{"id":"4626ef90-bdb5-41b0-94a0-1848c56aeb1b","arxiv_id":"2608.01367","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Moderate imprecision in belief fusion, a Jaccard threshold between about 1/3 and 1/2, improves collective learning accuracy when populations start strongly biased toward incorrect beliefs.","lead":"This paper models social learning where each agent's belief is a set of possible worlds, and agents combine beliefs using a fusion operator whose level of imprecision can be tuned. It finds that when most agents start firmly committed to a wrong belief, a moderate amount of fusion imprecision helps the whole group converge to the correct belief, provided evidence is only moderately noisy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central quantitative claim rests on an uncheckable stability-region calculation: Eqs. (9)-(13) and the 'only four fixed points' assertion are stated without derivation or code, so the R1-area benefit in Figure 3 is not independently verified.","rationale":"The paper is internally coherent: spot-checking the updating matrix (A.1), the gamma = 0 fusion matrix (B.1-B.2), and Eq. (8) shows consistency with the worked example for B5, modulo a harmless transposition convention. The mechanism by which moderate imprecision (gamma in [1/3, 1/2)) preserves the conflict-generated belief S = {all states}, whereas precise fusion collapses S intersect B4 to B4, is real and is qualitatively confirmed by the agent-based simulations. However, the headline strength of the abstract rests on a quantitative stability-region comparison. That comparison is not independently checkable from the manuscript: the Jacobian/eigenvalue derivation is absent, the 'only four fixed points' assertion is unproved, several captions and labels are inconsistent, and no code is shipped. The reader's verdict of CONDITIONAL is appropriate. My concern does not move the verdict but sharpens the condition: the authors should reproduce Eqs. (9)-(13) from Eq. (6), or ship the derivation/code, before the R1-area benefit claim can be taken as established.","tokens_in":22473,"tokens_out":11809,"duration_ms":119465,"concrete_test":"Independently derive the Jacobian of Eq. (6) analytically or with computer algebra at Pt, Ph and Pf for gamma = 0 and gamma = 1/3, compute eigenvalues, and reproduce the stability boundaries in Eqs. (9)-(13) and the R1-area curves in Figure 3. Separately, perform a dense numerical search for fixed points or attractors of Eq. (6) from thousands of random initial conditions at representative (sigma, rho) points in each claimed region, especially low-rho R3 points. If the derived boundaries differ from Eqs. (9)-(13), or if any stable attractor besides Pt, Ph, Pf is found, the central R1-area claim fails. If the boundaries match and no extra attractors appear, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative core of the abstract is the R1-area comparison in Figure 3: about 50% of (sigma, rho) at error rate 0.3 for gamma in [1/3, 1/2) versus about 30% for gamma = 0. This number comes entirely from the Section 4 stability analysis: the assertion that only Pt, Ph1, Ph2 and Pf are fixed points, and the eigenvalue thresholds summarized without derivation in Eqs. (9)-(13). No Jacobian calculation or code is provided, and Appendix B's fusion matrix is only for gamma = 0, so the gamma = 1/3 boundaries used for the purple curve cannot be checked. The Section 5 difference-equation 'simulations' iterate exactly the same Eq. (6), so they cannot validate the boundary equations; the Section 6 agent-based simulations are an independent qualitative check, but they only compare gamma = 0 versus gamma = 1/3 on a grid of (sigma, rho) and do not measure R1 area. If the eigenvalue derivation contains an error, or if additional stable fixed points or attractors exist—R3 is left uncharacterized—then the central quantitative benefit claim is unsupported. This is a checkability gap in the most load-bearing part of the argument, not an accusation of error.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-agent social learning model in which each agent's belief is a subset of possible worlds, agents update on noisy singleton evidence via Eq. (4), and agents combine beliefs pairwise using a Jaccard-similarity-threshold fusion operator ⊙_γ defined by Eq. (2). For the two-proposition case, the authors derive a 15-dimensional difference equation (Eq. (6)), assert that the only fixed points are the four precise consensus states P_t, P_{h1}, P_{h2}, P_f, and partition the (σ,ρ) parameter space into four stability regions R1–R4 (Table 3). The central quantitative claim is that for γ∈[1/3,1/2) the R1 region—where only the correct consensus is stable—occupies about 50% of the parameter space at error rate 0.3 versus about 30% for precise fusion γ=0 (Figure 3). This claim is checked using difference-equation simulations (Section 5) and agent-based simulations for γ=0 and γ=1/3 (Section 6). The authors conclude that moderate fusion imprecision helps initially wrong-biased populations reach the truth under moderate evidential noise, at the cost of slower consensus and degraded performance when the population is already correct.","tokens_in":22699,"tokens_out":7821,"duration_ms":78997,"significance":"If the stability analysis is correct, the paper reports a non-obvious and potentially useful result: imprecision in belief fusion can act as an escape mechanism from an absorbing wrong consensus, and this effect is robust enough to persist in agent-based simulations. The model is clearly specified, and the appendices provide explicit transition matrices for the updating operator U and for the γ=0 fusion operator, which allowed me to verify the internal consistency of those components and of the Hamming-distance accuracy vector in Section 5. The paper is also honest about several limitations, such as the well-mixed population assumption and the restriction to two propositions. However, the central quantitative result—the R1-area comparison—rests on an eigenvalue analysis that is not shown, and the agent-based simulations are only a qualitative check. The contribution is therefore significant but not yet fully verifiable in its present form.","major_comments":[{"comment":"The R1-area benefit quoted in the abstract and in Figure 3 comes entirely from the stability boundaries listed in Eqs. (9)–(13). These equations are stated without derivation, and Appendix B provides only the γ=0 fusion matrix; no Jacobian or eigenvalue calculation is supplied for γ∈[1/3,1/2), the case used for the purple curve in Figure 3. The Section 5 difference-equation simulations iterate exactly Eq. (6), so they cannot validate the boundary equations. Please provide the Jacobian/eigenvalue derivation, or a reproducibility artifact (e.g., symbolic or numerical code) for at least γ=0 and γ=1/3 at the error rates considered. Without this, the central 50%-vs-30% R1-area claim is uncheckable.","section":"Section 4, Eqs. (9)–(13), and Figure 3"},{"comment":"The assertion that P_t, P_{h1}, P_{h2}, P_f are 'the only fixed points' is made in one sentence and is not proved. The argument 'any difference in beliefs between agents will result in further fusion, and any belief imprecision will result in further evidential updating' does not exclude heterogeneous or imprecise equilibria. Moreover, R3 is left uncharacterized, and the γ=3/4 simulations in Figures 5c–5f show persistent diversity and high entropy/imprecision, indicating nontrivial attractors. If those attractors are fixed points rather than cycles, the 'only four fixed points' statement is false; if they are cycles, the classification of R3 should say so. Please clarify and prove the fixed-point claim, or restrict the stated classification accordingly.","section":"Section 4, paragraph on fixed points"},{"comment":"The reported advantage of imprecision depends critically on the asymmetric evidential update: a precise wrong belief B={s4} that receives correct evidence {s*} simply ignores it, so a wrong precise consensus is absorbing. Imprecise beliefs that already contain s* accept the evidence and shrink toward the truth. Consequently, the entire benefit of imprecision flows from this asymmetry. The paper would be strengthened by an explicit discussion of this scope condition and by a sensitivity test with a symmetric updating rule (e.g., contradictory evidence forces belief change) or disjunctive evidence. Without such a test, the claim 'imprecision improves social learning' is tied to a specific, arguably restrictive, evidential assumption.","section":"Section 2.3, Eq. (4)"}],"minor_comments":[{"comment":"The text says 'the correct belief, i.e. B={s4}', but s4=(0,0) is the completely incorrect state; the correct belief should be {s1} or {s*}. Similarly, Case 5 is described as adopting 'one of the half correct beliefs', but Table 4 assigns 0.85 probability to B3 alone. Please align the text and table.","section":"Section 6, Case 6"},{"comment":"Table 4 lists Case 1 as (0.05,0.01,0.01,0.97,0), but Section 5 defines P1_0=(0.01,0.01,0.01,0.97,0). Also, Section 6 says it 'applies a lower initial bias' of 0.85, but Table 4 includes both 0.97 and 0.85 cases; clarify which cases correspond to which initial probabilities.","section":"Table 4 and Section 5"},{"comment":"The sentence 'The full fusion transition matrix is given in Appendix A for γ=0' points to the wrong appendix: Appendix A is the updating matrix, while the fusion matrix appears in Appendix B. Also, the notation for the fusion matrix in Appendix B (F and F') should be aligned with F^P_γ in Eq. (6).","section":"Section 3, Eq. (8) and appendices"},{"comment":"The caption says 'The red and purple curves show the area of R1 for γ∈[1/3,1/2) and γ=0, respectively.' This is reversed with respect to the body text, which states that red is R1^0 and purple is R1^γ. Please correct the caption.","section":"Figure 3 caption"},{"comment":"Minor typos and formatting issues: 'tends to generates' (abstract), 'indpendent' (Section 6), 'Wehaveinvestigated' (Section 7), and inconsistent spacing in equations. These do not affect the technical content but should be cleaned up.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the model specification is largely workable, but the load-bearing stability-region calculation is not independently verifiable in the current manuscript. My recommendation of major_revision is driven by the need to see the Jacobian/eigenvalue derivation or a reproducibility artifact for the γ values used in Figure 3, and to clarify the fixed-point/attractor classification in R3. I do not see grounds for rejection: the appendices and simulations I checked are internally consistent, and the qualitative agent-based results support the claimed phenomenon. If the derivations are supplied and the scope condition of Eq. (4) is discussed, I would be willing to accept."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this as a moderately interesting result in social learning and opinion dynamics. The genuinely new piece isn't the fusion operator (from their ALIFE 2021 paper) or the set-of-worlds representation (from the literature); it's the systematic four-region stability partition of the (sigma, rho) plane and the claim that gamma in [1/3, 1/2) enlarges the region where only the correct consensus is stable, peaking around epsilon = 0.28. The paper does that part honestly: the model is specified precisely enough that the reader's check of the 15x15 update and fusion matrices goes through, and the difference-equation and agent-based simulations line up qualitatively with the stability regions. The mechanism—imprecision converts disagreement into larger belief sets that then accept correcting evidence—is real within this model, and the paper doesn't oversell it as more than that.\n\nThe soft spots are where the reader's stress-test lands. The eigenvalue analysis behind Eqs. (9)–(13) and the \"only four fixed points\" assertion is stated without derivation. No Jacobian, no code, and Appendix B only gives the gamma = 0 fusion matrix, so the gamma = 1/3 boundaries in Figure 3 can't be independently checked. The Section 5 difference-equation simulations are not an independent test because they iterate the same Eq. (6). The agent-based simulations are an independent qualitative check, but they don't measure R1 area. So the headline number—50% vs 30% at epsilon = 0.3—is unsupported as presented. That's a checkability gap, not evidence of fraud; still, if those boundary equations are wrong, the paper's central quantitative claim collapses. I'd want the derivation or the code before I'd sign off on it.\n\nOther soft spots are minor but real: R3 attractors are uncharacterized; the benefit depends on an evidence-updating asymmetry that lets a precise wrong belief ignore contradicting evidence; the well-mixed assumption is stated as a limitation in Section 7; and there are small internal inconsistencies (Table 4 sums to 1.04, Case 6 labels the correct belief as {s4}, Eq. (8) has a spurious P(i), a cross-reference points to the wrong appendix, Figure 3 caption swaps colors). None of these are load-bearing, but they suggest a cleanup pass.\n\nWho is this for? Researchers in collective decision-making and swarm intelligence who want an analytical handle on imprecision in fusion. It's not transformative, but it's a solid step within that subfield. I'd send it to peer review rather than desk-reject, because the model and the claim are serious and the verification gap is fixable. If the derivation is supplied, this becomes a credible paper; if not, the benefit result should be treated as a conjecture.","headline":"Worth a referee's time: the claim is plausible and internally consistent, but the load-bearing stability calculation is not shown, so I'd want the derivation or code before trusting the benefit curve.","tokens_in":23323,"tokens_out":1733,"would_cite":false,"duration_ms":18634,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Moderate imprecision in belief fusion helps a population that starts out strongly committed to a wrong belief converge collectively to the correct state of the world.","keywords":["social learning","imprecise belief fusion","propositional beliefs","collective decision-making","difference equations","stability analysis","agent-based simulation","Jaccard similarity"],"falsifier":"Rerun the Case-1 difference-equation and agent-based simulations (97% initially at the wrong state, e.g. sigma = rho = 0.5, epsilon = 0.4) with a modified update rule: when evidence E contradicts belief B, let the agent adopt E (or some combination) instead of ignoring it. If the accuracy gap between gamma = 1/3 and gamma = 0 closes or reverses, the reported benefit is a consequence of the ignore-contradictory-evidence asymmetry rather than of fusion imprecision itself. A second check: repeat the agent-based runs on a small-world or lattice network; if the R1 advantage disappears under local i","tokens_in":22196,"feed_emoji":"🎯","tokens_out":7151,"duration_ms":67498,"temperature":0.7,"pith_summary":"The paper tries to establish that fusing agents' beliefs imprecisely, so that disagreement produces broader, less committed beliefs rather than a sharpened one, can help a population that starts out strongly committed to a wrong belief converge to the truth. The claim is supported by a difference-equation model whose fixed points are analysed for stability, together with agent-based simulations for finite populations. At moderate fusion imprecision (a Jaccard-similarity threshold between 1/3 and 1/2 in a two-proposition language), the region of the fusion-rate/evidence-rate plane in which the correct consensus is the only stable fixed point roughly doubles compared with precise fusion at an evidence error rate of 0.3, growing from about 30% to about 50% of the parameter space. The same imprecision improves measured learning accuracy when 70-97% of agents start with incorrect beliefs, while slightly slowing consensus when the population is already correct, and too much imprecision prevents consensus altogether.","feed_headline":"Moderate fusion imprecision lets biased groups find the truth","feed_subtitle":"In noisy settings, imprecise belief merging beats precise merging when most agents start wrong.","key_machinery":"The engine is the similarity-threshold fusion operator B1 ⊙_gamma B2, which returns the intersection when Jaccard similarity J(B1, B2) > gamma and the union otherwise, with gamma in [0,1] controlling imprecision (gamma = 0 is precise intersection-union fusion, gamma = 1 is pure union). It works together with the evidential update B | E = B ∩ E when that intersection is non-empty and B otherwise, plus the rule that precise agents stop collecting evidence. Because a wrong precise belief rejects contradictory evidence, a false consensus is absorbing under precise fusion; moderate imprecision generates unions that contain the true state, and those imprecise beliefs then accept correct evidence a","core_discovery":"Within a model where each agent's belief is a set of possible worlds and fusion is a similarity-threshold operator, the paper shows that moderate imprecision can be actively beneficial rather than merely tolerated. For the two-proposition case it derives the eigenvalues of the Jacobian at the consensus fixed points (correct, partially correct, false) and maps four stability regions in the (fusion rate sigma, evidence rate rho) plane. As the imprecision parameter gamma moves from 0 (precise intersection-union fusion) into [1/3, 1/2), the region R1 where only the correct fixed point is stable expands, while the region where all fixed points are stable, and initial bias therefore decides the ou","pith_inferences":["A testable corollary of the absorbing-wrong-consensus explanation: if agents updated on contradictory evidence (e.g. switching to the evidence set), the reported gap between gamma = 0 and gamma = 1/3 should shrink substantially, since precise fusion would then escape false consensus on its own.","The mechanism suggests a design heuristic for multi-agent systems: tune fusion imprecision to the expected sensor noise level rather than minimising it, with the optimum somewhere around the error rate where evidence is informative but not trusted blindly.","The stability analysis rests on a well-mixed population; on networks with local structure, initial bias might be reinforced by spatial clustering, so the benefit of imprecision may depend on network topology, an extension the paper leaves for future work.","If the mechanism generalises to n > 2 propositions, imprecision acts as a way to keep the true world inside every agent's belief set until evidence can confirm it; the exponential state space makes the difference-equation route impractical, but agent-based sampling could test it."],"forward_implications":["Choosing gamma in [1/3, 1/2) instead of gamma = 0 nearly doubles the area of the (sigma, rho) plane in which the correct consensus is the only stable outcome at moderate noise (about 50% vs 30% at error rate 0.3).","A population with 70-97% of agents initially committed to a wrong belief can reach the truth under imprecise fusion in conditions where precise fusion converges to the wrong consensus.","The benefit grows with population size (400 agents outperform 100-200) and peaks at an evidence error rate around 0.28, vanishing near epsilon = 0 and declining sharply above about epsilon = 0.4.","The improvement comes with a cost: imprecise fusion slows the time to consensus and slightly reduces accuracy when the population is already correct, and gamma >= 1/2 prevents consensus altogether.","When evidence is plentiful relative to communication (high rho, low sigma), learning is good for all fusion settings; imprecision matters most when interactions dominate evidence."],"supporting_citations":[{"why":"Supplies the parameterised imprecise fusion operator that the paper's model is built on.","marker":"[30]"},{"why":"Defines the intersection-union operator used as the precise fusion baseline gamma = 0 and the comparison point for the claimed improvement.","marker":"[20]"},{"why":"Provides the evidential-updating scheme B -> B ∩ E and its use in best-of-n social learning, adopted in Section 2.3.","marker":"[39]"},{"why":"Shows that imprecise (three-valued) beliefs are more robust to noise than weighted voting, motivating the claim that imprecision aids accuracy.","marker":"[14]"},{"why":"Establishes the preceding result that fusion noise between robots can improve group accuracy, which the paper's stability analysis generalises to structured imprecision.","marker":"[27]"}],"fun_headline_variants":["Imprecise belief fusion rescues misinformed groups","Why imprecise merging beats precise in biased crowds","Moderate fuzziness in belief merging aids learning accuracy","Initial bias? Try imprecision when combining beliefs","Imprecision in fusion improves collective truth-finding"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The central benefit relies on an evidential rule in which agents ignore evidence that contradicts their current belief and stop gathering evidence once precise, so a wrong precise consensus is absorbing; if contradictory evidence instead forced a belief change, precise fusion would already escape false consensus and the advantage of imprecision would shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Imprecise belief fusion rescues misinformed groups","Why imprecise merging beats precise in biased crowds","Moderate fuzziness in belief merging aids learning accuracy","Initial bias? Try imprecision when combining beliefs","Imprecision in fusion improves collective truth-finding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000157,"raw_usage":{"total_tokens":1038,"prompt_tokens":706,"completion_tokens":332,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":255}},"tokens_in":450,"tokens_out":332,"duration_ms":4157,"temperature":1.0,"reasoning_tokens":255,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:18:42.115360+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the Case-1 difference-equation and agent-based simulations (97% initially at the wrong state, e.g. sigma = rho = 0.5, epsilon = 0.4) with a modified update rule: when evidence E contradicts belief B, let the agent adopt E (or some combination) instead of ignoring it. If the accuracy gap between gamma = 1/3 and gamma = 0 closes or reverses, the reported benefit is a consequence of the ignore-contradictory-evidence asymmetry rather than of fusion imprecision itself. A second check: repeat the agent-based runs on a small-world or lattice network; if the R1 advantage disappears under local i","supporting_citations":[],"review_version":1}