Paper deep dive
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
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 9/27/2026, 2:43:28 AM
Summary
The paper introduces WaveletECO, a closed-loop physical Engineering Change Order (ECO) platform that integrates large language models with a specialized local 9B model (Qwen3.5-9B). The model is trained using supervised fine-tuning (SFT) and preference optimization (CPO-SimPO) on execution demonstrations. WaveletECO-Policy (INT8) achieves a repair score of 79.65 across 22 designs, outperforming GPT-6 Astra (77.44) while reducing inference costs to approximately 1/147th of the commercial model.
Entities (8)
Relation Signals (6)
WaveletECO → solvestask → Engineering Change Order
confidence 95% · WaveletECO... enable agents to execute ECO decisions effectively
WaveletECO → usesmodel → Qwen3.5-9B
confidence 95% · We propose WaveletECO... We also train a local 9B model... Qwen3.5-9B
WaveletECO-Policy (INT8) → outperforms → GPT-6 Astra
confidence 92% · WaveletECO-Policy (INT8) score 79.65... compared with GPT-6 Astra's 77.44
WaveletECO-Policy (INT8) → haslowercostthan → GPT-6 Astra
confidence 90% · estimated inference cost of INT8 is about 1/147 of GPT-6 Astra's
Qwen3.5-9B → trainedwith → CPO-SimPO
confidence 90% · train a local 9B model through supervised fine-tuning and CPO-SimPO
TorchAO → enablesquantization → Qwen3.5-9B
confidence 85% · TorchAO [11] quantizes weights to INT8/INT4
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Engineering change order (ECO) is an important step in repairing timing and electrical violations during the late stages of chip design. Existing Agentic EDA methods primarily focus on tool invocation, with less attention to model decision quality and targeted training. A central challenge in ECO is multi-round decision-making: the model must use the results of each round to determine the next repair action. We propose WaveletECO, which integrates a closed-loop execution platform with large language models to enable agents to execute ECO decisions effectively. We also train a local 9B model through supervised fine-tuning and CPO-SimPO using execution demonstrations and decision-preference data, enabling ECO decision-making with a locally deployed model. Across 594 evaluation runs on 22 designs, WaveletECO-Policy (BF16) and (INT8) score 79.63 and 79.65, respectively, compared with GPT-6 Astra's 77.44. The estimated inference cost of INT8 is about 1/147 of GPT-6 Astra's. These results show that specialized model training supports effective, low-cost multi-round ECO repair, with repair quality retained under INT8 quantization.
Tags
Links
- Source: https://arxiv.org/abs/2609.23444v1
- Canonical: https://arxiv.org/abs/2609.23444v1
Trouble viewing inline? Open PDF directly →
Full Text
13,400 characters extracted from source content.
Expand or collapse full text
WaveletECO: A Closed-Loop Physical ECO Platform and a Specialized Local Language Model Guoxiang Xu Affiliation: Zhejiang University Email: qisunchn@zju.edu.cn Guozhen Ji Affiliation: Zhejiang University Zijian Luo Affiliation: Zhejiang University Zhengrui Chen Affiliation: Zhejiang University Affiliation: ChipFlux Qi Sun Affiliation: Zhejiang University Affiliation: ChipFlux Affiliation: Corresponding authors. Cheng Zhuo Email: czhuo@zju.edu.cn*Equal contribution. Affiliation: Zhejiang University Affiliation: Corresponding authors. Abstract Engineering change order (ECO) is an important step in repairing timing and electrical violations during the late stages of chip design. Existing Agentic EDA methods primarily focus on tool invocation, with less attention to model decision quality and targeted training. A central challenge in ECO is multi-round decision-making: the model must use the results of each round to determine the next repair action. We propose WaveletECO, which integrates a closed-loop execution platform with large language models to enable agents to execute ECO decisions effectively. We also train a local 9B model through supervised fine-tuning and CPO-SimPO using execution demonstrations and decision-preference data, enabling ECO decision-making with a locally deployed model. Across 594 evaluation runs on 22 designs, WaveletECO-Policy (BF16) and (INT8) score 79.63 and 79.65, respectively, compared with GPT-6 Astra’s 77.44. The estimated inference cost of INT8 is about 1/147 of GPT-6 Astra’s. These results show that specialized model training supports effective, low-cost multi-round ECO repair, with repair quality retained under INT8 quantization. 1 Introduction and Related Work Physical ECO [1][2] repeatedly repairs timing and electrical violations under area constraints, requiring model decisions to be executed, verified, and retained across rounds. Recent work brings language models into this process. ChipNeMo [3] and ChatEDA [4] explore model adaptation for chip design and EDA, while RSR [5] and AgenticECO [6] provide ECO agent workflows. These ECO frameworks establish execution workflows but do not report dedicated training for repair decisions (Table 1). WaveletECO trains a local 9B model on ECO demonstrations and action preferences, comparing commercial, trained, and quantized models under shared execution rules. Table 1: ECO methods: models, training, and task scope. Method Model / training Task / scale IR-aware RL [7] (2024) R-GCN; DQN training Gate sizing: timing/power. 7 designs, 45 nm; up to 26k gates. RL-CTD [8] (2026) Regional net; RL training Regional ECO: timing/DRC. 16 designs, 14 nm; up to 232k gates. RSR [5] (2026) GPT-5-mini; RAG + reflection. No ECO fine-tuning reported. Timing/power/area. 8 designs, ASAP7; up to 33k μ 2. AgenticECO [6] (2026) Opus 4.8 / GPT-5.6; frozen LLMs + agents. 3D spacing DRC. 9 cases, ASAP7; up to 1.2k nets. WaveletECO (Ours) Qwen3.5-9B, local; SFT + CPO-SimPO. Setup / hold / electrical DRV. 22 designs; up to 202k cells. Five ECO rounds; three repeats per model/design. Sizes are rounded maxima in source-specific units. RSR also tests GPT-4o-mini and open-weight DeepSeek-V3.2. 2 Workflow and Model Training Each round analyzes timing, electrical, and library information (Figure 1). The model selects a repair and its targets; the shared workflow validates, executes, and measures it, retaining accepted changes and rolling back rejected ones. Commercial and local models use the same interface. Figure 1: WaveletECO combines a shared ECO workflow with training of a local repair model. Commercial and local models use the same action checks, physical measurements, and rollback mechanism. Data collection tests three proposals per state for up to five rounds. Training and validation data for SFT and CPO-SimPO [9][10] are split by physical design state. Offline preference labels compare same-state actions by execution success or recorded branch quality. Each response specifies one repair action. TorchAO [11] quantizes weights to INT8/INT4 with BF16 activations. 3 Experimental Results 3.1 Protocol and Repair Quality We evaluate nine policy variants on 22 designs, with up to five repair rounds per run. Three repeats per model and design yield 594 runs. The benchmark contains 13 designs used for local adaptation or model selection and nine held-out designs. All variants use the same EDA backend, one candidate per round, and a 1% area-growth limit. Let bd,kb_d,k and x denote the initial and terminal values of metric k for design d. Violation magnitude is vk(x)=max(0,−x)v_k(x)= (0,-x) for WNS/TNS and vk(x)=xv_k(x)=x for violation counts. Each metric receives qd,k(x)=100clip(1−vk(x)vk(bd,k),0,1),vk(bd,k)>0,100,vk(bd,k)=0,vk(x)=0,0,vk(bd,k)=0,vk(x)>0.q_d,k(x)= cases100\,clip\! (1- v_k(x)v_k(b_d,k),0,1 ),&v_k(b_d,k)>0,\\[4.0pt] 100,&v_k(b_d,k)=0,\ v_k(x)=0,\\ 0,&v_k(b_d,k)=0,\ v_k(x)>0. cases (1) The terminal ECO score equally weights setup, hold, and electrical repair: Sd,r=13[qWS+qTS+qNS3+qWH+qTH+qNH3+qD].S_d,r= 13 [ q_W_S+q_T_S+q_N_S3+ q_W_H+q_T_H+q_N_H3+q_D ]. (2) Here, W,T,NW,T,N denote WNS, TNS, and violation count; subscripts S/HS/H indicate setup/hold, and D denotes the deduplicated electrical violation count. Scores use unrounded measurements and are averaged over three repeats, then equally across designs. Table 2: Repair quality and estimated inference cost on 22 designs. ECO score (mean ± SD) All-22 cost / efficiency Model Seen Unseen All (22) USD/task C¯m C_m Score/USD EmE_m Relative RmR_m Initial state 51.28±0.0051.28± 0.00 62.96±0.0062.96± 0.00 56.06±0.0056.06± 0.00 — — — Qwen3.5-9B (direct)† 66.41±0.0066.41± 0.00 83.00±0.0083.00± 0.00 73.20±0.0073.20± 0.00 0.007 9,784.1 41.93×41.93× Qwen3.5-9B (thinking)† 66.71±0.2966.71± 0.29 80.08±0.6480.08± 0.64 72.18±0.4372.18± 0.43 0.088 820.5 3.52×3.52× WaveletECO-SFT† 70.92±0.0070.92± 0.00 86.49±0.0086.49± 0.00 77.29±0.0077.29± 0.00 0.005 14,339.0 61.45×61.45× WaveletECO-Policy (BF16) 74.44±0.3474.44± 0.34 87.13 ± 0.08 79.63±0.1779.63± 0.17 0.008 10,135.8 43.44×43.44× WaveletECO-Policy (INT8) 74.84 ± 0.00 86.58 ± 0.75 79.65 ± 0.31 0.002 35,208.6 150.89×150.89× WaveletECO-Policy (INT4) 73.43±0.0273.43± 0.02 86.24±0.0086.24± 0.00 78.67±0.0178.67± 0.01 0.001 53,085.4 227.50×227.50× GPT-5.6 Sol 70.36±0.1970.36± 0.19 84.52±0.7384.52± 0.73 76.15±0.3676.15± 0.36 0.161 471.7 2.02×2.02× GPT-6 Astra 71.84±0.1171.84± 0.11 85.54±0.1585.54± 0.15 77.44±0.0077.44± 0.00 0.332 233.3 1.00×1.00× DeepSeek V4 68.85±1.0668.85± 1.06 82.41±0.2982.41± 0.29 74.40±0.6574.40± 0.65 0.011 7,055.9 30.24×30.24× Notes: Scores report mean ± population SD over three repeats, with designs equally weighted within each repeat. SD denotes standard deviation. Seen comprises 13 designs used for adaptation or model selection; Unseen comprises 9 held-out designs. Cost and efficiency use all 22 designs. DeepSeek costs use the arithmetic mean of peak and off-peak prices. † Costs estimated using WaveletECO-Policy BF16 throughput. Bold marks column-best values and additionally highlights the INT8 row. Table 2 shows that WaveletECO-Policy (BF16) scores 79.63 across all 22 designs, compared with 77.29 for WaveletECO-SFT and 73.20 for Qwen3.5-9B (direct). Preference training improves the mean score over SFT by 2.34 points. GPT-6 Astra scores 77.44 under the same execution protocol. The INT8 and INT4 variants score 79.65 and 78.67, respectively, retaining most of the BF16 repair quality. On the nine held-out designs, WaveletECO-Policy (BF16) scores 87.13, compared with 86.49 for WaveletECO-SFT and 85.54 for GPT-6 Astra. These results support transfer to unseen designs within this benchmark. 3.2 Inference Cost and Cost Efficiency Mean cost over K=66K=66 tasks includes every request j, including corrections and rejected actions: C¯m=1K∑j∑bpm,bNm,j,b106,API,hm3600∑j(Nm,jinvm,jpre+(Nm,jout−1)+vm,jdec),local. C_m= 1K cases _j _b p_m,bN_m,j,b10^6,&API,\\[6.0pt] h_m3600 _j ( N^in_m,jv^pre_m,j+ (N^out_m,j-1)_+v^dec_m,j ),&local. cases (3) Here j indexes requests, b API billing categories, and N token counts; p is USD/million tokens, h USD/hour, and v tokens/s. Prefill includes the first output token, with (x)+=max(x,0)(x)_+= (x,0). For mean ECO score S¯m S_m, we define Em=S¯mC¯m,Rm=EmEGPT-6.E_m= S_m C_m, R_m= E_mE_GPT -6. (4) Here EmE_m is score per dollar and RmR_m uses GPT-6 as 1×1×. API costs use the fixed September 19, 2026 reference prices, including cache discounts and billed reasoning: high for GPT-5.6/GPT-6 [12], none for DeepSeek V4-Flash-0731 [13]; DeepSeek costs use the arithmetic mean of off-peak and peak prices. Local single-GPU vLLM costs use token counts and independent throughput measurements on RTX 6000D (BF16) and RTX 4090 (INT8/INT4), excluding EDA and idle time. Rental proxies are CNY 6.98/hour (6000D) and 1.88/hour (4090) [14], at 6.6996 CNY/USD. Figure 2: Repair quality versus estimated inference cost on 22 designs (66 tasks/model). Star: WaveletECO-Policy INT8, our highest-scoring local variant. DeepSeek: arithmetic mean of off-peak and peak prices; † : BF16 throughput proxies. WaveletECO-Policy INT8 scores 79.65 at USD 0.00226/task versus GPT-6’s 77.44 at USD 0.33188/task (Figure 2), yielding 150.89×150.89× the score per dollar. INT4 scores 78.67 at USD 0.00148/task, trading a small score decrease for lower cost. Repair scores are from the original evaluations; the deployments used for cost estimation have not been re-evaluated for repair quality. 4 Conclusion We present WaveletECO, which supports multi-round ECO decision-making through a closed-loop execution platform and improves the decision quality of a local 9B model through supervised fine-tuning and preference training. Experiments show that WaveletECO achieves higher overall repair scores than GPT-6 Astra across 22 designs. In particular, the INT8 variant achieves approximately 103% of GPT-6 Astra’s repair score at an estimated inference cost of about 1/147 of GPT-6 Astra’s. These results show that combining closed-loop execution with specialized model training enables locally deployed models to perform effective multi-round ECO repair at low cost. However, generalization to larger-scale and more diverse designs and technology nodes requires further validation. References [1] A. B. Kahng, J. Lienig, I. L. Markov, and H. Jin (2011) VLSI physical design: from graph partitioning to timing closure. Springer. Cited by: §1. [2] S. Sapatnekar (2004) Timing. Springer. Cited by: §1. [3] M. Liu, T. Ene, R. Kirby, et al. (2023) ChipNeMo: Domain-Adapted LLMs for Chip Design. arXiv preprint arXiv:2311.00176. Cited by: §1. [4] Z. He, H. Wu, X. Zhang, X. Yao, S. Zheng, H. Zheng, and B. Yu (2024) ChatEDA: A Large Language Model Powered Autonomous Agent for EDA. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. External Links: Document Cited by: §1. [5] Y. Ouyang, Y. Luo, D. Zuo, and Y. Ma (2026) Retrieve, Schedule, Reflect: LLM Agents for Chip QoR Optimization. Note: arXiv:2603.13767v2 Cited by: Table 1, §1. [6] S. Ren, Y. Han, L. Shen, Z. Jia, R. Fu, B. Yu, and T. Ho (2026) AgenticECO: An Agentic Framework for ECO on 3D Integrated Circuits. Note: arXiv:2608.03738v1 Cited by: Table 1, §1. [7] W. Jiang, V. A. Chhabria, and S. S. Sapatnekar (2024) IR-Aware ECO Timing Optimization Using Reinforcement Learning. In 2024 ACM/IEEE International Symposium on Machine Learning for CAD (MLCAD), External Links: Document Cited by: Table 1. [8] J. Xu, L. Wu, S. Wu, Y. Zhang, L. Ren, Y. Zuo, H. Yan, and L. Shi (2026) Towards Efficient ECO Automation: Timing and DRCs Concurrent Optimization Using Reinforcement Learning. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. Note: Accepted manuscript External Links: Document Cited by: Table 1. [9] H. Xu, A. Sharaf, Y. Chen, et al. (2024) Contrastive preference optimization: pushing the boundaries of LLM performance in machine translation. In International Conference on Machine Learning (ICML), Cited by: §2. [10] Y. Meng, M. Xia, and D. Chen (2024) SimPO: simple preference optimization with a reference-free reward. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §2. [11] PyTorch Contributors (2026) TorchAO: Native PyTorch Quantization and Sparsity. Note: Version 0.16.0 used for the reported experiments Cited by: §2. [12] OpenAI (2026) GPT-6 Astra and GPT-5.6 Sol: Model Reference Prices. Note: Reference rates recorded September 19, 2026 Cited by: §3.2. [13] DeepSeek (2026) DeepSeek V4: Historical Flash Pricing. Note: August 13 announcement; off-peak and peak reference-price scenarios Cited by: §3.2. [14] AutoDL (2026) GPU Rental Member Prices. Note: PRO 6000 96GB and RTX 4090 displayed member prices, recorded September 19, 2026 Cited by: §3.2.