acceptodds
Under review as a conference paper at ICLR 2027

From Pressure to Priority: Correctable Particle Transport for Dynamic Partitions

Abstract

Distributed systems expose two views of contention that age at different rates: an occasional membership snapshot records which workloads shared a resource, while fresher local telemetry reports only how crowded each workload's own group appears. The latter behaves like a noisy observation of latent block cardinality, not like an identity-bearing edge measurement. A bootstrap partition filter ignores that distinction and spends its particles on split-merge transitions that fresh telemetry has already contradicted; replacing the belief by the most plausible single partition discards the ambiguity that same-cardinality groups necessarily carry. We introduce Pressure-Residual Partition Transport (\prpt), a full-support proposal for sequential inference over dynamic set partitions. factorizes a pressure-conditioned move into an event, a target cardinality, and a product-weighted fixed-cardinality subset. An elementary-symmetric-polynomial recursion samples splits and evaluates their proposal probabilities without enumerating exponentially many subsets, so exact importance correction remains available. The resulting kernel is permutation equivariant and costs for blocks. With 32 particles under contaminated observations, lowers posterior total variation from to on an exactly enumerable benchmark, and removing correction raises it to . In delayed, lossy multi-step experiments, it improves a prespecified decision objective in all six test conditions. On 600 held-out Alibaba trace-derived episodes with mostly exact pressure, it reduces total variation by and raises the effective-sample-size ratio by relative to an equal-budget bootstrap filter; a separate 397-episode non-exact slice retains the ESS gain but exposes day-dependent posterior accuracy. These results delimit when cardinality telemetry improves finite-particle inference without pretending that it identifies membership.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Sequential Monte Carlo, Dynamic Partitions, Structured Proposals, Combinatorial Inference, Permutation Equivariance, Partial Identification

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_jEBhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies inference over changing partitions when membership snapshots are delayed or missing but fresh measurements provide noisy estimates of each node’s group size. It proposes Pressure-Residual Partition Transport (PRPT), which uses discrepancies between observed and predicted group sizes to guide the choice of a stay, merge, or split event, followed by a target size and membership subset. An elementary-symmetric-polynomial recursion samples fixed-size subsets without enumerating all possible splits. The proposal is mixed with the transition prior and corrected through importance weights. The main contribution is this structured proposal for cardinality observations, together with its permutation-equivariance and polynomial-cost construction.

The evaluation includes exact small-state posterior comparisons, correction ablations, multi-step decision tasks, random-topology tests, five held-out days of Alibaba trace-derived data, and runtime measurements. With 32 particles under contaminated synthetic observations, PRPT reduces posterior total variation from 0.2062 to 0.0645 relative to bootstrap SMC. Decision results and mostly-exact trace observations also show improvements, while the non-exact trace results reveal that posterior accuracy gains depend on the observation regime.

Critical strengths

The following are the critical strengths in this paper:

  1. A tractable proposal tailored to the observation structure. PRPT connects cardinality residuals to event selection, target size, and membership sampling. The fixed-size subset sampler avoids enumerating exponentially many splits while retaining an evaluable proposal probability for importance correction. This is a useful combination of established tools for the stated problem: it makes observation-guided sampling computationally feasible without reducing uncertain membership to a single estimate. For a block of size 16, the reported median proposal time is 0.211 ms, compared with 127.1 ms for explicit split-support construction.

  2. Controlled experiments provide meaningful evidence of finite-particle gains. The exact small-state benchmark measures posterior error directly, rather than relying on ESS alone. Across 200 contaminated-pressure cases with 32 particles, PRPT reduces TV from 0.2062 for bootstrap SMC to 0.0645, close to the exact-posterior proposal’s 0.0619. Removing correction increases TV to 0.3778 despite retaining 99% true-state coverage, illustrating why guided sampling alone is insufficient. The six delayed-communication conditions also show an average decision-objective improvement of 6.36 over bootstrap. These results connect the proposal design to both posterior quality and downstream decisions, which is consequential to the paper’s practical value.

Critical weaknesses

The following are the critical weaknesses that I would like to mention:

  1. The treatment of delayed snapshots is not sufficiently explained to establish filtering correctness. Section 2 states that an arriving snapshot reveals an earlier partition, while Algorithm 1 resets the particles to that snapshot before applying a single transition. The algorithm does not describe how the snapshot’s timestamp, intervening transitions, or intervening pressure observations are handled. Proposition 1 establishes support and importance correction for a local update, but does not resolve this temporal issue. Since delayed communication is central to the problem, the paper needs a precise account of how this operation targets the current filtering posterior. As written, this remains a substantive correctness concern rather than an established implementation error.

  2. The proposal specification leaves important reproducibility and verification gaps. Section 3.1 describes event and size logits through residual costs without giving their complete expressions, and describes node log-odds without an explicit formula. Section 4 reports “exploration 0.2” but does not define how it enters the proposal. Algorithm 1 also leaves the ESS threshold unspecified. More consequentially, Equation (5) defines a mixture of complete proposals, whereas the Figure 1 caption describes mixing at every level; these constructions generally differ. The sampling procedure and probability used for correction therefore need to be stated unambiguously. Appendix B does not provide the detailed specification promised in the reproducibility statement.

  3. The experiments do not sufficiently isolate the contribution of the proposed design. The correction ablation establishes the importance of proper weighting, but does not distinguish PRPT’s hierarchy from other correctly weighted informed proposals. In the random-topology tests, Huber-PRPT and equally tuned L2 guidance tie in four conditions; L2 performs better by 0.71 objective units in the low-churn n=12n=12 condition and by 5.82 in the hardest n=12n=12 condition (§5.2, p. 7). These findings leave the additional benefit of the robust residual choice unresolved. Likewise, Table 5’s timing comparison gives 0.211 ms versus 127.1 ms at n=16n=16, which is against explicit support construction (p. 9). A comparison with alternative tractable informed proposals and ablations of the event, size, and subset guidance would better establish the method’s added value.

  4. The real-data and scaling results leave important limits on the empirical case. In Table 3’s non-exact stratum, TV reduction is 0.0345 with interval [−0.0238,0.0761][-0.0238,0.0761], utility gain is 0.1134 with [−0.0300,0.2642][-0.0300,0.2642], and coverage gain is −0.0403 (p. 8). Thus, posterior and utility improvements remain uncertain when pressure is an imperfect cardinality proxy, although ESS improves. Section 4 and Appendix B also leave the request-to-pressure mapping and episode-selection criteria insufficiently detailed to assess the retained cases. Finally, Table 5 reports proposal timing at n=32n=32 and n=64n=64, while the independent-domain experiment supplies the decomposition to all methods (§5.2). These evaluations do not establish posterior accuracy on larger, coupled partitions.

Questions for authors

The following are my questions for authors:

  1. Clarification of the delayed-snapshot update. When a snapshot from time t−dt-d arrives, does the implementation reset particles to that partition and advance once, as Algorithm 1 appears to describe, or does it replay the intervening transitions and observations? Please specify the posterior targeted by this update and how timestamps are handled. An exact small-state comparison with multiple transitions during the delay would help resolve whether the procedure preserves the intended filtering distribution.

  2. Clarification of the actual proposal probability. Is the prior/guided mixture sampled once for the complete transition, as in Equation (5), or independently at each level, as the Figure 1 caption suggests? Please provide the corresponding probability used in Equation (6), together with the explicit event scores, target-size scores, node log-odds, and the role of “exploration 0.2.” These details are needed to verify that sampling and importance correction use the same distribution.

  3. Evidence for the hierarchy’s specific contribution. What evidence separates the benefit of event–size–subset guidance from the general benefit of using pressure-informed, correctly weighted proposals? Could you provide a focused ablation of these three guidance levels, or a comparison with another tractable informed proposal under matched tuning and runtime? The reported Huber/L2 comparison leaves the residual choice’s advantage unresolved, while the timing comparison against explicit enumeration does not establish an advantage over other non-enumerative constructions. This is a request to substantiate the current contribution.

  4. Clarification of the trace-derived observation construction. How exactly are the available request fields converted into node-wise cardinality pressure, and what rules select the mostly-exact and non-exact episodes? Please specify whether current server membership is used anywhere in feature construction, what mapping is frozen after calibration, and whether all compared variants use identical episodes and target likelihoods. Given the non-exact results in Table 3, these details would materially affect my confidence that the observed gains reflect inference from available telemetry and clarify the scope of the real-data evidence.

These requests concern the validity and interpretation of the current results. Clear answers, with supporting evidence where available, could move my assessment toward weak acceptance.

1 comment
AuthorsOpenReview ↗

Response for correctness, specification, and trace interpretation

Thank you for the concrete questions. New exact-posterior, ablation, and trace tests support PRPT; we specify the implemented law for verification.

W1/Q1 (delayed snapshots). On the actual delayed/lossy channel, submitted PRPT lowers TV to the exact assumed-model current-time posterior from 0.1465 (prior-only) to 0.0830 over 6,000 four-node steps; after two or more changes, it lowers TV from 0.3981 to 0.2528 over 992 steps. A delivered partition value different from the last delivered value resets particles to that partition, followed by one pressure-guided transition; a newer snapshot with the same value does not reset. Its one-step weight targets this transition under current pressure; the exact current-time target also uses source time and later observations. The submitted update does not replay intervening events; timestamp replay reaches 0.0433 TV overall. The submitted approximation thus improves inference under delay, although replay is more accurate. We will clarify these semantics in Algorithm 1.

W2/Q2 (implemented proposal). Both proposal densities normalize within 2.22×10−162.22\times10^{-16} (80 exhaustive cases); 20,000 draws match qinnerq_{\rm inner} at TV 0.00466. The implemented law follows. Pressure xix_i estimates block size minus one (yi=xi+1y_i=x_i+1 in the paper). Let hδ(r)=r2/2h_\delta(r)=r^2/2 for ∣r∣≤δ|r|\leq\delta and δ(∣r∣−δ/2)\delta(|r|-\delta/2) otherwise; set C(B,s)=∑i∈Bhδ(xi−s+1)C(B,s)=\sum_{i\in B}h_\delta(x_i-s+1) and DB(k)=∑i∈Bmin⁡{hδ(xi−k+1),hδ(xi−∣B∣+k+1)}D_B(k)=\sum_{i\in B}\min\{h_\delta(x_i-k+1),h_\delta(x_i-|B|+k+1)\}. Event scores are Lstay=∑BC(B,∣B∣)L_{\rm stay}=\sum_B C(B,|B|), Lmerge(A,B)=Lstay−C(A,∣A∣)−C(B,∣B∣)+C(A,∣A∣+∣B∣)+C(B,∣A∣+∣B∣)L_{\rm merge}(A,B)=L_{\rm stay}-C(A,|A|)-C(B,|B|)+C(A,|A|+|B|)+C(B,|A|+|B|), and Lsplit(B)=min⁡kDB(k)L_{\rm split}(B)=\min_k D_B(k); gE(e)∝e−Le/Tg_E(e)\propto e^{-L_e/T} and gT(k)∝e−DB(k)/Tg_T(k)\propto e^{-D_B(k)/T}, with uniform split-size prior.

For split (m,k)(m,k), the anchor is the smallest persistent key (dynamic) or node ID (ASI). Other nodes have ℓi=clip⁡((∣xi+1−(m−k)∣−∣xi+1−k∣)/T,−20,20)\ell_i=\operatorname{clip}((|x_i+1-(m-k)|-|x_i+1-k|)/T,-20,20); a guided size-(k−1)(k-1) subset SS has mass ∏i∈Seℓi/ESP⁡k−1(eℓ)\prod_{i\in S}e^{\ell_i}/\operatorname{ESP}_{k-1}(e^\ell), versus uniform prior over (m−1k−1)\binom{m-1}{k-1} subsets.

Each level mixes qj=0.8gj+0.2pjq_j=0.8g_j+0.2p_j, giving qinner=qEqTqSq_{\rm inner}=q_Eq_Tq_S and one-step weight pL/qinnerpL/q_{\rm inner}.

Dynamic settings are (T,δ)=(0.5,0.5)(T,\delta)=(0.5,0.5), pE(stay)=0.75p_E(\mathrm{stay})=0.75 with the rest uniform over valid changes (sole stay has mass one), and systematic resampling after ESS calculation at every propagation.

ASI uses (T,δ)=(1,100)(T,\delta)=(1,100), uniform event prior, 16 prior plus 16 guided draws, qASI=0.5p+0.5qinnerq_{\rm ASI}=0.5p+0.5q_{\rm inner}, weight pL/qASIpL/q_{\rm ASI}, and no resampling. The paper's Eq. (5), Algorithm 1, and Section 4 default 50-50 sentence differ from this implemented law; we will correct them and Appendix B in an allowed revision.

W3/Q3 (hierarchy-specific value). At matched policy CPU, PRPT (32 particles) exceeds tuned APF (80) by 20.88 [17.35, 24.46] objective units on 64 coupled episodes; APF itself beats bootstrap. In 3,000 matched factorized cases, removing event or subset guidance raises TV by 0.0121 [0.0025,0.0209] or 0.0126 [0.0047,0.0201]; under a joint nonfactorized likelihood, the corresponding pooled increases across 1,600 cases are 0.0345 and 0.0116. Thus the gain over generic guidance is traceable to event and subset conditioning. The size level gives a normalized fixed-size conditional, although its chosen score has no isolated TV gain. All six residual-family/exploration settings on fixed Day 170 improve mean TV and ESS over bootstrap, so the gain there is not specific to Huber.

W4/Q4 (trace and coupled scale). Across all 5,924 eligible one-arrival cases on five held-out days and two particle streams, PRPT gains 0.0421 [0.0382,0.0460] TV and 0.0258 [0.0210,0.0306] truth coverage over bootstrap; all five days have positive TV point estimates. On 774 naturally non-exact cases in a separate three-stream test, TV still gains 0.0402 [0.0133,0.0633], while coverage falls 0.0454 [0.0137,0.0818]. In coupled n=8n=8 delayed/lossy tests, submitted PRPT's exact current-time TV is 0.0659 versus 0.2452 for prior-only; under higher ambiguity, 0.3611 versus 0.4268.

For pod ii on current server jj, let rhr_h denote pod hh's recorded gpu_request. The adapter computes

xi=max⁡(0,(∑h∈jrh−ri)/cspec),x_i=\max(0,(\sum_{h\in j}r_h-r_i)/c_{\rm spec}),

where cspecc_{\rm spec} is the Day-30 median positive same-spec request. Current membership constructs the proxy upstream; the filter sees xx and stale membership.

Cases have one arrival, no removal, 2-7 old same-spec pods, two stable same-spec distractor servers, and at most 16 nodes; fixed hash order selects them. Day 60 fixes settings; Days 170/180 provide 300 frozen cases each. A separate 397-case non-exact, globally new-arrival challenge uses Days 160/175/184. Methods share episodes and likelihoods. The trace inference is conditional on this proxy: 86.9% of cases have exact pressure; all 600 main cases have a unique size-signature candidate; online availability is unverified.

Reviewer_hZ2VOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Pressure-Residual Partition Transport (PRPT), a sequential Monte Carlo framework for dynamic partition inference using delayed membership snapshots and noisy cardinality observations. The method constructs a hierarchical proposal consisting of event selection, target block size, and fixed-cardinality subset sampling. An elementary symmetric polynomial recursion enables efficient subset sampling without exponential enumeration while maintaining exact importance correction. Experiments on synthetic benchmarks and Alibaba trace-derived data demonstrate improvements in posterior approximation and downstream decision quality compared with bootstrap particle filtering, although performance is less consistent under imperfect cardinality observations.

Critical strengths
  1. The proposed hierarchical proposal effectively combines pressure-based guidance with exact importance correction, maintaining full support over valid transitions and preserving the intended posterior target.
  2. The polynomial time subset sampling strategy avoids exponential enumeration and provides a computationally efficient approach to partition splitting.
  3. The experimental evaluation is reasonably comprehensive, including exact-posterior comparisons, component ablations, multi-step decision experiments, random-topology scenarios, and trace derived evaluations.
  4. The authors establish theoretical properties, including permutation equivariance and computational complexity, and transparently report negative experimental results and limitations.
Critical weaknesses
  1. Limited methodological novelty: Although PRPT introduces a structured and exactly correctable proposal for dynamic partition inference, its main components, including SMC, split merge proposals, defensive mixtures, and weighted subset sampling, are established techniques. While the authors acknowledge these connections, the extent of the methodological advance over existing observation-informed particle filtering approaches could be more convincingly demonstrated through stronger comparative baselines.
  2. Restrictive transition assumptions: The transition model permits only one split, merge, or stay event per observation interval. Although the authors evaluate different churn rates and acknowledge this limitation, the method's effectiveness when multiple partition changes occur between observations remains unclear.
  3. Limited large-scale global evaluation: Although the paper reports end-to-end inference latency for up to 16 nodes and proposal-level computational costs for configurations up to 64 nodes, larger-scale experiments primarily examine proposal efficiency or independently decomposable domains. End-to-end posterior accuracy, runtime, and memory behavior for larger globally interacting partition systems remain insufficiently evaluated.
  4. Sensitivity to imperfect observations: Table 3 shows that PRPT's posterior accuracy improvements are less consistent under non-exact pressure observations, with confidence intervals crossing zero for total variation and MAP accuracy gains and reduced coverage in some cases. Although these limitations are transparently discussed, further analysis of observation-model misspecification would strengthen the robustness claims.
  5. Limited practical validation: The Alibaba benchmark uses server co-residency as a proxy for contention rather than directly measuring resource interference or production scheduling outcomes. Although the authors acknowledge this limitation, additional validation using actual contention measurements would strengthen the practical relevance of the framework.
Questions for authors
  1. How would PRPT handle multiple split merge events occurring within a single observation interval, particularly in rapidly changing environments?
  2. Could the authors extend the end-to-end evaluation to larger globally interacting partition systems and report posterior approximation accuracy, inference latency, and memory consumption under matched particle budgets?
  3. While the paper compares different residual guidance functions, could the authors provide a more systematic sensitivity analysis of the defensive mixture coefficient, proposal temperature, and residual-loss parameters, particularly under misspecified cardinality observations?
  4. Comparisons with stronger adaptive or auxiliary particle filtering approaches under matched computational budgets would help clarify whether PRPT's improvements arise specifically from its hierarchical partition proposal or from observation-informed proposal adaptation more generally.
1 comment
AuthorsOpenReview ↗

Response: requests for informed controls, coupled scale, and observation sensitivity

Thank you for the requests for informed controls, coupled scale, and observation sensitivity. The new comparisons measure where PRPT improves inference and decisions at comparable compute.

W1/Q4 (informed baselines). On 64 coupled test episodes, 32-particle PRPT exceeds an 80-particle auxiliary-prior filter (APF) by 20.88 [17.35, 24.46] objective units at approximately matched CPU; PRPT leads in all eight test cells by point estimate. APF itself improves on bootstrap by 12.48 [7.42, 17.67], so this is a competitive control. We selected APF's particle count by CPU on separate calibration episodes and tuned its lookahead on disjoint validation episodes. APF minus PRPT policy CPU is -8.55 [-20.77, 4.71] ms per 80-step episode. We also implemented the ESP exact-local comparator, which leads under the factorized model. The APF result therefore supports PRPT's value over the tested generic informed filter at comparable compute.

W2/Q1 (multiple events). PRPT retains a decision benefit when more than one hidden change occurs between delivered snapshots. In the high-churn/delay-four cell, 34.21% of snapshot windows contain at least two changes. Across 1,090 decisions following two or more changes, PRPT makes 13.94 fewer capacity violations [11.13, 16.90] and 18.81 more productive components [9.49, 28.11] per 100 decisions than prior-only SMC. The submitted code applies one pressure-guided transition after a changed delivered snapshot; it does not replay intervening events. True change counts define only this post hoc evaluation stratum, never a policy input. Timestamp replay further improves current-time posterior accuracy, while the submitted update retains the measured decision gain.

W3/Q2 (global scale, posterior, CPU, memory). In 64 paired globally coupled episodes at n=8,12,16,20n=8,12,16,20, with 32 particles per SMC method, PRPT gains 33.36 [28.56, 38.08] objective units, avoids 6.78 [5.58, 7.97] violations, and adds 5.88 [4.64, 7.09] completions over prior-only SMC. In a separate n=8n=8 delayed/lossy test, its exact multi-step current-time posterior TV is 0.0659 versus 0.2452 for prior-only over 960 steps. At n=20n=20, PRPT/exact-local policy-act CPU is 462.6/131.8 ms; first-episode Python allocation is about 147-150/260-351 KB. A separate higher-noise, lower-stay one-step stress test at n=12n=12-20 exposes a submitted inner-PRPT limitation: mean TV is 0.2424 versus 0.1871 for bootstrap, while a post-review 16/16 defensive variant reaches 0.1602. We keep one-step and multi-step targets separate; CPU is not wall latency and Python allocation is not RSS. The coupled results establish decision gains across four sizes, while the stress test identifies where defensive mixing helps.

W4/Q3 (imperfect observations and sensitivity). In the submitted contaminated-pressure benchmark, 32-particle PRPT lowers TV from 0.2062 to 0.0645 and raises truth coverage from 84.5% to 99.0% versus bootstrap. On fixed Day-170 trace cases, all nine temperature/prior-fraction settings and six residual-family/exploration settings yield positive mean TV and ESS changes. The defensive prior fraction matters on non-exact Day 175: 0.25 loses mean TV at all three temperatures, whereas 0.75 gives small positive point estimates; the frozen coverage loss remains. A separate assumed-likelihood sweep distinguishes proposal sensitivity from model mismatch. These tests show a broad Day-170 gain and identify the prior mixture as a useful control under non-exact pressure.

W5 (trace interpretation). The Alibaba adapter uses a request-derived local aggregate constructed upstream from current server membership. Co-residency is a partition label rather than measured physical interference, and public records do not establish online availability of this aggregate. Under that stated proxy, PRPT improves equal-day TV by 0.0421 [0.0382, 0.0460] and adapter-defined utility by 0.0317 [0.0204, 0.0430] over bootstrap across all 5,924 eligible one-arrival episodes on five held-out days and two particle streams. The five-day test therefore supports inference and decision gains under the observation actually evaluated.

Reviewer_cm9NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to track a grouping of workloads (i.e. jobs that share a server) that changes over time by splitting and merging. Specifically, the controller gets two signals: a delayed snapshot of the true grouping, and a noisy number from each workload saying how large its own group is (“pressure” - not sure if this is a standard term). The authors keep a probability distribution over groupings using a particle filter. The contribution is called PRPT, and it’s a better way to propose the next grouping for each particle. On an 8-node synthetic test where the exact solution is available, PRPT with 32 particles gets much closer to it than a standard filter. It also helps on episodes built from an Alibaba cluster trace. On the trace, though, most of the gain disappears when the size readings are not precise.

Critical strengths

Positives:

  • Clear problem setup
  • Results seem sound
  • Clearly stated limitations
Critical weaknesses

Cons:

  • I really don't think this is a good fit for ICLR: nothing is learned.
  • The method seems to be a combination of existing pieces
  • It seems to me like the solution can be computed exactly?!
Questions for authors

Q1: Can’t you compute the best possible proposal exactly? If I understand correctly, each node’s reading depends only on its own group size. So your equivariance assumption means the split prior can only depend on group sizes. If this is the case, then the posterior over splits of a group C into sizes (s, |C|−s) is also a product-weighted fixed-size distribution, with node weights p(y_i | s) / p(y_i | |C|−s). So wouldn't the same ESP recursion gives the exact one-step posterior in roughly cubic time per group? This is your “exact” baseline, but it seems to me like this could be implementable?

Q2: Whu is this a good fit for ICLR? Nothing in the paper is learned; there are no representations and no training.

1 comment
AuthorsOpenReview ↗

Response for methodological contribution and the ESP comparator

Thank you for the ESP argument and for pressing us to identify where PRPT adds value.

W1/Q2 (ICLR fit). PRPT addresses sequential inference over a changing partition when each workload supplies a noisy block-size estimate but no membership labels. This is a structured latent-state problem: the method defines an observation-conditioned, full-support split-merge proposal whose density supports one-step importance correction. The workload setting tests whether the inferred posterior improves decisions. With 32 particles, PRPT lowers exact-benchmark posterior TV from 0.2062 (bootstrap) to 0.0645 and gains 6.36 objective units per episode on average across the submitted six delayed/lossy conditions. The contribution fits probabilistic methods through its inferential mechanism and measured posterior quality; it does not depend on training a policy or representation. Structured inference of this kind also has direct ICLR precedent (auto-encoding SMC, Le et al., 2018, in our references). Moreover, PRPT is a structured amortization of a local Bayes update whose residual scores can be replaced by a learned equivariant scorer trained with filtering variational objectives while preserving the normalized factors (Section 3.2); learning is compatible future work, not a prerequisite.

W2 (method). The methodological object is a normalized law that turns cardinality residuals into linked event, anchor-side size, and fixed-size subset choices while retaining all prior-supported transitions. This permits guided sampling with an evaluable importance weight. SMC and weighted subset sampling are established ingredients; their observation-conditioned coupling for partition transitions is what we test. Removing event or subset guidance raises mean TV by 0.0121 or 0.0126 in 3,000 matched factorized cases, and by 0.0345 or 0.0116 in 1,600 joint-likelihood cases. These ablations isolate gains from the event and subset levels; the chosen size score has no isolated TV gain.

W3/Q1 (exact local proposal). Yes. Under the factorized likelihood, fixing a split size turns its likelihood-weighted subset mass into an ESP; summing over sizes, stay, and merge events gives a polynomial-time fully adapted one-step proposal. Our implementation matches exhaustive normalizers in 45 cases (maximum relative error 4.59×10−154.59\times10^{-15}). This shortcut requires node-factorized likelihoods; the 1,600-case joint-likelihood test in W2 shows that PRPT's normalized proposal and exact correction still apply when likelihoods do not. This comparator is strong: on the paper's six frozen random-topology blind-test configurations (100 paired episodes each), exact-local exceeds submitted PRPT by 4.59 [3.74, 5.47] objective units. PRPT nevertheless exceeds bootstrap by 13.35 [12.03, 14.66] on these same configurations. These differ from the submitted six delayed/lossy conditions above. PRPT's measured advantage is the cost of constructing a correctable proposal for a fresh observation: in cached single-block tests with 32 draws, it is faster in all ten cases at each n=64,128,256n=64,128,256 (median PRPT/exact CPU ratios 0.269/0.135/0.062, including likelihood and weighting). At matched CPU in 100 ambiguous n=32n=32 cases, 64 PRPT particles also reduce anchor co-membership MAE from 0.0716 to 0.0607 relative to 32 exact-local draws (7.23 versus 7.34 ms). These are conditional post-review results. The n=20n=20 full-policy test uses smaller blocks and the uncached submitted policy, where exact-local is faster. A locally exact step also leaves the multi-step ancestor distribution particle-approximated, so exact-local optimality is one-step rather than end-to-end. Together, the comparisons identify a measured large-block regime in which PRPT gives a lower-cost, importance-correctable alternative to exact-local construction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.