debajyotidasgupta/orderplace-repro-bundle

Reproduction: Order Matters — Unveiling the Hidden Impact of Macro Placement Sequences via Proxy-Guided LLM Evolution

Independent reproduction of ICML 2026 paper #12180 — OrderPlace (Shibing Mo, Jing Liu, Jianchu Xu, Ruilin Wu). arXiv 2606.08904 · OpenReview fPjkFDW9j7 · official code (placeholder only, no runnable source): Explorermomo/OrderPlace.

Part of the Hugging Face × AlphaXiv ICML-2026 reproduction challenge.

What this is

OrderPlace claims that the order in which macros are placed — not just where — decisively affects wirelength, and proposes a proxy-guided LLM evolution framework to discover ordering strategies. The official repository ships only result PNGs (no code), so this is an independent reproduction from scratch: we reimplement the wire-mask-guided greedy placer (the paper's "sensitive probe"), the ordering strategies (Table 1 + the discovered gravity_entropy), a WireMask-EA baseline, and the proxy-guided evolution loop, and we run them on the real ISPD-2005 netlists.

What reproduces

Claim	Result
C1 34.04% / 14.08% wirelength reduction	🟡 direction & mechanism reproduce (order matters: random 3–7× worse; OrderPlace beats WireMask-EA 4–7% on real data); exact % not uniquely reconstructable from the tables
C2 avg rank 1.17, best on 7/8 (Wilcoxon)	✅ 1.17 reproduces exactly excluding the 3 analytic placers (strict all-12 = 1.50); best on 7/8 macro methods
C3 Theorems Ω(l) chain, Θ(g) star	✅ exact — simulation gives ratio ∝ l^1.03 and ∝ g^0.97 (R²≈1.0)
C4 evolution Z=4, τ=0.7, 𝒱=50, 𝒲=8	✅ loop reproduced with exact hyperparameters; monotonic 6.7% proxy improvement (LLM backend substituted)
C5 top-1 always LLM dynamic, 0 static	✅ verified from Table 8 + dynamic beats static on real data
C6 24 strategies: 75% entropy, 62.5% closure	🟡 frequencies exact (18/24, 15/24); but Table 8 lists 22 LLM strategies not 24 (off by 2)

Layout

src/orderplace/       reproduction package
  instance.py         Instance (M,E,G) + chain/star/synthetic topology generators
  placer.py           wire-mask-guided greedy placer (Eq. 1-2, 6-7), separable O(g·deg) mask
  strategies.py       Table-1 static+dynamic strategies + discovered gravity_entropy (Eq. 21-25)
  wiremask_ea.py      WireMask-EA baseline (EA over placement permutations)
  evolution.py        proxy-guided LLM evolution loop (Z=4, τ=0.7, 𝒱=50, 𝒲=8)
  ispd.py             ISPD-2005 bookshelf parser (macros = 'terminal' nodes)
experiments/          one script per claim + plotting
outputs/              result JSON/CSV + interactive HTML/PNG figures per claim
paper.txt             extracted paper text (used for the Table 2/3/7/8 audit)
make_poster.py        builds the reproduction poster
build_logbook.py      populates the Trackio logbook

Reproduce

pip install numpy scipy plotly kaleido pandas openai
# theorems (Claim 3), ~30s
PYTHONPATH=src python3 experiments/claim3_theorems.py --scale full
# table audit (Claims 1,2,5,6), instant
PYTHONPATH=src python3 experiments/claim124_arithmetic.py
# real-data mechanism (Claims 1,2,5) -- needs ISPD-2005 under data/ispd2005/
wget https://www.cerc.utexas.edu/~zixuan/ispd2005dp.tar.xz
tar -xf ispd2005dp.tar.xz -C data/  # gives data/ispd2005/{adaptec1,bigblue1,...}
PYTHONPATH=src python3 experiments/claim125_mechanism.py --scale full --circuits adaptec1 bigblue1
# proxy-guided evolution (Claim 4)
PYTHONPATH=src python3 experiments/claim4_evolution.py --circuit adaptec1 --V 50 --generations 4

Compute

HF Jobs and HF Inference Providers were unavailable for this account (402 Payment Required, no credits). All experiments are CPU-bound (numpy) and ran locally; one scaled theorem sweep (l,g up to 512) ran on a Vast.ai instance as the cloud record. The LLM strategy generator (paper uses claude-sonnet-4.5) is backend- substituted with a programmatic elite-guided generator, per the challenge's backend-substitution policy. Total cloud spend < $1.

Faithfulness notes

• The wire-mask greedy placer is a from-scratch reimplementation of WireMask-BBO's probe (Shi et al., 2023): each macro placed at the grid cell minimising incremental HPWL; validity via integral-image overlap check; the bounding-box cost is separable in x/y so the mask is O(g·deg).
• ISPD-2005 macros are the terminal nodes (543 for adaptec1, 560 for bigblue1, matching Table 6); nets are macro-only hyperedges; geometry is scaled onto a 224×224 grid (the paper's resolution) at the reported utilisation.
• The evolution loop reproduces the paper's procedure and hyperparameters exactly; the LLM's creative strategy generation is approximated by an elite-guided parametric search that still passes through the V=50 / W=8 proxy evaluator.