WaveletECO: A Closed-Loop Physical ECO Platform and a Specialized Local Language Model
Guoxiang Xu, Guozhen Ji, Zijian Luo, Zhengrui Chen, Qi Sun, Cheng Zhuo
cs.AR · cs.AI
Quick verdict
Trained local 9B model beats GPT-6 Astra on 22-design ECO benchmark at ~1/147 cost; plausible, but zero-SD repeats and seen-design overlap need scrutiny.
Summary
WaveletECO couples a closed-loop physical ECO execution platform with a locally trained 9B language model for multi-round repair of timing and electrical violations in late-stage chip design. Built on Qwen3.5-9B, the policy is trained with supervised fine-tuning on 368 execution demonstrations plus CPO-SimPO preference optimization on 759 pairs, using data collected from the platform's validate-execute-verify loop with rollback. Across 594 runs on 22 designs (13 seen, 9 held-out; up to five rounds, three repeats), WaveletECO-Policy scores 79.63 (BF16) and 79.65 (INT8) on an equal-weight setup/hold/electrical score, versus 77.44 for GPT-6 Astra and 76.15 for GPT-5.6 Sol under an identical action space and 1% area-growth cap; the base Qwen3.5-9B scores 73.20. On the nine held-out designs, the BF16 policy reaches 87.13 versus 85.54 for GPT-6 Astra. Using dated API prices and measured or estimated local throughput, INT8's estimated inference cost is about 1/147 of GPT-6's (~150x score-per-dollar). Limitations: training data is small (368 demonstrations, 759 preference pairs); the All-22 headline includes 13 designs used for adaptation or model selection; several reported standard deviations are exactly zero; local cost estimates exclude EDA and idle time; no code release or classical ECO-tool baseline is reported.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 12 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A closed-loop ECO platform with validation, re-measurement, and rollback enables LLM agents to make effective multi-round repair decisions.
- SFT plus CPO-SimPO training lifts a Qwen3.5-9B policy from 73.20 (direct base) to 79.63 (BF16), exceeding GPT-6 Astra's 77.44 under the same protocol on 22 designs.
- INT8 quantization retains repair quality (79.65) at an estimated 1/147 of GPT-6 Astra's per-task inference cost (USD 0.00226 vs 0.33188).
- The trained policy transfers to nine held-out designs, scoring 87.13 versus 85.54 for GPT-6 Astra.
Strengths
Closed-loop, verifiable execution protocol
Section 2 / Figure 1: every proposed action is validated, executed against a shared EDA backend, re-measured, and rolled back if rejected, under coupled guardrails (setup/hold/electrical DRV, DRC, area growth <=1%), so score gains cannot come from unverified claims.
Controlled cross-model comparison
Section 3.1 / Table 2: nine variants (base direct/thinking, SFT, three quantizations, three commercial APIs) run through the same interface, one candidate per round, and the same 1% area cap, isolating decision quality from workflow differences; training stages show a clean progression 73.20 -> 77.29 -> 79.63.
Quantization robustness with cost transparency
Held-out generalization is reported separately
Weaknesses
Very small training corpora
Headline mixes seen and held-out designs
Anomalous variance reporting
Cost comparison rests on estimates and proxies
No classical baseline, no code release
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This five-page systems paper (arXiv 2609.23444, cs.AR/cs.AI; Zhejiang University and ChipFlux) targets engineering change orders (ECO): late-stage repair of timing and electrical violations. It makes three claims: (1) a closed-loop execution platform lets LLM agents make verifiable multi-round repair decisions; (2) a Qwen3.5-9B model fine-tuned with SFT (368 demonstrations) and CPO-SimPO (759 preference pairs) outperforms frontier commercial APIs on this task; (3) INT8 quantization preserves repair quality at an estimated 1/147 of GPT-6 Astra's inference cost. Table 1 positions WaveletECO against ChipNeMo, ChatEDA, RSR, AgenticECO, and two RL-based ECO methods, arguing prior LLM frameworks lack dedicated decision-model training.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is WaveletECO?
A system from Zhejiang University and ChipFlux that pairs a closed-loop physical ECO platform (validate, execute, re-measure, rollback) with a locally trained 9B language model, based on Qwen3.5-9B, that decides which repair actions to apply when fixing timing and electrical violations in late-stage chip design.
How does WaveletECO compare with GPT-6 Astra?
On the paper's 22-design benchmark under identical action rules and a 1% area cap, WaveletECO-Policy scores 79.63 (BF16) and 79.65 (INT8) versus GPT-6 Astra's 77.44, and its estimated INT8 inference cost is about 1/147 of GPT-6's (USD 0.00226 vs 0.33188 per task).
Does INT8 quantization hurt the repair quality?
Not in this paper: the INT8 variant scores 79.65, essentially equal to the BF16 model's 79.63, and INT4 scores 78.67. The authors do note that the deployments used for cost estimation were not themselves re-evaluated for repair quality.
Is code or data available for WaveletECO?
The paper contains no code, prompt, or dataset release statement. Results depend on a proprietary EDA backend and non-public designs (up to 202k cells), so independent reproduction is currently not possible from the published material.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1 and 2 appear only as garbled text fragments; the architecture diagram and the cost-quality scatter could not be inspected.
- Equation 3's local-cost branch renders mangled in the extract; the cost interpretation relies on the surrounding prose.
- This is a five-page extract with no supplementary material: per-design results, prompts, and hyperparameters are absent, so several judgments rest on the summary tables alone.
- Some +/-0.00 SDs could be extraction artifacts; they are raised as questions for the authors rather than asserted as errors.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.