{"id":"0fcd2dea-b11c-410f-a157-0cb32e1ecc3d","arxiv_id":"2607.08335","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"SCOREBED isolates EIG double intractability in a policy-independent score-matching stage, then trains design policies with a singly intractable gradient estimator, enabling cheap multi-policy selection.","lead":"The paper shows how to train adaptive Bayesian experimental design policies by first learning a policy-independent score network for the intractable marginal likelihood, then using it for cheap policy gradients. This converts a costly nested sampling burden into an additive two-stage procedure, so multiple policies can be trained and selected under a fixed likelihood budget.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged design-support / gradient-bias assumption.","rationale":"The reader's weakest_assumption already names the precise load-bearing condition (q-support + reparameterisation Lipschitz / Jacobian conditioning). The manuscript itself supplies the supporting theory (Theorem 3), the bias-variance decomposition (Appendix C), and the empirical diagnostics (Figures 2–7, Tables 1–2, E.3–E.5). No stronger internal flaw appears: the gradient derivation is standard reparameterisation + chain rule, the MSM identity is classical, and the additive-cost accounting is explicit. Residual score instability on dynamical systems is already mitigated by multi-seed selection and does not overturn the isolation claim when the score is accurate. Therefore the CONDITIONAL verdict with high confidence stands; no adjustment is warranted.","tokens_in":40318,"tokens_out":535,"duration_ms":5816,"concrete_test":"On the cart-pole task, retrain the score network under the mixture q of Appendix D.4.1, then evaluate mean score MSE and EIG-gradient squared bias (Appendix C) on trajectories from the final trained policy versus trajectories from the original q; if policy-induced MSE or bias rises by more than a factor of ~3 relative to the q-support values used for the main Table 2 results, the support-transfer assumption is the binding limitation on that task.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that double intractability can be isolated into a policy-independent MSM stage (Eq. 6) whose learned scores plug into the reparameterised gradient (Theorem 1 / Eq. 3), converting multiplicative NLE cost into additive cost and enabling multi-policy training under a fixed budget. The only place this can fail is exactly the condition the reader already isolates: that s_ψ trained under a hand-chosen q(ξ) remains accurate on trajectories visited by later policies, so that the Lipschitz-controlled gradient error of Theorem 3 stays small enough for stable SGD. Appendices E.3–E.5 already document the failure modes (isotropic Gaussian sampler collapses EIG; dynamical-system Jacobians become ill-conditioned and score MSE correlates with condition number; early score seeds produce unstable policies). These are residual practical risks, not an internal inconsistency or a hidden assumption that would invalidate the isolation argument when the support condition holds. Math, code release, and budgeted experiments support the claim as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes SCOREBED, a two-stage method for policy-based Bayesian experimental design. It first trains a policy-independent score network via marginal score matching (MSM, Eq. 6) to approximate the Stein and Fisher scores of the marginal p(y|ξ). These scores are then substituted into a reparameterised EIG gradient (Theorem 1 / Eq. 3), yielding a singly intractable estimator for policy SGD. The claim is that this isolates double intractability from policy learning, converting the usual multiplicative nested-sampling NLE cost into an additive cost and thereby enabling cheap multi-policy training (architecture search, restarts) under a fixed likelihood budget. Experiments on source location finding, three dynamical systems, and gravimetry compare against PCE, pre-trained and joint variational bounds, and IO-SMC2 under matched NLE budgets, reporting both lower and upper EIG bounds for P=1 and multi-restart protocols.","tokens_in":40581,"tokens_out":1050,"duration_ms":9816,"significance":"If the isolation argument holds in practice, SCOREBED is a useful engineering contribution for policy-based BED: amortising the hard nested terms once, then reusing them for many policy trainings, is a natural and previously under-exploited separation. Theorem 1 cleanly isolates the intractable scores; the static-design simplification and the Lipschitz error bound (Theorem 3 / Appendix B) are sound; the bias–variance analysis (Appendix C) explains the observed gradient plateaus. Code release, fixed-NLE protocols, and dual EIG bounds strengthen the empirical case. The main practical value is the ability to train multiple competitive policies without repeating the expensive nested work—an advantage that is real under the paper’s own budget accounting and that existing nested or co-trained methods do not share.","major_comments":[{"comment":"§3, Appendix D.4.1 and E.3: the central practical claim—that a single policy-independent score network yields usable gradients for later policies—rests on the support of the hand-chosen design sampler q(ξ). Table 14 shows that an isotropic Gaussian sampler collapses EIG to ~2.86 while scale-aware or low-rank samplers recover ~11. The paper documents the failure mode but does not give a systematic way to choose or validate q(ξ) for a new problem; without that, the additive-cost advantage is conditional on a non-trivial modelling choice that is currently problem-specific.","section":"§3 / App. E.3"},{"comment":"§5.2 and Appendix E.5 (Figs. 5–7): on the dynamical systems the score-network stage is unstable (multiple seeds required; Jacobian condition numbers of trained policies correlate with score MSE). The multi-seed P=3/50 protocol mitigates this, but it means part of the “policy” budget is still spent on score restarts. The paper should quantify how much of the claimed additive saving survives when score training itself needs restarts, and whether the Lipschitz constant L in Theorem 3 becomes large enough on non-Markovian trajectories to make the error bound uninformative.","section":"§5.2 / App. E.5"}],"minor_comments":[{"comment":"Clarify early that the “marginal” p(y|ξ) is not the policy-induced data distribution p(y;π_ϕ); the distinction is made in §3 but is easy to miss and is load-bearing for the policy-independence claim.","section":"§3"},{"comment":"Table 1 / Fig. 1: for d=3,K=10, SCOREBED (P=1) slightly outperforms P=5; a short discussion of when extra restarts help versus when the score budget is still the bottleneck would help readers allocate NLE.","section":"Table 1"},{"comment":"Appendix A.2: the reparameterisation map g is deferred; a short explicit form for the common additive-noise case (already treated in A.3) would make the gradient expression easier to implement from the main text alone.","section":"App. A.2"},{"comment":"Related work: concurrent Huang et al. (2026a) is noted as complementary; a one-sentence comparison of what is amortised (belief representations vs. scores) would sharpen the positioning.","section":"§4"}],"recommendation":"minor_revision","confidential_remarks":"The isolation idea is clean and the math is careful; the residual risk is purely practical (design support / Jacobian conditioning). I would not block on that if the authors add a short “how to choose q(ξ)” checklist and a clearer accounting of multi-seed score cost. Fit for UAI/stat.ML is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful move here is simple: the information gain depends on the designs, not on how the policy chose them. They write the reparameterised EIG gradient (Theorem 1) so the only nested terms are the Stein and Fisher scores of the marginal p(y|ξ), train one score network on a policy-free MSM regression against the known conditional score, then plug it into a singly intractable policy gradient. Multiplicative nested cost becomes additive. That is new relative to PCE nesting and to co-trained variational bounds, and it is the right place to put the amortisation.\n\nMath is solid. The static-design simplification is clean; the Lipschitz bound (Theorem 3) correctly ties score MSE to gradient error; Appendix C’s bias-variance split explains the plateaus they plot. Code is released. Experiments fix total likelihood evaluations, report both PCE lower and NMC upper bounds, and show that under the same budget they can train several competitive policies while baselines must thin their per-policy effort. On the high-dimensional location task SCOREBED is clearly best; on the dynamical systems it is competitive once they select among score seeds.\n\nSoft spots are practical, not foundational. The score network is only as good as the design sampler q(ξ). Isotropic Gaussians collapse performance; they need scale hyperpriors or low-rank trajectories. On the non-Markovian dynamical systems the reparameterisation Jacobian becomes ill-conditioned under the trained policy, score MSE correlates with condition number, and early seeds produce unstable policies—hence the three-seed split. Absolute gains on the easier tasks are modest. None of this breaks the isolation argument when support is adequate; it just means the free parameters (sampler, architecture, budget split) still matter.\n\nThis is for people who actually train adaptive BED policies and care about likelihood budgets. It deserves a serious referee. I would engage with it and expect to cite the gradient form and the two-stage cost argument.","headline":"Clean isolation of EIG double intractability into policy-independent score matching; additive NLE cost is real and the math holds.","tokens_in":41187,"tokens_out":486,"would_cite":true,"duration_ms":7263,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Score matching isolates EIG double intractability so policy training becomes singly intractable and reusable.","keywords":["Bayesian experimental design","expected information gain","score matching","policy networks","amortised inference","reparameterisation","marginal likelihood"],"falsifier":"Train the score network once, then train many policies under a matched likelihood budget; if the best-of-P ScoreBED policies systematically underperform best-of-P nested or co-trained baselines on the same tasks, or if measured EIG-gradient bias fails to fall with score error, the claim fails.","tokens_in":41193,"feed_emoji":"📊","tokens_out":784,"duration_ms":14412,"temperature":0.7,"pith_summary":"Policy-based Bayesian experimental design trains networks that choose the next experiment from past data, but the expected information gain is doubly intractable, so every gradient step usually needs nested sampling or a co-trained variational model. That multiplies cost by the number of policies one wants to try and leaves little budget for restarts or architecture search. This paper shows the information gain itself depends only on the realised designs and data, not on the policy parameters. The double intractability therefore lives in two policy-independent scores of the marginal likelihood. Those scores can be learned once by ordinary regression (marginal score matching) and then plugged into a reparameterised gradient that is only singly intractable. The expensive step becomes additive rather than multiplicative, so many competitive policies can be trained under a fixed likelihood budget and the best one selected.","feed_headline":"Score matching turns EIG double cost into additive cost","feed_subtitle":"Learn the intractable scores once, then train many policies cheaply under a fixed likelihood budget","key_machinery":"Theorem 1 (reparameterised EIG gradient) isolates the intractable terms as the Stein score ∂/∂y log p(y|ξ) and Fisher score ∂/∂ξ log p(y|ξ); marginal score matching then regresses a shared network onto the conditional likelihood scores, recovering both targets without sampling the posterior.","core_discovery":"The double intractability of the expected information gain can be separated from policy optimisation: first solve a single policy-independent score-matching problem for the Stein and Fisher scores of the marginal likelihood, then substitute the learned scores into a reparameterised EIG gradient that is only singly intractable. This converts the usual multiplicative likelihood cost into an additive cost and makes multi-policy training cheap.","pith_inferences":["Because the score targets live in design and observation space rather than parameter space, the method should scale more gracefully than nested posterior sampling when the latent dimension grows.","If the design sampler q(ξ) can be adapted online to the current policy support without destroying amortisation, the coverage assumption becomes far weaker.","The same score network could be reused across related models that share the same observation and design spaces, amortising design cost over a family of experiments."],"forward_implications":["Under a fixed likelihood budget one can train many policy restarts or architectures and simply keep the best, without re-paying the double-intractability cost.","Gradient bias, not variance, becomes the dominant error once outer-sample size is moderate; score-network capacity and training budget therefore control final policy quality.","The same two-stage pattern applies to any variational approximation of the marginal or posterior that is independent of the policy.","Static-design special cases drop the Fisher-score term, lowering bias and variance even without score matching."],"fun_headline_variants":["Score matching isolates EIG double intractability for cheap multi-policy BED","Separate score matching from policy training cuts EIG cost multiplicatively","Learn scores once then train BED policies under additive likelihood cost","Policy-independent score matching converts EIG double cost to additive","Score-matched EIG gradients enable cheap multi-policy Bayesian design"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"A score network trained only on samples from a hand-chosen design distribution must stay accurate on the trajectories later visited by trained policies, otherwise gradient bias ruins policy optimisation.","fun_headline_variants_meta":{"raw":{"variants":["Score matching isolates EIG double intractability for cheap multi-policy BED","Separate score matching from policy training cuts EIG cost multiplicatively","Learn scores once then train BED policies under additive likelihood cost","Policy-independent score matching converts EIG double cost to additive","Score-matched EIG gradients enable cheap multi-policy Bayesian design"]},"model":"grok-4.5","effort":"low","cost_usd":0.00396,"raw_usage":{"total_tokens":1184,"prompt_tokens":738,"num_sources_used":0,"completion_tokens":72,"cost_in_usd_ticks":39600000,"prompt_tokens_details":{"text_tokens":738,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":374,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":738,"tokens_out":72,"duration_ms":3540,"temperature":1.0,"reasoning_tokens":374,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T09:10:52.835992+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the score network once, then train many policies under a matched likelihood budget; if the best-of-P ScoreBED policies systematically underperform best-of-P nested or co-trained baselines on the same tasks, or if measured EIG-gradient bias fails to fall with score error, the claim fails.","supporting_citations":[],"review_version":1}