solver.press

← All replicated papers

Replicated under our protocolAuthors reviewedOct 26, 2026

TreeWalker: Partial Evaluation for Grouped Tree-Ensemble Inference

Karatay and Newman · arXiv 2610.03939

The outcome is the verdict on the one claim below, under the protocol we fixed in advance. It is not a verdict on the paper as a whole.

Claim tested

Partial evaluation speeds up tree-model inference over a row-independent walk.

What we ran and what came back

Tested twice, at the paper's first code and at its camera-ready code, with the authors' own benchmark harness on the public SUPPORT and FLCHAIN data (LightGBM and XGBoost, 15 fresh timing runs on one pinned hyperthread of a shared-cloud CPU, SMT and ASLR on). Group of 16: 3.7x at both tags (bounds 1.9x and 1.8x). Group of 128: 7.3x at the first tag and 10.9x at the camera-ready tag (bounds 4.2x and 5.0x). Our ratios are higher than the paper's at group size 16 and, for the camera-ready tag, at 128; the cause is not known. One machine, two of the paper's three datasets, wall-clock ratios; at the camera-ready tag XGBoost at group size 1 shows about 1.2x, because the baseline path is slower on XGBoost on this CPU, so part of its gain is not row sharing. The authors reviewed the wording and noted that the first code's native-XGBoost timings are invalid, which our comparison (TreeWalker against its own full walk) does not use.

What this result supports

On the authors' own harness the code gives a large speedup over its row-independent walk for SUPPORT/FLCHAIN, LightGBM/XGBoost at G=16 and G=128, at both the first and the camera-ready tag.

What it does not support

The paper's exact magnitudes on other hardware; the Arm machine or Expedia; that XGBoost's gain at larger G is all row amortisation (G=1 is 1.2x).

Record

Protocol, deposited before the run: 10.5281/zenodo.23462095, 10.5281/zenodo.23639314

The authors reviewed this result before it was listed.

How the protocol works is on the method page. If you are an author of this paper and something here is wrong, write to customers@aegismind.app and we will correct it on this page.