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a Blackwell-informativeness perspective on the aggregation of LLM agents predictions, viewing any aggregation as a garbling of the data that agents had access to during training. Based on this perspective, the paper proposes a connection (under conditional independence and common uniform prior) between two classical ways (PoE and LOP) of aggregating individual agent's predictions and the direct Bayesian inference from the pooled training data. Main experimental results suggest competitiveness of these approaches with predictions from individual agents, debate-based aggregation approaches and voting-based baselines. Experimental setup consists of several QA datasets and multiple LLMs acting as agents predicting labels.

Critical strengths

The paper brings a decision-theoretic, information-structure view (Blackwell) to the aggregation of predictions in multi-LLM systems. To my knowledge, this is the first such application specifically to LLMs.

The empirical study is broad: heterogeneous pools, as well as homogeneous ones, four debate families at three round counts, and three voting rules, six datasets, calibration, dependence and token-cost diagnostics.

As reported, the proposed aggregators match or exceed the best baseline in every configuration.

Critical weaknesses

As written, the theory does not support the method or explain the results:

The Blackwell analysis gives a common upper bound for all aggregation protocols, not an ordering among them. PoE/LOP estimates are themselves garblings of the agents’ information, so nothing leads to their superiority from this perspective. The connection to the best possible garbling in the form of exact bayesian inference from the pool data is tenuous. The proportionality in eq. 7 only holds for am=1a_m = 1, thus only for PoE. The factorization of likelihood is only valid under conditional independence of each agent's private data given (x,y)(x, y), which is hard to assume for modern LLMs. Moreover, each LLM’s predictive distribution must be the exact Bayesian posterior under a common prior, which is untrue, as training isn’t Bayesian conditioning. Calibration itself makes the comparison unfair, as it introduces new data to the information structure, thus making it not a garbling of the original private training data of the agents, and thus falls outside of the common upper bound on the informativeness of PoE/LOP (as implemented) and baselines. Simply put, the paper compares supervised and unsupervised approaches. Table 11 contains results under common calibration, and should probably be stated as the main experimental setting of this submission.

Generally, theoretic presentation is highly redundant. Propositions 1–2 say that any statistic of the agents’ data computed without the test label is a garbling, and hence no more informative. This holds for any predictor and for M=1M=1. Nothing in it is specific to LLMs, agents or aggregation, and it follows in two lines from Definition 4 and Theorem 1.

Questions for authors

Some of the table entries cannot be the mean of 1, 3 or 5 runs on the stated test sizes. For example, in Tab. 2 (Formal Logic) PoE is 0.7367 and LOP is 0.7445, but 3-seed means over 126 items must be multiples of 1/378. How were these computed, and were any items excluded?

5 comments
AuthorsOpenReview ↗

Derivation of Eq. (7)

Derivation of Eq. (7) We thank the reviewer for the comment, and we agree that some further clarifications are needed. We reiterate that Eq. (7) is to approximate the pooled posterior inspired from the generalized Bayes and logarithmic opinion pooling. In particular, the pooled posterior of interest can be written as: \begin{equation} \Pr(y | x, d_{1:M}) \propto \Pr(y | x) \Pr(d_{1:M} | y, x) \propto \Pr(y | x) \prod_{m = 1}^{M} \Pr(d_{m} | d_{< m}, y, x) \propto \Pr(y | x) \prod_{m = 1}^{M} [ \Pr(d_{m} |, y, x) ]^{\alpha_{m}}, \end{equation}

where: αm=ln⁡Pr⁡(dm∣d<m,y,x)ln⁡Pr⁡(dm∣y,x)\alpha_{m} = \frac{\ln \Pr(d_{m} | d_{< m}, y, x)}{\ln \Pr(d_{m} | y, x)}. This exponent depends on yy, on the shared context xx, and on the private information d1:md_{1:m}. This makes it both data dependent and hard to estimate. We, therefore, approximate αm\alpha_{m} by treating it as a hyper-parameter and learn it from data. This approximation is justified when the dependence among the private information is mostly redundant, i.e., double-counting of evidence. Furthermore, the approximation closely resembles the generalised Bayes and logarithmic opinion pooling. We also add this clarification and justification in the revision of our paper.

AuthorsOpenReview ↗

Conditional independence

Conditional independence. We agree that this is the limitation and we measure directly how far the agents are from it (Tab. 13, p. 30). Mean pairwise Yule's QQ between the agents' correctness indicators is 0.73 to 0.86 in the heterogeneous pools and rises to 0.87 to 0.98 in the homogeneous ones, and the gain from pooling shrinks accordingly: PoE is 2 to 3 points above the best baseline in Tab. 2 and within 1 to 2 points in the homogeneous and larger-model pools (Tabs. 6 to 9). The difficulty analysis in App. P shows the residual effect directly: on items that every agent gets wrong the pooled posterior is as concentrated as on items that four agents get right, because the agents share the same wrong answer and the product sharpens it.

AuthorsOpenReview ↗

Agents' distributions not being exact posteriors

Agents' distributions not being exact posteriors. We agree that training is not Bayesian conditioning, the agents do not share a prior, and an LLM's sampled responses are not draws from Pr⁡(y∣x,dm)\Pr(y\mid x,d_m). In the paper the Blackwell results are statements about information structures and hold whatever the agents' output is, since any label-free statistic of the private data is a garbling regardless of how it was produced. Different from MV or MAD, we directly approximate the pooled posterior, and treat the calibrated estimate Pr⁡~(y∣x,dm)\widetilde{\Pr}(y\mid x,d_m) in Eq. 9 as an approximation for the agent's posterior. Raw NLI scores and raw LLM logits are not probabilities in any Bayesian sense, and Tab. 3 shows what happens when the product is applied to them: PoE w/o calibration falls below output-level self-consistency (Tab. 2) on five of six datasets, all except HellaSwag, so the multiplicative rule has nothing to work with unless the per-agent estimates are reasonably calibrated. With calibration the per-agent ECE drops from 0.19 to 0.29 to 0.03 to 0.11 (App. K, Tab. 10). We also thank the reviewer read our experiments for comparison under the same calibrated supervision. We move the original Table 11 into the main text (new Table 3, Page 9). Majority voting and self-consistency applied to the calibrated outputs perform below PoE, suggesting that retaining and combining the approximated posteriors from the agents is useful relative to these hard-decision and frequency-based rules.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.