{"id":"83e4e9ca-610c-41b8-8d6e-c87ac5585d81","arxiv_id":"2608.03566","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Pairwise sharding plus consensus MCMC lets a single neural posterior estimator fit drift diffusion models across designs with accuracy close to full MCMC.","lead":"This paper splits decision-making data into pairs of experimental conditions, runs one neural network on each pair, and combines the resulting posteriors to fit the drift diffusion model. The method aims to make fast amortized Bayesian inference work across study designs without requiring a newly trained network for every design.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The exactness claim in §4 covers only the factorization step; Eq. 2's diagonal-Gaussian consensus is approximate, and the real-data a_accuracy gap (1.65 vs 2.09) shows this approximation is load-bearing, not negligible.","rationale":"The reader's weakest assumption identifies the consensus MCMC step as the load-bearing approximation, and the manuscript's own real-data results provide direct evidence that this approximation can fail badly when shards are small. I agree with that assessment. The likelihood factorization in §4 is exact, but the paper's strongest claim extends exactness to the full pipeline, including the diagonal-Gaussian consensus and importance-sampling calibration. That extension is not supported by the theory and is contradicted by the a_accuracy discrepancy in the real-data example. The proposed simulation test targets exactly this gap: it holds the design and shard size fixed and asks whether the pipeline reproduces the full posterior when the data-generating parameters are known. This would distinguish a pipeline approximation error from model misspecification in the real data. The reader's CONDITIONAL verdict remains appropriate because the simulated evidence is credible and the central method is useful, but the advertised exactness and design-universality need qualification.","tokens_in":25743,"tokens_out":4774,"duration_ms":57839,"concrete_test":"Simulate datasets with known parameters from the rr98 design (same 33×2 imbalanced condition structure, 10-trial mini-blocks, same Q), with true a_accuracy=2.09; run the full pairwise ABI pipeline and compare the calibrated consensus posterior of a_accuracy with the true full posterior. If the bias approaches the real-data gap (~0.44), the small-shard/diagonal-Gaussian consensus is the load-bearing source of error and the 'mathematically equivalent' claim fails in this regime. If bias is small, the real-data gap is due to model misspecification or MCMC discrepancy, and the exactness claim is not the primary issue.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in §4 that the pairwise decomposition is 'mathematically equivalent to analyzing the full dataset at once' is only true for the product of exact shard posteriors ∏ p(θ|S_q). The implemented consensus MCMC (Eq. 2) replaces each shard posterior by a Gaussian and then uses a diagonal precision-weighted average of draws, explicitly forgoing the full covariance matrix. This step is not exact and is not covered by the factorization argument. The paper's own real-data analysis (Table 7 vs Table A3) shows the consequence: for participant 1, a_accuracy is 1.653 (ABI) and 1.620 (MCMC) under consensus, versus 2.092 for full MCMC. The authors attribute this to non-Gaussian 10-trial shard posteriors, but that is precisely the regime where Eq. 2's Gaussian/diagonal approximation is unjustified. Additionally, the importance-sampling correction uses weights p(θ)^{1-Q}; in the real-data example Q is on the order of hundreds or more, so these weights can be extremely variable and may not be reliably approximated by finite posterior draws. Thus the 'exact' design-universality guarantee is overstated for the actual pipeline; the method is a useful approximation whose error grows as shards shrink.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a divide-and-conquer framework for amortized Bayesian inference (ABI) of the drift diffusion model (DDM). The full dataset is partitioned into pairwise condition shards, a single neural posterior estimator (NPE) trained on two-condition data is applied to each shard, and the shard posteriors are combined via consensus MCMC with an importance-sampling correction for prior over-counting. The authors claim that, because the DDM likelihood factorizes exactly over disjoint shards, the procedure is mathematically equivalent to full-data inference. They evaluate the method in two simulation studies (four-condition and 3×2 designs, n=100/200/500, 100 replicates each) with both ABI and MCMC, reporting accuracy (MAE, RMSE) and posterior SDs comparable to full-model inference, and orders-of-magnitude speedups for ABI. They also apply the method to the Ratcliff & Rouder (1998) dataset with 33×2 conditions, where they observe larger discrepancies from full MCMC, which they attribute to small (10-trial) mini-block shards producing non-Gaussian posteriors.","tokens_in":26125,"tokens_out":2844,"duration_ms":36217,"significance":"If the central claim holds, the paper offers a practical solution to a well-known limitation of amortized inference: design-specific NPEs do not generalize across experimental designs. The pairwise decomposition is an appealing idea that could extend ABI to arbitrary numbers of conditions without retraining, and the computational advantage over MCMC is substantial. The paper is also commendable for benchmarking against full-model MCMC, for including two distinct designs and multiple sample sizes, and for explicitly discussing limitations in the real-data example. However, the paper's strongest theoretical claim—that the implemented pipeline is 'mathematically equivalent to analyzing the full dataset at once'—is not supported for the actual consensus step. The discrepancy in the real-data analysis (Table 7 vs. Table A3) shows that the approximation is consequential in small-shard regimes, which is precisely the setting the method is advertised to handle. The contribution is valuable as an approximate and scalable inference strategy, but the exactness framing needs to be corrected and the approximation error characterized more carefully.","major_comments":[{"comment":"The statement in §4 that the decomposition is 'mathematically equivalent to analyzing the full dataset at once' is not supported by the implemented pipeline. The exact factorization applies to the product of exact shard posteriors p(θ|S_q), but Eq. (2) replaces each shard posterior with a Gaussian and then uses a diagonal precision-weighted average, explicitly forgoing the full covariance matrix. The consensus step is therefore approximate, not exact. The paper's own real-data analysis illustrates the consequence: for participant 1, a_accuracy is 1.653 (ABI) and 1.620 (MCMC) under consensus versus 2.092 for full MCMC (Table 7 vs. Table A3), and the authors attribute this to non-Gaussian 10-trial shard posteriors. This is the regime where Eq. (2)'s Gaussian/diagonal approximation is unjustified. Please revise the exactness claim to state precisely what is exact (the likelihood factorizati","section":"§4, Eq. (2)"},{"comment":"The real-data analysis uses 10-trial mini-blocks, chosen as the minimum trial count in rr98, but no sensitivity analysis is reported for this free parameter. The number of shards is very large (1056 pairs for participant 1), and the importance-sampling correction of Appendix B uses weights w(θ) ∝ p(θ)^{1-Q}. With Q on the order of hundreds, these weights can be extremely variable, and finite posterior draws (1000 per pair) may not reliably approximate the reweighted target. The paper does not report any diagnostic for the importance weights (e.g., effective sample size) or the stability of the calibrated consensus posterior. Please add such diagnostics and, at minimum, a small robustness check varying the mini-block size (e.g., 10 vs. 20 vs. 50 trials) to show how the gap relative to full MCMC depends on Q and shard size.","section":"§6, mini-block size"},{"comment":"The claim that 'a single pairwise NPE trained on pairwise data can be applied to any shard regardless of the original design' is subject to a range restriction that is acknowledged implicitly but not as a limitation. In the simulations, the pairwise NPE is trained on shard sizes U(30,300) trials per condition, yet the real data contains conditions with only 10 trials, so the authors retrain a separate NPE with training sizes 10–250. This means the design-universality claim only holds for shard sizes within the training range; the same network cannot be applied to arbitrarily small shards. Please state this explicitly and clarify that the real-data analysis uses a new pairwise NPE, not the one from the simulation studies.","section":"§5, Estimation; §6"},{"comment":"The abstract's claim that the approach 'achieves accuracy and uncertainty comparable to MCMC' is supported by the simulation studies but is not supported by the real-data example, where the consensus estimates of a_accuracy differ from full MCMC by about 20% (1.653 vs. 2.092 for participant 1) and credible intervals are visibly wider and shifted at the tails of the brightness-strength range (Figure 14 vs. Figure A14). The paper's own text acknowledges this discrepancy. The conclusion should be moderated to distinguish the well-behaved large-shard regime from the small-shard regime, where the consensus approximation introduces nontrivial bias and overdispersion. As it stands, the abstract overstates the general empirical performance.","section":"Table 3; Table 7"}],"minor_comments":[{"comment":"Typographical issues: 'T er' appears in several places where 'Ter' is meant; '0 anda' should be '0 and a'; the diffusion constant notation σ is used but the parameter is not listed in the prior table. Please proofread.","section":"Section 2"},{"comment":"The caption says 'The bottom row reports the consensus estimate obtained from the relevant shards,' but the bottom row contains parameter name labels, not estimates. Clarify that the bottom row is the consensus parameter set, not numerical estimates.","section":"Table 2"},{"comment":"The derivation of the importance weights is correct but terse. It would help to state explicitly that the draws from p(θ|x) in the first line are the consensus posterior draws, not the calibrated draws, and to define C and C* consistently. Also, 'θ⋆ l' in the text appears misindexed.","section":"Appendix B"},{"comment":"The sentence 'Obtaining 3,000 posterior draws from the full-model NPE required approximately 195 milliseconds for 100 simulated datasets' is ambiguous: 195 ms per dataset or total? From the context it seems per dataset, but please make it explicit, because the subsequent pairwise time of 730 ms is also per 100 datasets.","section":"Section 5, Computational time"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid applied contribution with an honest limitations section, but the 'mathematically equivalent' framing in §4 is likely to mislead readers into thinking the consensus step is exact. The real-data gap is not a minor artifact; it is the expected consequence of applying a Gaussian/diagonal consensus to small shards. If the authors correct the exactness claim, add diagnostics for the importance-sampling step, and temper the abstract, the paper would be a valuable contribution to the ABI/DDM literature. I do not see a need for rejection, as the simulations are extensive and the method is useful as an approximate, fast alternative."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper is worth reading: the authors show that by splitting any DDM dataset into pairwise shards, a single neural posterior estimator can be applied regardless of the original design, with the shard posteriors combined by consensus MCMC. The idea is not deep—likelihood factorization plus consensus MCMC plus NPE—but the demonstration is solid. Two simulation settings (four-condition and 3x2 design), three sample sizes, 100 replicates, and both ABI and MCMC variants show the calibrated consensus posterior tracks the full posterior well, and the computational gain is real: ABI inference stays near-instantaneous while MCMC grows. The paper also spends attention on NPE calibration and honestly reports where the method struggles. That is real evidence, not just a proof of concept.\n\nThe soft spot is the paper's own exactness claim. Section 4 argues that because the likelihood factorizes exactly, the decomposition is \"mathematically equivalent\" to analyzing the full dataset. That is only true for the product of exact shard posteriors. The actual combination uses Eq. 2, a precision-weighted average of draws with a diagonal covariance approximation—explicitly an approximation, and one acknowledged to be suboptimal. The real-data example shows the consequences: for participant 1, boundary separation under accuracy instructions is 1.653 (ABI) and 1.620 (MCMC) in the consensus pipeline versus 2.092 from full MCMC. The authors attribute this to non-Gaussian 10-trial shard posteriors, which is exactly the regime where the Gaussian/diagonal consensus is unjustified. The importance-sampling correction, weights p(θ)^(1-Q) with Q on the order of hundreds, is also likely unstable. So the paper's own evidence undermines the \"mathematically equivalent\" phrasing.\n\nThese issues are fixable. The paper could present the method as a controlled approximation whose error grows as shards shrink, and support that with a comparison of full-covariance versus diagonal consensus or some diagnostic for Gaussianity. It also needs to release code and data; without them, the extensive simulation results cannot be checked. The central empirical finding—that a design-agnostic pairwise NPE plus consensus MCMC gives posteriors close to full MCMC in reasonable settings—survives my reading.\n\nFor a cognitive modeling or amortized inference audience, this is a genuinely useful methodological contribution. I would send it to peer review, with the expectation that the authors temper the exactness claim, add the missing reproducibility artifacts, and ideally report the real-data discrepancy more prominently rather than as an aside.\n\nRecommendation: engage with it; it deserves referee time.","headline":"Useful divide-and-conquer ABI for DDM, but the 'exact' claim is overstated: the real-data boundary-separation gap shows the consensus step is approximate in practice.","tokens_in":26520,"tokens_out":2445,"would_cite":true,"duration_ms":28738,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62-08","65C05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single pairwise neural posterior estimator, applied to disjoint pairs of conditions and combined by consensus MCMC, can approximate the full drift-diffusion posterior for any experimental design.","keywords":["drift diffusion model","amortized Bayesian inference","neural posterior estimation","consensus MCMC","conditional flow matching","divide-and-conquer","design generalization","Bayesian computation"],"falsifier":"Simulate a dataset whose conditions are tiny (about 10 trials per shard), run the pairwise divide-and-conquer pipeline, and compare the calibrated consensus posterior to the full MCMC posterior; if the boundary-separation posterior mean and credible interval deviate beyond Monte Carlo error, the Gaussian consensus assumption is the failing link.","tokens_in":25682,"feed_emoji":"🧠","tokens_out":7863,"duration_ms":83661,"temperature":0.7,"pith_summary":"Amortized Bayesian inference (ABI) for the drift diffusion model (DDM) is nearly instantaneous, but a trained neural network is locked to the experimental design it saw during training. This paper proposes a divide-and-conquer fix: split any dataset into disjoint pairs of conditions, run one pairwise-trained network on every pair, and combine the pairwise posteriors through consensus MCMC with an importance-sampling correction. Because DDM trials are independent, the product of shard posteriors is mathematically equivalent to the full posterior, so the decomposition is not an approximation in principle. In two simulation settings, the calibrated consensus posterior matched full MCMC in both posterior means and uncertainties, with inference time in milliseconds rather than minutes. The real-data test showed the boundary of the method: when conditions shrink to about 10 trials, consensus estimates of boundary separation depart visibly from full MCMC, a failure the paper traces to non-Gaussian shard posteriors.","feed_headline":"One pairwise network fits any drift-diffusion design","feed_subtitle":"Splitting data into condition pairs lets a single trained network match full MCMC accuracy in milliseconds.","key_machinery":"The central mechanism is the exact factorization of the DDM likelihood over disjoint shards: p(x|θ) = ∏_q p(S_q|θ), which holds because trials are independent given the parameters. The paper combines this with a pairwise neural posterior estimator and consensus MCMC, where each shard's posterior draws are fused by a diagonal precision-weighted average (Eq. 2), and importance weights proportional to p(θ)^{1−Q} undo the prior overcounting introduced by analyzing Q shards separately. The Gaussian-product assumption in the combination step is what makes the whole pipeline tractable.","core_discovery":"The paper establishes that the DDM's independence across trials lets any experimental design be partitioned into pairwise shards that all share one common model structure: two conditions, two drift rates, and shared remaining parameters. A single neural posterior estimator trained only on two-condition data can therefore be applied to shards from any design, regardless of the number of conditions. The full posterior is recovered by multiplying the shard posteriors, correcting the Q-fold reuse of the prior with importance weights, and combining draws through a precision-weighted consensus average. The calibrated consensus posterior approximates the full posterior closely in simulations, demon","pith_inferences":["Beyond the paper, the same sharding trick should transfer to any generative model whose likelihood factorizes over independent units, such as hierarchical or multi-condition cognitive models beyond the DDM, provided a pairwise estimator can be trained for the smallest exchangeable unit.","The small-shard failure seen in the real-data analysis suggests a non-Gaussian combination step—for example, fitting a normalizing flow or a copula to the shard draws—could extend the framework to designs with very few trials per condition.","Because the pairwise network is trained on a broad range of trial counts, it could support adaptive experimental designs where condition sizes change mid-experiment, without retraining."],"forward_implications":["A single pairwise-trained network can serve any number of conditions without retraining, removing the main generalization bottleneck of ABI for DDM studies.","Inference time stays constant in dataset size and several orders of magnitude below MCMC; the paper reports roughly 730 ms for 100 datasets across six shards versus tens of minutes for MCMC.","The sharding strategy lowers the dimensionality of the inference problem, avoiding the biased posteriors the authors encountered when training high-dimensional full-model networks.","For MCMC, the same pairwise decomposition is not a computational win because each shard needs its own chain; the gains are specific to the amortized setting.","Accuracy and uncertainty of the consensus posterior approach those of full MCMC as trials per condition grow, with small residual over-dispersion in boundary separation parameters."],"supporting_citations":[{"why":"Introduces the drift diffusion model, the observation model whose inference this paper accelerates.","marker":"Ratcliff, 1978"},{"why":"Provides the consensus MCMC algorithm used to combine shard posterior draws via a precision-weighted average.","marker":"Scott et al., 2022"},{"why":"Supplies the asymptotically exact embarrassingly parallel MCMC framework behind the Gaussian combination of subposteriors.","marker":"Neiswanger et al., 2013"},{"why":"Provides the importance-sampling machinery used to cancel the prior overcounting introduced by analyzing Q shards.","marker":"Tokdar & Kass, 2010"},{"why":"Supplies the Bernstein-von Mises theorem justifying the Gaussian approximation of each shard posterior.","marker":"Le Cam, 2012"},{"why":"Defines the amortized Bayesian inference workflow used to train the neural posterior estimators.","marker":"Radev et al., 2023"},{"why":"Provides conditional flow matching, the generative network architecture used for the posterior estimator.","marker":"Lipman et al., 2022"},{"why":"Contributes the real-world dataset on which the divide-and-conquer approach is stress-tested.","marker":"Ratcliff & Rouder, 1998"}],"fun_headline_variants":["Pairwise shards make one DDM network fit all designs","One network trained on pairs, any DDM design","DDM inference in milliseconds via pairwise shards","Divide and conquer DDM: pairwise shards, consensus"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The combination step assumes each shard posterior is close to Gaussian, so a diagonal precision-weighted average of posterior draws recovers the full posterior after importance weighting; with very small shards this premise fails, as the paper's 10-trial mini-block analysis shows.","fun_headline_variants_meta":{"raw":{"variants":["Pairwise shards make one DDM network fit all designs","One network trained on pairs, any DDM design","DDM inference in milliseconds via pairwise shards","Divide and conquer DDM: pairwise shards, consensus"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001076,"raw_usage":{"total_tokens":4321,"prompt_tokens":705,"completion_tokens":3616,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":3563}},"tokens_in":449,"tokens_out":3616,"duration_ms":28393,"temperature":1.0,"reasoning_tokens":3563,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:44:09.905267+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a dataset whose conditions are tiny (about 10 trials per shard), run the pairwise divide-and-conquer pipeline, and compare the calibrated consensus posterior to the full MCMC posterior; if the boundary-separation posterior mean and credible interval deviate beyond Monte Carlo error, the Gaussian consensus assumption is the failing link.","supporting_citations":[],"review_version":1}