acceptodds
Under review as a conference paper at ICLR 2027

FOCON: Formula-Conditioned Receiver-Aware Clause Communication Network for Parallel SAT Solving

Abstract

Learned-clause exchange is a key mechanism in parallel conflict-driven clause learning (CDCL), whereas existing sharing policies largely assess clause utility through global statistics, overlooking the dependence of utility on formula structure and the evolving search state of the receiver. We formulate clause exchange as formula-conditioned, receiver-aware communication and propose FOCON, a lightweight learned router that selects clause imports without modifying symbolic CDCL. A polarity-equivariant factor network processes each CNF once to produce signed structural context and formula-specific routing coefficients. Each candidate transfer is then scored from clause structure and directed sender-receiver states through a compact event-level scorer. FOCON learns from latency-discounted clause reuse, regulates communication through adaptive budget control, and falls back to native sharing outside the learned support. Experiments demonstrate that FOCON achieves aggregate gains over state-of-the-art learning-guided methods and solver baselines while accepting only about of candidate transfers and remaining effective across unseen SAT families and distinct problem encodings.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
SAT solving, learned-clause sharing, clause routing, graph neural networks, neural algorithmic reasoning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Gdy6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents FOCON, a lightweight formula‑conditioned, receiver‑aware router for parallel CDCL SAT solvers. FOCON employs a polarity‑equivariant factor network and an event‑level scorer with latency‑discounted clause reuse, enabling adaptive communication control and robust out‑of‑distribution (OOD) fallback. Evaluations demonstrate strong performance in both solving efficiency and communication overhead, while maintaining correctness and cross‑family generalization.

Critical strengths
  1. The problem modeling is clear and has practical significance: shifting the clause exchange from "clause quality ranking" to a joint decision of "sender-receiver-formula-search phase." The example in Figure 1 intuitively illustrates the receiver's dependency on utility, providing sufficient motivation.
  2. The method demonstrates good design in calculating amortization and deployment constraints: formula-level graph inference is performed only once, and online decision-making degrades to low-dimensional inner products, making it suitable for high-frequency clause exchanges. Additionally, budget control and OOD (Out-of-Distribution) fallback are introduced to balance efficiency, communication constraints, and conservativeness under distributional shifts. The experiments also report ablation studies and theoretical analyses.
Critical weaknesses
  1. The Baseline parallel backend is outdated and has a single scale. The main experiment only used the 8-worker Glucose-Syrup (circa 2014) and did not compare it with modern parallel solvers from the recently SAT Competition (such as MallobSat, Painless-PRS-Kissat, PRS) or 16/32/64-worker setups, making it difficult to assess the method's actual incremental value in the current parallel SAT ecosystem.
  2. The communication budget is not strictly aligned with the baseline. The FOCON acceptance rate is approximately 20%, but Glucose-Syrup's native lazy clause exchange already has LBD filtering and a budget mechanism; the paper's "20% vs 100%" may overestimate selectivity, as a fair comparison was not conducted under the same candidate stream and equivalent communication budget.
  3. The presentation of the paper includes a high density of symbols and formula derivations, with relatively limited intuitive explanations. Recommend to optimize the expression to improve readability.
Questions for authors
  1. Can FOCON be verified on modern parallel backends (MallobSat / Painless / PRS) and with 16~64 workers? Do the benefits perceived by the receiver remain constant as the number of workers increases, and how does the router overhead ratio change? 2.Glucose-Syrup's native sharing strategy already incorporates LBD filtering and budget control. Could we consider adjusting its acceptance rate to approximately 20%, similar to FOCON, and run another end-to-end control experiment? This will allow us to more clearly separate the contributions from "routing selectivity" and "backend differences".
4 comments
AuthorsOpenReview ↗

Modern Parallel Solver Integration, Worker Scalability, and Budget-Matched Comparisons

We thank the reviewer for the insightful comments and constructive suggestions, which have helped strengthen the presentation and experimental evaluation. All revisions and additional experiments are highlighted in blue in the updated manuscript. We address the weaknesses (W) and questions (Q) below.

1. [W1/Q1. Modern parallel backends and worker scalability]

Re. We have integrated FOCON into MallobSat, Painless, and PRS at the clause-import interface, preserving native CDCL search, clause export, and communication topology (Appendix D.2). FOCON improves SR and PAR-2 across all four backends, with PAR-2 reductions of 4.80%, 3.47%, 2.23%, and 1.00% on Glucose-Syrup, PRS, Painless, and MallobSat, respectively (Table 5). FOCON further reduces PAR-2 in 7 of 8 backend-timeout configurations on SAT Competition 2025-2026 (Table 23), demonstrating the complementary value of learned routing across modern parallel architectures.

We have also extended the evaluation to 16 and 32 workers on Glucose-Syrup (Appendix E.7, Table 19):

WorkersSR (%) Native → FOCONPAR-2 reduction vs. NativePAR-2 reduction vs. LBD-20Reuses/acc. Native → FOCONRouter share
884.35 → 85.064.8%5.1%4.38 → 20.511.43%
1685.29 → 86.298.7%5.6%3.63 → 17.481.79%
3286.12 → 87.2910.6%4.8%2.92 → 16.482.21%

As the portfolio grows, the candidate-event rate increases 4.4 times, from 64290 to 284522 events/s. Although absolute reuse per accepted transfer decreases, FOCON's relative reuse advantage over native sharing increases from a factor of 4.7 to 5.6. At 8 workers, FOCON retains 98.7% of native reuses while accepting only 152.09K of 721.68K clauses (Table 3). The PAR-2 reduction over native sharing increases from 4.8% to 10.6%, while the improvement over LBD-20 remains between 4.8% and 5.6%. These results highlight the increasing value of receiver-aware selection under higher communication loads and the benefit beyond LBD-based selection.

FOCON maintains acceptance near 20%, with online routing accounting for only 1.43-2.21% of solver CPU time.

2. [W2/Q2. Budget-matched comparison with native sharing]

Re. Glucose-Syrup already incorporates LBD-based filtering and communication regulation. The reported 100% acceptance refers to native-eligible transfers after native filtering, rather than all generated learned clauses. FOCON therefore provides an additional layer of receiver-aware selection.

To isolate the contribution of learned selection, we evaluate two budget-matched controls (Table 4, Appendix D.2):

  • Random-20: Uniform random scores replace learned scores.
  • LBD-20: Negative origin-worker LBD scores prioritize low-LBD clauses.

Both controls share FOCON's backend, native eligibility criteria, import interface, and adaptive controller at ρ=0.2\rho=0.2, while varying the selection mechanism. Realized candidate streams remain policy-dependent.

PolicyAccept. (%)SR (%)PAR-2 (s)Data (GiB)
Glucose-Syrup100.0084.3524.37241.13
Random-2020.0082.1225.6853.65
LBD-2020.0083.8224.4533.48
FOCON19.9585.0623.2028.26

Under matched acceptance targets, FOCON reduces PAR-2 by 9.66% and 5.11% over Random-20 and LBD-20, respectively, while achieving higher SR. FOCON also transfers 15.59% less data than LBD-20 and 88.3% less than native sharing, and is the only evaluated 20% policy that improves both SR and PAR-2 over native sharing. Both heuristics fall below native sharing, indicating that a 20% budget alone degrades performance unless transfers are selected by receiver-specific utility. Despite transferring 1.9 times as much data as FOCON, Random-20 exhibits the lowest SR and the highest PAR-2.

On SAT Competition at 1000s (Table 23), Random-20 and LBD-20 solve 5 and 8 fewer instances than native Glucose-Syrup, respectively, whereas FOCON solves 5 additional instances and reduces PAR-2 by 4.27%. These results demonstrate the added value of receiver-aware clause selection under matched communication budgets.

3. [W3. Presentation and notation density]

Re. We have revised Sections 4 and 5 to establish algorithmic intuition before mathematical formalization. Section 4 now introduces three operations linked to Figure 2: formula conditioning, receiver-aware scoring, and support-aware budget control. We clarify the roles of literal and clause embeddings (hℓ\mathbf{h}_\ell, gj\mathbf{g}_j), explain how worker-state features (sw\mathbf{s}_w) characterize receiver dependence, and explicitly define all training losses (Eqs. 11-12). Section 5 now introduces the main theoretical results before their formal statements and summarizes their practical implications.

We hope that our responses have adequately addressed the reviewer's concerns and helped improve the quality of the paper. We welcome any further questions or discussion.

Dear Reviewer,

As the discussion period approaches the end, we would be grateful if you could review our rebuttal and let us know whether our responses have adequately addressed your concerns. If any concerns remain, please feel free to share them, and we would be pleased to provide further clarification.

Thank you very much for your time and consideration.

Best regards,

The Authors

The authors have supplemented multi-backend experiments on MalloSAT/Painless/PRS, extended the evaluation to 16/32 workers, and added random-20/LBD-20 budget-matched controls, which substantively address my core concerns. I'll raise my score accordingly.

Dear Reviewer,

Thank you very much for revisiting our response and raising your score. We sincerely appreciate your time, constructive feedback, and positive reassessment of our work.

Best regards,

The Authors

Reviewer_YjwXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FOCON, a lightweight learned router that selects clauses in portfolio-based parallel SAT solving framework. The existing clause sharing policies overlook the dependence of clause utility on formula structure and the receiver’s evolving search state. FOCON uses a polarity-equivariant GNN to encode each formula once, followed by lightweight scoring of candidate transfers using clause features and sender–receiver states. The experimental results show PAR-2 reduction compared to the baseline methods.

Critical strengths
  1. The paper clearly explains why the same learned clause can provide different benefits to different receivers, motivating recipient-specific clause-sharing decisions.
  2. FOCON shows better results than the other baselines in Table 1 on the proposed benchmarks.
Critical weaknesses
  1. The experimental evaluation is insufficient. In Sec. 6.1, the evaluation benchmark contains only small-scale CNF instances, and most of them can be solved within 60 seconds.
  2. The baseline method is not strong enough. In Table 1, Glucose-Syrup is the only parallel baseline.
Questions for authors
  1. Please provide the results with annual SAT competition benchmarks. (https://benchmark-database.de/)
  2. For the larger problem instances mentioned above, how about comparing FOCON with other related portfolio-based parallel SAT solvers, such as MallobSat or Painless?
  3. What’s the additional runtime overhead of the proposed sharing policy?
5 comments
AuthorsOpenReview ↗

Expanded SAT Competition Evaluation, Modern Solver Comparisons, and Runtime Scalability Analysis

We thank the reviewer for the valuable feedback and constructive suggestions. All revisions and additional experiments are highlighted in blue in the updated manuscript. We address the weaknesses (W) and questions (Q) below.

1. [W1/Q1. Benchmark scale and SAT Competition evaluation]

Re. We have extended the evaluation to the SAT Competition 2025-2026 main tracks (Appendix E.10, Table 23), using the GBD benchmark database suggested by the reviewer. The router is trained on the 2023-2024 main tracks, ensuring a strict temporal split. We evaluate 240 CNFs, with 60 SAT and 60 UNSAT instances per year. All methods are evaluated under 600s and 1000s end-to-end timeouts.

Backend (w/o → w/ FOCON)SR (%) 600sPAR-2 (s) 600sSR (%) 1000sPAR-2 (s) 1000s
Glucose-Syrup41.25 → 42.92758.29 → 732.9247.92 → 50.001149.19 → 1100.08
MallobSat67.92 → 68.33424.63 → 423.3579.17 → 79.17546.63 → 549.56
Painless61.25 → 61.67553.75 → 549.2964.58 → 65.42822.31 → 806.04
PRS60.00 → 60.83571.35 → 560.2966.67 → 67.92808.83 → 782.32

FOCON improves both SR and PAR-2 across all four backends at 600s and reduces PAR-2 in 7 of 8 backend-timeout configurations. At 1000s, FOCON solves 5 additional instances over native Glucose-Syrup and reduces PAR-2 by 4.27%. In contrast, Random-20 and LBD-20 solve 5 and 8 fewer instances than native sharing at the same acceptance target.

These results demonstrate the effectiveness of learned clause selection on challenging competition instances beyond the benefits of communication-budget control. Table 18 (Appendix E.6) shows that FOCON outperforms native sharing at all deadlines starting from 1s, with the largest SR improvement at 5s (62.88% vs. 60.84%).

2. [W2/Q2. Comparison with modern parallel SAT solvers]

Re. We have integrated FOCON into MallobSat, Painless, and PRS at the clause-import interface, preserving native CDCL search, clause export, and communication topology (Appendix D.2). Comparing each backend with and without FOCON isolates the contribution of receiver-aware routing from architectural differences.

On the in-distribution benchmark (Table 5), FOCON improves SR and PAR-2 across all four backends:

BackendPAR-2 w/o → w/ FOCONReduction
Glucose-Syrup24.37 → 23.204.80%
MallobSat13.94 → 13.801.00%
Painless21.50 → 21.022.23%
PRS25.36 → 24.483.47%

On competition instances at 1000s (Table 23), FOCON solves 3 additional instances with PRS and 2 with Painless, reducing PAR-2 by 3.28% and 1.98%, respectively, while MallobSat maintains identical coverage (190/240) with a minor PAR-2 increase of 0.54%.

Together, these results establish that FOCON complements modern clause-sharing mechanisms rather than relying on a specific backend; its gains on MallobSat at 600s and in-distribution show applicability even to advanced parallel SAT architectures.

3. [Q3. Runtime overhead]

Re. We have added detailed formula-level and event-level profiling on in-distribution benchmark (Appendix E.8, Tables 20-21) and a worker-scaling analysis (Appendix E.7, Table 19).

ComponentMeasured cost
Graph construction (once per CNF)2.24 ms (median)
GNN inference (once per CNF)320 ms (median); 365 ms (95th percentile)
Online routing (per scored event)2.01 μs (mean)
Router CPU share1.43% (1.02-1.52% across families)
Amortized preprocessing per event0.229-0.999 μs
Median peak memory~1 GB

FOCON separates one-time formula reasoning from high-frequency clause routing. Each online decision costs O(∣Ce∣+dϕ)O(\lvert C_e\rvert+d_\phi) for clause aggregation and a fixed-dimensional inner product. Formula-level message passing scales linearly with factor-graph size for fixed network width and depth (Appendix B.4).

Across formula groups with median literal occurrences from 4153.5 to 58366, mean formula-level time remains approximately 0.3s, amortized preprocessing stays below 1 μs per event, and median peak memory stays near 1 GB (Table 21).

As the portfolio scales from 8 to 32 workers, the candidate-event rate increases from 64290 to 284522 events/s, while router CPU share increases only from 1.43% to 2.21%. Meanwhile, the PAR-2 reduction over native sharing increases from 4.8% to 10.6% (Table 19), showing that learned routing benefits more under higher communication loads.

All reported solving times are end-to-end, including graph construction, GNN inference, and online routing. FOCON reduces mean RT from 14.99s to 14.25s and PAR-2 from 24.37s to 23.20s relative to Glucose-Syrup (Table 1).

We hope that our responses have adequately addressed the reviewer's concerns and helped improve the quality of the paper. We welcome any further questions or discussion.

Dear Reviewer,

As the discussion period approaches the end, we would be grateful if you could review our rebuttal and let us know whether our responses have adequately addressed your concerns. If any concerns remain, please feel free to share them, and we would be pleased to provide further clarification.

Thank you very much for your time and consideration.

Best regards,

The Authors

I thank the authors for their response and clarifications. I therefore lean towards acceptance ; since I am not an expert in the field of SAT solving and have therefore a limited appreciation of the improvement on the field, I will carefully review the other reviews and consider updating my rating during the reviewer-AC discussion.

Reviewer_a9tHOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces FOCON, a trained router for assisting in SAT solving. When solving a SAT problem, several solvers attack the problem in parallel, and solve conflicts by introducing new clauses and sharing them with the other solvers. FOCON trains a router that determines the relevance of sharing a clause between two instances, by taking into account the current state of the receiver. FOCON achieves state of the art results with only 20% of the clauses being transfered. The router is trained through a surrogate measure of relevance computed as an score that decreases the longer the candidate clause remains unused.

Critical strengths
  • the addressed problem is relevant: preventing the sharing of redundant clauses in parallel SAT problem solving and speeding up the process, with results indicating that this is particularly true to identify UNSAT cases.
  • the work is well motivated and extensively described, with the different parts of the framework detailed.
  • the authors propose theoretical support that motivates their approach, by showing that receiver-aware transfer is bound to improve performance compared to an agnostic one.
Critical weaknesses
  • as a general comment the paper (especially section 4) is very dense and notation-heavy. Many variables and formulas are introduced as-is, without any proper high-level motivation. For instance, in 4.1, h_l and g_l are introduced without any comment as to their role in the framework, leaving the reader to piece everything from the text and equations below. I would recommend introducing such variables (especially given the number of different moving parts) properly to make reading more linear. This is to a lesser extent applicable to section 5, where a high-level takeaway should be provided before the equation-heavy machinery.
  • the losses are not detailed, but they are a critical part of the framework. They should be provided in full, with their inputs for the readers to understand their individual role in detail.
Questions for authors
  • what are the explicit losses(equation 10) for the model?
  • the surrogate introduced in 4.3 can only be computed from the clauses that were added to the model. How to evaluate the utility of a clause that is never added during training (and therefore never reused)?
2 comments
AuthorsOpenReview ↗

Clarified Methodology, Explicit Training Objectives, and Clause Utility Validation

We thank the reviewer for the constructive feedback and insightful suggestions. All revisions and additional analyses are highlighted in blue in the updated manuscript. We address the weaknesses (W) and questions (Q) below.

1. [W1. Notation density]

Re. We have revised Sections 4 and 5 to introduce intuition before formalization. Section 4 now presents three operations linked to Figure 2: formula conditioning, receiver-aware scoring, and support-aware budget control. Section 4.1 explains how literal and clause embeddings derive the signed coordinate pip_i, while Section 4.2 connects worker-state features to the receiver dependence in Figure 1. Section 5 now introduces the theoretical results before their formal statements and summarizes their implications. The computational pipeline is summarized below:

SymbolRoleComputed
hℓ,gj\mathbf{h}_\ell,\mathbf{g}_jLiteral and clause embeddingsOnce per CNF
pF,wF,zF\mathbf{p}_F,\mathbf{w}_F,\mathbf{z}_FSigned context, routing coefficients, support embeddingOnce per CNF
so,sr\mathbf{s}_o,\mathbf{s}_rOrigin and receiver statesOnline
se=wF⊤ϕes_e=\mathbf{w}_F^\top\boldsymbol\phi_eReceiver-aware scorePer event
θ\thetaAdaptive threshold tracking ρ\rhoPer event

FOCON amortizes graph reasoning across the search, leaving clause aggregation, an inner product, and threshold updates online. The theory establishes symmetry invariance, the value of receiver conditioning under a fixed budget, and acceptance-rate tracking.

2. [W2/Q1. Explicit training losses]

Re. We have fully specified the objective in Eqs. 11-12 of Section 4.3 and Appendix B.2: J=Jrob+λbJbudget+λiJinv+λaJac.\mathcal{J}=\mathcal{J}_{\mathrm{rob}}+\lambda_b\mathcal{J}_{\mathrm{budget}}+\lambda_i\mathcal{J}_{\mathrm{inv}}+\lambda_a\mathcal{J}_{\mathrm{ac}}.

(1) Robust prediction. The censor-aware domain loss is Jg=−1∣Eg∣∑eωe[β+yelog⁡p^e+(1−ye)log⁡(1−p^e)].\mathcal{J}_g=-\frac{1}{\lvert\mathcal{E}_g\rvert}\sum_e\omega_e[\beta_+y_e\log\hat p_e+(1-y_e)\log(1-\hat p_e)]. The smoothed worst-domain loss is Jrob=τrob[log⁡∑gexp⁡(Jg/τrob)−log⁡∣DB∣].\mathcal{J}_{\mathrm{rob}}=\tau_{\mathrm{rob}}[\log\sum_g\exp(\mathcal{J}_g/\tau_{\mathrm{rob}})-\log\lvert\mathcal{D}_B\rvert]. Here, ye=exp⁡(−Δe/τu)y_e=\exp(-\Delta_e/\tau_u) for observed reuse and zero otherwise; ωe\omega_e discounts censored negatives, β+\beta_+ offsets class imbalance, and DB\mathcal{D}_B denotes mini-batch domains with event sets Eg\mathcal{E}_g. Robust aggregation emphasizes high-loss domains and limits dominance of frequent families.

(2) Budget shaping. It aligns the mean batch score pˉB\bar p_B with the target ρ\rho: Jbudget=(pˉB−ρ)2.\mathcal{J}_{\mathrm{budget}}=(\bar p_B-\rho)^2. Online thresholding then regulates realized acceptance.

(3) Domain invariance. It aligns domain-level support embeddings: Jinv=1q∣DB∣∑g∥zˉg−zˉ∥22.\mathcal{J}_{\mathrm{inv}}=\frac{1}{q\lvert\mathcal{D}_B\rvert}\sum_g\lVert\bar{\mathbf{z}}_g-\bar{\mathbf{z}}\rVert_2^2. Here, zˉg\bar{\mathbf{z}}_g is the domain mean and zˉ\bar{\mathbf{z}} their average.

(4) Anti-collapse. It preserves embedding dispersion: Jac=1q∑j[νmin⁡−νj]+2,\mathcal{J}_{\mathrm{ac}}=\frac{1}{q}\sum_j[\nu_{\min}-\nu_j]_+^2, where νj\nu_j is the batch dispersion, νmin⁡\nu_{\min} is the minimum target.

Together, these losses align domains while keeping dispersion informative for OOD detection; removing the OOD gate raises OOD PAR-2 from 1.24s to 1.65s (Table 6).

3. [Q2. Utility estimation for clauses]

Re. FOCON learns from native solver traces collected before learned filtering (Section 4.3). Thus, native imports that FOCON would later reject still provide reuse supervision. Earlier reuse receives higher utility, while censor-aware weighting accounts for limited follow-up. Clauses never imported natively are excluded, consistent with routing over native-eligible transfers.

To evaluate predictive quality, we add a native-trace reuse audit (Appendix E.9, Table 22), grouping 81.9M held-out native imports by hypothetical FOCON decisions without altering the observed trajectories.

MetricWould acceptWould rejectRatio
Reuse rate6.07%1.43%4.24 times
Mean yˉe\bar y_e0.04200.00755.60 times

The routing score achieves an AUC of 0.776 (0.666-0.810 across families), consistently separating high-utility from low-utility transfers. This translates into communication efficiency: FOCON reduces accepted transfers from 721.68K to 152.09K while retaining 98.7% of native reuses (3.12M vs. 3.16M; Table 3). Under matched budgets, FOCON reduces PAR-2 by 9.66% and 5.11% over Random-20 and LBD-20, respectively (Table 4). This confirms the benefit of learned clause selection beyond communication-budget control.

We hope that our responses have adequately addressed the reviewer's concerns and helped improve the quality of the paper. We welcome any further questions or discussion.

Dear Reviewer,

As the discussion period approaches the end, we would be grateful if you could review our rebuttal and let us know whether our responses have adequately addressed your concerns. If any concerns remain, please feel free to share them, and we would be pleased to provide further clarification.

Thank you very much for your time and consideration.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Point-by-Point Responses

Dear Area Chair and Reviewers,

As the discussion period approaches its conclusion, we would like to sincerely thank the Area Chair and all reviewers for their time, careful evaluations, and constructive feedback. We have made substantial revisions and conducted extensive additional experiments to directly address the concerns raised in the reviews. All changes and new results are highlighted in blue in the revised manuscript. We summarize our responses and key improvements below.

Strengths

  • The reviewers recognize the practical significance and strong motivation of efficient learned-clause communication in parallel SAT solving. They acknowledge the value of receiver-aware clause selection, the lightweight and efficient architecture of FOCON, and its compatibility with existing CDCL solvers. The reviews also highlight the theoretical justification, adaptive communication-budget control, OOD fallback mechanism, and promising experimental results demonstrating improvements in solving efficiency and communication overhead.

Questions and Weaknesses

  • Notation density and methodological clarity (a9tH W1, Gdy6 W3): Restructured Sections 4-5 to introduce algorithmic intuition before formalization, clarify key components, and improve readability.

  • Explicit training objectives (a9tH W2/Q1): Fully specified all training losses and their roles in Section 4.3 and Appendix B.2.

  • Clause utility estimation and validation (a9tH Q2): Clarified native-trace supervision and added a clause-reuse audit over 81.9M held-out imports, achieving an AUC of 0.776 (Table 22).

  • Benchmark scale and SAT Competition evaluation (YjwX W1/Q1): Extended evaluation to 240 CNFs from SAT Competition 2025-2026 under 600s and 1000s timeouts, with a strict temporal training-evaluation split. FOCON reduces PAR-2 in 7 of 8 backend-timeout configurations (Table 23).

  • Modern parallel solver comparisons (YjwX W2/Q2, Gdy6 W1/Q1): Integrated FOCON into MallobSat, Painless, and PRS. FOCON improves both SR and PAR-2 across all four evaluated backends on the in-distribution benchmark (Table 5).

  • Worker scalability (Gdy6 W1/Q1): Extended experiments to 8, 16, and 32 workers. PAR-2 reduction over native sharing increases from 4.8% to 10.6%, while router CPU usage remains between 1.43% and 2.21% (Table 19).

  • Budget-matched comparisons (Gdy6 W2/Q2): Introduced Random-20 and LBD-20 under matched acceptance targets. FOCON reduces PAR-2 by 9.66% and 5.11%, respectively, demonstrating the benefit of receiver-aware selection beyond communication-budget control (Table 4).

  • Runtime overhead analysis (YjwX Q3, Gdy6 Q1): Added detailed profiling of preprocessing, online routing, CPU usage, and memory consumption, with an average online routing cost of only 2.01 microseconds per event (Tables 19-21).

We believe that our substantial revisions, extensive additional experiments, and detailed clarifications have directly addressed the reviewers' concerns and significantly strengthened both the technical contributions and empirical validation of FOCON. We sincerely hope that the reviewers and Area Chair will take these improvements and additional evidence into careful consideration during the next stage of the review process and the final evaluation.

Thank you once again for your time and thoughtful consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.