- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper introduces ParetoRouter, an ML-based, multi-objective global router for VLSI design using flow matching. It explicitly balances wirelength (WL) and overflow (OF) using a "Average Flow" objective and a Das-Dennis-based Pareto sampling method to generate preference-compliant routing solutions.
Critical strengths
Novel Multi-Objective Approach: Applying flow matching and Pareto sampling to explicitly balance WL and OF trade-offs is an innovative step that addresses a major gap in existing single-metric ML routers.
Generation Efficiency: The method significantly reduces generation time compared to existing end-to-end ML-based solvers (like DSBRouter) while maintaining highly competitive overflow metrics.
Critical weaknesses
-
The absence of open-source code limits the ability to verify and independently reproduce the claimed experimental results.
-
Missing Ablation Studies: The hyperparameters introduced by the loss function and the classifier function lack clear explanations and ablation studies, making it difficult to assess model sensitivity.
-
Ambiguous Preference Selection: The paper does not adequately elaborate on the methodology or practical criteria for choosing between OF-preferred and WL-preferred configurations. It just enumerates a series of weights.
-
Classifier Complexity vs. Benefit: Introducing a classifier function increases the overall model complexity, but it appears to provide only marginal benefits to the final routing performance. The necessity and trade-offs of this added complexity are not well justified.
Questions for authors
See the weaknesses above.
Response to Reviewer dLeN (1/3)
We sincerely thank the reviewer for the two precise and insightful questions. Both concern the division of roles between training and sampling: ParetoRouter learns a preference-agnostic prior during training, while topology selection and the multi-objective trade-off are determined at sampling time using the current routing state. We explain this design principle below and list the corresponding revisions.
Q1. Eq. 7's weighted-mean minimizer and topology selection.
We thank the reviewer for this sharp observation, and we agree with it. We will revise Sec. 4.1 to make the role of the loss explicit.
What Eq. 7 does. π is computed per training net from the two teachers' OF and WL and is held constant (no gradient), so Eqs. 6–7 are regression losses whose sum has a pixel-wise weighted-mean minimizer. Ldis therefore only shapes a learned prior, while the discrete topology choice is made at sampling time, as described below. We will make the phrase "follow the better route" in Sec. 4.1 more precise accordingly. Over all 221,919 training nets, the teachers disagree on 61.0%. Among these, π equals exactly 1/2 on 64.0% (identical OF and WL, e.g., the two equal-length L-shapes of a two-pin net), where Eq. 7 reduces to Eq. 6. Our implementation also includes a small entropy term 0.01p(1−p), which we will add to Sec. 4.1.
What selects the topology. The discrete choice is made at sampling time by the route construction inside the guidance (Alg. 3, line 4). It builds a connected reference route r′ on the current capacity under the ω-weighted cost, with the network's high-confidence pixels as candidates, and the guidance gradient (Eq. 13) moves the prediction toward r′. To verify this division of roles, we ran this sampling step with the paper model on all 221,919 training inputs at ω=(0.5,0.5):
The prediction is thus highest where the teachers agree, lower where they disagree, and close to −1 off-route: agreed pixels become candidates about 5× more often than disputed ones and about 1,000× more often than off-route pixels, and the construction completes the rest. The resulting r′ connects all pins on 100% of the nets. On the disputed segments of the 135k disagreeing nets, r′ follows one teacher in 61.5% of cases (covering at least half of that teacher's disputed pixels and less than half of the other's), takes a route different from both in 35.3%, and keeps both teachers' segments in only 3.3%. The superposition implied by the weighted-mean minimizer is thus resolved by the construction into a single connected route under the requested preference. The learned candidates also matter: compared with constructing from the pins alone ("w/o NN"), they reduce WL by 2.2% on ibm01 (64,648 vs. 66,119).
Why the choice is deferred to sampling. For the 64.0% exact ties, the teachers' OF and WL provide no basis for preferring either topology. Congestion, which could distinguish them, emerges as nets are routed sequentially on a shared capacity map; this information is available at sampling time, through the remaining capacity and the preference ω. During training, the adversarial term complements Eqs. 6–7: unlike them, it is not minimized by the pixel-wise mean, because the discriminator's real samples are individual routes rather than their averages.
On weighted summation in multi-objective loss design. A weighted sum of regression losses toward conflicting targets is minimized by a compromise. When the feasible set is non-convex, that compromise need not be a valid solution. The set of routes is non-convex: the average of the two L-shapes of a two-pin net is a half-intensity rectangle. ParetoRouter therefore uses the weighted loss only to learn a prior, and the topology is selected at sampling time by the construction under the preference ω. The scalarization in Eq. 11 is linear, as in ParetoFlow, and thus targets supported Pareto points; a Tchebycheff scalarization could replace it without changing the rest of the pipeline, which we will discuss as a limitation.