Question about External Connectivity (T2B) Distance Distribution in the Hidden Dataset

Dear ICCAD 2026 FloorSet Challenge Organizers,

We are currently trying to understand why our Beta hidden results differ significantly from our results on the released datasets.

For our submission, the score on the Beta hidden set is roughly **2× worse on almost every testcase** compared with our results on the 1M training data and the 100 public validation cases. Because this behavior is quite consistent across the testcases, we suspect that our method may be overfitting to a particular feature of the released FloorSet data.

One feature we found our current model relies heavily on is the **locality of external connectivity (T2B)** — specifically, the tendency for a terminal to connect to a block that is geometrically nearby in the ground-truth placement.

After looking at Algorithm 4 of the FloorSet paper, we analyzed the distance distribution of both T2B and B2B connections in the released datasets.

For each connection, we normalized the distance and grouped it by relative distance range:

| Normalized distance | Validation 100 T2B | Train 1M T2B | Validation 100 B2B | Train 1M B2B |
| ------------------- | -----------------: | -----------: | -----------------: | -----------: |
| 0–10% nearest | 68.397% | 72.333% | 14.757% | 16.787% |
| 10–20% | 28.606% | 25.457% | 27.992% | 30.447% |
| 20–30% | 2.779% | 2.039% | 28.242% | 27.807% |
| 30–40% | 0.191% | 0.149% | 18.848% | 16.386% |
| 40–50% | 0.0213% | 0.0178% | 7.845% | 6.535% |
| 50–60% | 0.00284% | 0.00333% | 1.949% | 1.714% |
| 60–70% | 0.00142% | 0.00054% | 0.315% | 0.290% |
| 70–80% | 0% | 0.000044% | 0.0519% | 0.0319% |
| 80–90% | 0% | 0.000058% | 0% | 0.00178% |
| 90–100% farthest | 0% | 0% | 0% | 0.000041% |

The difference between **T2B and B2B is very large**.

For T2B, around **97%–98% of the connections fall within the nearest 20% distance range**, and connections beyond 30% are almost nonexistent. In comparison, B2B connections have a much broader distance distribution.

This suggests that external connectivity in both the training and validation datasets has a particularly strong nearby-block locality feature, rather than simply following the general connectivity distribution of the floorplan.

We also compared this behavior with Algorithm 4 of the FloorSet paper, which defines:

`TSim = 1 - Normalize(TDist)`

and samples terminal-to-block connectivity according to `TSim`.

The observed T2B locality in the released datasets appears substantially stronger than what we obtain from a straightforward linear interpretation of this formula.

This feature has a very large impact on our method. As a stress test, if we make all `TSim` values equal, so that the connected block is selected without any distance preference, our score can degrade to around **4**.

Therefore, we would like to ask:

**In the hidden test set, does the external connectivity (T2B) preserve a similarly strong nearby-block locality property as the 1M training set and the public validation set?**

Thank you very much for your clarification.

Start by reading Algorithm 4 of the FloorSet paper and the issue's reported T2B and B2B distance distributions. Compare the stated sampling rule with the released-data observations; done means an organizer-authored clarification of whether the hidden set preserves the reported T2B locality.

Written by the indexing model from the issue text.