acceptodds
Under review as a conference paper at ICLR 2027

A Unified Information-Geometric Framework for Collective Belief Dynamics in Multi-Agent LLMs

Abstract

Communication among LLM agents sometimes adds information and sometimes collapses a population onto a wrong consensus. We isolate one step of this question: a population of beliefs is pooled and is also updated on common evidence, and we ask when the order matters. Logarithmic (-) pooling commutes with a common Bayes update, whereas arithmetic (-) pooling reallocates the social weights in proportion to the evidence normalizers; we characterize the exact zero set of this order gap, bound it, and show that it is zero or fourth order near consensus. Because the order gap is not a quality ordering, we score both orderings against an independently declared finite observation model: update-then-pool has higher expected log loss in all 18 cells (exact enumeration), and at matched observation budget no single rule is best across six information structures. Stated as operators on an information-geometric map of predictive states, the same language also closes the second moments of sampled communication, which includes the sampled convex-combination rule used to model memetic drift. A preregistered one-step transfer test on native Qwen2.5 receivers failed its gate, so we do not claim that native agents follow these operators.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_8u3bOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how belief pooling interacts with Bayesian updating in multi-agent systems. It derives conditions and bounds for the differences between the two operations and evaluates them mainly on synthetic models. The tests on Qwen2.5 agents fail to support the proposed framework.

Critical strengths
  1. The paper provides some useful mathematical results on the differences between pooling and updating.
  2. The authors openly report the failed LLM experiments and acknowledge the limitations of their results.
Critical weaknesses
  1. Much of the paper builds on existing opinion-pooling theory. While there are additional mathematical results, I am not convinced they offer enough new insight. The authors should explain more clearly what is genuinely new and why it matters. The revised paper also withdraws its claimed improvement over Tanaka (2026).

  2. Most experiments use synthetic models to confirm mathematical results. Both Qwen2.5 models fail the transfer tests, and the planned population-level evaluation is not completed. The authors should provide stronger evidence that the framework works for real LLM agents and explain why these older models were chosen.

  3. I still do not see how the proposed results would help improve real multi-agent systems. The authors should provide concrete examples or experiments showing when the framework is useful. The paper is also unnecessarily long and complicated given its main findings.

Questions for authors

see Weaknesses

1 comment

Thank you for your careful review and constructive comments.

Q1. §§2–3 give global two-sided symmetric-KL bounds, a scale-invariant lower certificate, and quartic-or-zero behavior on the stated linear family; the exact zero set allows dependent posteriors. §4.1's quality ordering is restricted. The normalizer-only upper bound (§3) follows from divergence contraction; priority checking is incomplete.

Q2. No native mechanism is validated: both Qwen2.5-3B/7B transfers failed; population transfer was not run (§7; general response, point 2). We ran no newer model; model age is not an established cause.

Q3. It was not necessary: the revision keeps three propositions, one corollary and one theorem (§§2–4), using direct simplex proofs and separating discrepancy from quality. Unification/population extensions are removed; §5/App. B correct attribution/movement.

Q4. None for native agents. Table 2 compares rules against a joint-law oracle; three product entries are uncomputed and the local baseline is not all-information/token-matched. App. D is a finite synthetic check, not native validation; speed/horizon limits remain (general response, points 3–4). The retained title exceeds our explicit-probability scope.

Reviewer_tL3jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the problem of belief (only on a finite hypothesis set) aggregation and update problem on multi-agent systems. Two classic procedures are: first update then aggregate, and first aggregate then update. Belief update uses a shared likelihood and follows Bayes law. Aggregation has two styles: multiplicative (e-pool) and additive (m-pool), the paper show that these can be unified as special cases (with deterministic kernels) of a more general multi-agent communication model with two different settings, where in the communication agents "talk to" each other and update their beliefs accordingly. The framwork also consider the QSG theory as a special case with probabilisitic kernel. Moreover, the limit under infinite sampling in a single QSG commnication can be considered as a two-agent m-pool with given weights.

The paper proved that e-pool commute with update, while m-pool and QSG fails to commute. For m-pool, the paper proves 1. the communator (update after aggregation minus aggregation after update) can be represented as a linear combination of the posteriors of each agent in a update, 2. the coefficents in the communator decomposition has certain bounds, 3. when agents agree enough, the bounds are 4-th order close to 0 in disagreement. For QSG, paper also derives closed second-moment dynamics, consensus probabilities and N-scaling.

Experiments are conducted on a 2-agent 18-cell experiment setup, with hypotheses size 3 or 5. Theoretical results are numerically verified. Besides, further experiment show that there is no "better" aggregation-update order, they are just different in m-pool / QSG (for e-pool, commutativity make two orders exactly the same). The real-LLM experiment fails to provide evidences that LLM agents follow the constructed operators.

Critical strengths
  1. The paper turns the aggregation-update ordering problem into a computable argument, by writing the commutator of m-pool exactly and decompose them into linear combination of posteriors.

  2. The bound estimation and order argument on m-pool are sharp and strong.

  3. The paper shows in experiment that the non-commutativity does not lead to the conclusion that one order better than the other.

Critical weaknesses
  1. The core mathematical results has a limited novelty and contribution. The most novel part, the bound estimation of commutator coefficients, are unfortunately not a key evidence to a better aggregation-update order.

  2. The paper claimed a unified framework among m-pool, e-pool and QSG, but there are no shared theorems among the three, and the nontrivial results work on each case themselves.

  3. There is no theoretical strong connection between the unified theory and their LLM experiment setup.

  4. In the title, beliefs dynamics in multi-agent LLMs are claimed, but the theory is not related to that claim. The only LLM experiment results in a negative result on its relation to the theory, making the title misleading.

  5. Only one-step update are considered, and the likelihood are force to be shared exactly, the setup is too simple for any further applications.

Questions for authors
  1. Is there a concrete application of the bound χ\chi?

  2. What can be proposed as the reason of failure in the LLM experiment?

  3. Could there be similar bounds in QSG? χ\chi can also be defined in a 1-round communication in QSG, but the dynamics in communication for more than 2 agents may be too complicated.

  4. There could be multiple generalizations of this thoery, if only deterministic kernels are considered, such as multi-round update, private likelihood, etc.

1 comment

Thank you for your careful review and constructive comments.

Q1. §3 gives a TV-tolerance/margin-preservation check, not correctness or an order-quality decision (§4). App. D is synthetic, not population validation; screens were slower. No shared theorem unifies pooling and QSG; the retained LLM title exceeds our explicit-probability scope (general response, points 1, 3, 5).

Q2. Native failures remain unexplained; a history-rich state is an untested possible remedy, not a diagnosis. Newer-model success and answer-only versus history are untested (§7; general response, point 2).

Q3. No sampled order-gap theorem is proved: §5 attributes evidence-free equations to Tanaka; a transfer needs a specified evidence kernel, coupling and divergence target, and strict positivity can fail.

Q4. Eq. 8/App. G bounds TV from identical initial states using reference-path local errors, predetermined pooling and bounded private likelihoods shared across paths, not the common-likelihood certificate. It excludes adaptive inputs and can be vacuous (from round 7 when 2b/a = 2 and ε = 0.01); it guarantees neither KL nor utility. App. G's separate squared-loss example does not prove universal utility ordering.

Reviewer_r2caOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether pooling categorical beliefs commutes with updating them on a common likelihood. Logarithmic pooling commutes with the update, whereas arithmetic pooling changes the influence weights in proportion to the agents' evidence normalizers. The paper characterizes the exact zero set of the resulting symmetric-KL order gap, supplies bounds, and derives a quartic-or-zero expansion along a specified near-consensus family. Exact finite experiments compare the pooling orders against a separately constructed task posterior. A second component expresses communication as a speaker kernel followed by a receiver update and analyzes sampled convex-combination dynamics, including moment equations, consensus probabilities, and relaxation rates. A log-likelihood-vector embedding supplies a geometric representation. The attempted transfer to native LLM receivers failed its first validation gate, and native results are reported as exploratory.

Critical strengths

1. A clean, precisely answered pooling question. The operator-order question is well posed and the answer is exact. Equal evidence normalizers are both necessary and sufficient for a zero gap, and the proof is short and neat. The bounds and the near-consensus expansion are explicit, and the authors are careful to separate "the coefficient is positive" from "the gap is identically zero."

2. Honest separation of what is shown from what is not. The quality target comes from a declared joint observation law rather than from one of the pooling outputs, which is the right way to compare the two orders. The authors also report their failed preregistered transfer test and their rejected predictors openly. This makes it easy to see exactly what the synthetic results do and don't support.

Critical weaknesses

1. A central population result is already in prior work, and the "improvement over Tanaka" is measured against the wrong baseline.

Proposition 7's exact second-moment system is already in Tanaka (2026), Appendix A.8–A.9, equations (32)–(33). The only difference is normalization: Tanaka sums disagreement over agents, while this paper averages it. Substituting

UTanaka=R,VTanaka=NVU_{\text{Tanaka}} = R, \qquad V_{\text{Tanaka}} = N V

turns his equation (32) into the paper's first moment equation, and his equation (33), divided by the population size, into the second one, term for term.

The "added value over Tanaka" gate in Table 3 and the RMSE comparison in Figure 11 are both against Tanaka's simplified mean-field model, not his exact equations. So they show an advantage over an approximation, not over what Tanaka actually derived. As written, this overstates the advance over prior work.

Some later results may still be new: the relaxation-rate approximation and its N-scaling (Prop. 9), the consensus theorem for general sample counts and adaptation rates (Thm. 2), the conserved weighted mean (Prop. 8), and the bounds and expansion (Props. 3–4). These should be positioned on their own. But the exact moment closure itself is not a new contribution, and that removes a large part of what the paper presents as its advance.

2. The pieces don't add up to an explanation of the problem the paper sets out to address.

Two separate theories under one notation. The deterministic part assumes explicit numerical pooling and a known common likelihood. The population part assumes evidence-free sampled messages and a fixed convex receiver rule. Writing both as "kernel then update" is fine, but Theorem 1 mostly lists their separate properties. No result uses evidence, sampled communication and predictive quality together. The geometric embedding is not used in either the order-gap results or the population equations, which work in plain simplex coordinates.

The state can lose the information communication is supposed to carry. The paper's own XOR example (Prop. 15) shows that agents can have identical answer distributions while holding complementary information that solves the task once it is shared. A theory built on answer distributions can therefore miss exactly what makes communication useful. The authors acknowledge this and propose a richer history/probe state, but that state is not tested.

The quality results lack an explanation. In E1, pool-then-update wins in all 18 cells, but the independence product is the exact posterior there. So the comparison ranks two approximations of a known optimum, with no condition explaining why one wins. The random-target experiment adds little: any two different predictions can be ranked either way by a suitably chosen target.

The LLM connection is not established. The native transfer gate fails for both models. A theory paper doesn't need benchmark gains, but it does need either a justified abstraction of a real LLM protocol or a substantial result in its narrower setting. The caveats in the paper are appropriate, but they don't supply that missing link.

Questions for authors

Clarifications that could change my assessment:

  1. Prior work. Given that Proposition 7 matches Tanaka's exact appendix equations, what is new in the population analysis? Why is the "added value" comparison made against his mean-field approximation rather than his exact equations?

  2. Unification. Is there a result that needs the pooling theory, the sampled dynamics and the geometric map together? Is there a concrete LLM protocol where a common likelihood, an explicit pooling step and a sufficient belief state are justified?

  3. Why pool-then-update wins in E1. Can you give a condition on the observation model that explains this? For example, if the pooled prior were the true prior and the likelihood correctly specified, pool-then-update would be exactly Bayes-optimal. E1 is not that model, so a condition that fits E1 or another motivated task family would be much more convincing than the random-target experiment.

  4. Response coefficient versus actual movement. Section 3.2 argues that a listener-only response that increases with size makes the largest listener move the most. Here is a counterexample with binary beliefs, four sampled messages per interaction, and a uniformly chosen speaker:

x=(0.2,0.8,0.5),s=(1,2,3),g(s)=14+s2(1+s)x = (0.2, 0.8, 0.5), \qquad s = (1, 2, 3), \qquad g(s) = \frac{1}{4} + \frac{s}{2(1+s)}

The expected movements of the three agents when listening are

(E∣Δx1∣,E∣Δx2∣,E∣Δx3∣)=(0.2314,0.2700,0.1965)\left( \mathbb{E} \lvert \Delta x_1 \rvert, \mathbb{E} \lvert \Delta x_2 \rvert, \mathbb{E} \lvert \Delta x_3 \rvert \right) = (0.2314, 0.2700, 0.1965)

The largest agent has the largest response coefficient but moves the least, because it sits near the others. What exactly is P2 meant to claim: a statement about coefficients, about movement under matched incoming discrepancies, or about an average over some initial-state distribution? Under which assumptions does the claimed need for speaker–listener dependence hold?

1 comment

Thank you for your careful review and constructive comments.

Q1. Nothing new remains in the population analysis: Prop. 7 rederived Tanaka's Eqs. 32–33. §5/Table 3/Fig. 2 compare his exact equations; mean field is a labelled approximation. We withdraw superiority and remove relaxation, conserved-mean and consensus extensions pending prior-work comparison.

Q2. No to both. Pooling uses a known common likelihood; sampled dynamics are evidence-free; no geometric embedding is needed. XOR exposes lost information; native sufficiency/history remain unvalidated (§7; general response, points 1–2).

Q3. Yes: all 18 E1 cells satisfy Theorem 1 (§§4.1–4.2): uniform label, two equally accurate symmetric private channels, a symmetric common channel, all independent given the label, equal weights, and private accuracy above chance. PU has no larger conditional expected log loss, not a realized-label guarantee; the independence product is the oracle. Table 1 gives failures outside these assumptions. The random-target experiment is removed.

Q4. App. B retains your counterexample: coefficients order movement only for matched incoming discrepancies, norm and sampling; unconditional speaker–listener necessity is withdrawn.

Comments on the paper

1 comment

We thank all reviewers for their careful and constructive reviews.

  1. We withdraw population novelty/superiority over Tanaka (§5), coefficient-to-movement claims (App. B), and unification (§§1,7). Pooling identities are classical; priority checks incomplete.
  2. Native transfers, gain-sign and receiver-gain predictors failed; population transfer was not run. The pilot evaluated zero worlds without full probabilities; answer-only versus history is untested (§7/App. A).
  3. All four screens were slower than direct TV; the normalizer-only screen ran at 0.300× its speed (3.3× slower; Table 4).
  4. Eq. 8 bounds finite-horizon TV for predetermined inputs; it may be vacuous and guarantees neither KL nor utility.
  5. The title is broader than our results: explicit probabilities, not native agents. Blue marks cited changes in this substantially rewritten manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.