acceptodds
Under review as a conference paper at ICLR 2027

ECOpilot: A Circuit-Aware Multimodal LLM Representation Learning Framework for Timing Reasoning and Optimization

Abstract

Digital-circuit timing depends on connectivity, cell behavior, and operating conditions, which are difficult to capture through text alone. We introduce ECOpilot, a multimodal representation learning framework that makes these complementary signals available to a pretrained large language model (LLM). A graph encoder and projection align circuit structure with a frozen language backbone through self-supervised masked gate modeling and tool-supervised arrival-time regression. A cell representation combines learned delay-surface features with logical effort, an analytical descriptor of cell driving behavior. Together with textual operating conditions, these representations support timing prediction and tool-guided optimization of cell replacements. On held-out circuits implemented in a TSMC 7 nm technology, ECOpilot achieves 8.6% mean absolute percentage error in arrival-time prediction. Downstream timing engineering change order (ECO) reduces worst-negative-slack magnitude by up to 85.9% and violating-path count by up to 96.0%, measured from per-corner averages across three test designs. Prediction, feature-ablation, and optimization results support combining explicit circuit structure with physical and logical cell knowledge.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
multimodal representation learning, circuit-aware large language models, Electronic Design Automation, multi-corner timing Engineering Change Order

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_4VMJOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ECOpilot, a framework for post-route timing ECO that connects a graph attention encoder over critical-path subgraphs to a frozen Llama 3.1 8B model through a trainable projection. The encoder and projection are pretrained with masked gate modeling (MGM) and arrival-time regression (ATR) on PrimeTime labels. They are then fine-tuned to regress the tool-measured delay change of a candidate cell replacement.

Separately, a CNN trained with contrastive, ranking and regression objectives embeds each standard cell’s NLDM delay table. The embedding is compressed with PCA, concatenated with logical-effort parameters, and stored in a database that is retrieved into the prompt as text.

A four-agent loop applies these components to gate sizing and threshold-voltage swapping. A Netlist Analyzer selects critical instances with a corner-weighted heuristic. A Circuit Optimizer proposes replacements. A Tool-Based Evaluator obtains PrimeTime estimate_eco feedback, and a Script Generator emits Innovus commands.

Experiments use 13 OpenCores designs implemented in TSMC 7 nm, with 8 for training and 5 held out. The paper reports 97.2% MGM accuracy and 8.6% ATR MAPE on held-out designs, and a reduction in cell-delay MAE from 3.57 to 1.36 ps with combined physical and logical cell features. Post-ECO results on three held-out designs at FF, TT and SS corners are compared with a greedy cell-sensitivity heuristic (CS) and with ECO-GNN.

Critical strengths

S1. The evaluation uses designs implemented in a commercial 7 nm flow.

S2. On held-out designs, the model degrades more gracefully on ATR than the Q-Former baseline.

Critical weaknesses

W2. Post-route timing ECO by gate sizing and Vt swapping is a standard capability of the commercial engines already used in this flow, such as PrimeTime fix_eco_timing and Innovus post-route optimization. No comparison with them is provided. The paper therefore does not establish a practical role for the method, for example repairing violations that remain after the tool converges, or reaching equal QoR with fewer cell changes or lower leakage. The near-complete repair achieved by the five-iteration greedy CS baseline, which reduces corner-average TNS by 97.6% to 98.9%, also suggests the evaluated violations are not challenging. Learned sizing methods the paper cites, TransSizer and LEGO-Size, are not compared either.

The related work omits several directly relevant lines of research. For circuit representation learning, CircuitFusion (ICLR 2025), NetTAG (DAC 2025) and GenEDA (2025) already combine netlist graphs with text or LLM latent spaces, use sub-circuit splitting at sequential boundaries and masked graph modeling, and align representations with implementation stages. For timing prediction, TimingPredict (DAC 2022) and PreRoutGNN (AAAI 2024) encode cell lookup tables inside timing GNNs. For learned sizing and ECO, RL-Sizer (DAC 2021) performs post-route timing sizing with multi-corner multi-mode information encoded by GNNs, and RL-LR-Sizer (MLCAD 2024), GNN-based Lagrangian-relaxation imitation (ICCAD 2022) and GNN-based Vth assignment methods address the same action space. For LLM agents, Retrieve-Schedule-Reflect (2026), the Timing Analysis Agent (2025) and ORFS-agent (MLCAD 2025) apply tool-using LLM agents to ECO and timing closure. In addition, TransSizer and LEGO-Size are cited but not compared, although both target post-route gate sizing and LEGO-Size casts sizing as language modeling over timing paths. Neither the ISPD 2012/2013 gate-sizing benchmarks nor a Lagrangian-relaxation sizer is used for evaluation. Given this body of work, the contribution of combining a graph encoder, a frozen LLM, cell features and an agentic loop for sizing and Vt swapping is incremental, and the paper should position itself against these methods empirically, not only in prose.

Questions for authors

Q1. ECO-GNN (Lu et al., 2023) is a signoff power predictor. How was it converted into a timing-ECO method that proposes replacements, and in what sense is TT its training corner?

Q2. How does ECOpilot compare with PrimeTime fix_eco_timing, or Innovus post-route optimization, restricted to the same sizing and Vt-swap actions, in QoR, number of changed cells, leakage and runtime? Can ECOpilot repair violations that remain after the commercial engine converges?

Q3. What are the output heads for MGM, ATR and Stage 2, and are they trained? What are the MGM, ATR and Stage 2 results for the same GAT, features and heads with the LLM removed?

Q4. Stage 2 is trained with an MSE loss on delay change, yet the optimizer selects the action with the highest model output probability while the LM head is frozen. How is this probability computed? How does the pipeline perform if the LLM proposal is replaced by direct ranking of legal candidates by predicted delay change, or by a random legal candidate filtered with PrimeTime feedback?

Q5. How many graph tokens represent a subgraph? What are the typical and maximum prompt lengths per agent, broken down into graph tokens, cell features, few-shot examples and trajectory, as a function of K and design size? Why does the optimizer also receive textual circuit graphs?

Q6. What is the downstream ECO quality when MGM, ATR or both are removed from pretraining? How does ATR compare with a sum-of-mean-cell-delays baseline, TimingPredict and PreRoutGNN?

Q7. How does the full pipeline perform with a more recent backbone? How does it compare with a strong text-only reasoning model given the same path descriptions, cell features, candidate sets and PrimeTime feedback, but no graph tokens?

Q8. How do you position ECOpilot relative to CircuitFusion, NetTAG and Retrieve-Schedule-Reflect?

Reviewer_HbKBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ECOpilot, a multimodal framework for timing reasoning and automated timing ECO. It addresses the challenge of jointly representing circuit topology, standard-cell delay behavior, and operating conditions for LLM-based optimization. A graph attention encoder is aligned with a frozen LLM through masked gate modeling and arrival-time regression, followed by adaptation to predict the timing impact of cell replacements. The framework also combines learned delay-surface features with logical-effort descriptors to represent standard cells. A tool-guided multi-agent workflow then selects critical instances, proposes legal replacements, evaluates timing effects, and generates ECO scripts across PVT corners. The main contributions include task-aligned circuit representation learning, domain-aware cell representations, and representation-guided multi-corner optimization. Experiments on held-out circuits demonstrate improved timing prediction and ECO performance against existing baselines, while ablation studies evaluate the contributions of key components.

Critical strengths
  1. Effective integration of multimodal circuit information for timing prediction. ECOpilot integrates circuit topology, learned physical delay features, and logical-effort information to improve timing prediction. The experimental results demonstrate strong performance against existing baselines, while the ablation studies support the complementary contributions of different feature types. These results highlight the practical value of incorporating circuit-specific knowledge into LLM-based timing reasoning.
  2. Practical validation through a realistic physical-design flow. The proposed framework is evaluated using a TSMC 7 nm technology, with physical ECO implementation performed in Cadence Innovus and final timing analysis conducted using Synopsys PrimeTime. Experiments across three held-out circuits and multiple PVT corners demonstrate improvements in total negative slack (TNS) and the number of violating paths compared with the conventional cell-sensitivity baseline. This provides practical evidence of the framework's effectiveness in downstream timing ECO optimization.
Critical weaknesses

1.Insufficient validation of the contribution of timing prediction to ECO optimization. Although the paper describes an action selection mechanism based on model output probabilities, the relationship between the timing prediction tasks (ATR and action-impact prediction) and the final ECO decisions is not sufficiently clarified. Moreover, the current ablation studies focus on the graph encoder and cell feature database, without isolating the contributions of ATR and action-impact fine-tuning. Consequently, it remains unclear to what extent the learned timing representations contribute to the reported ECO improvements. 2. Insufficiently established methodological novelty over closely related work. Several key components of ECOpilot, including graph-to-LLM alignment, masked modeling, arrival-time prediction, and CNN-based cell timing feature extraction, have been explored in prior studies such as BRIDGES and LEGO-Size. Although ECOpilot introduces domain-aware cell representations and a multi-corner agentic ECO workflow, the paper does not sufficiently distinguish its methodological contributions from these closely related approaches. More explicit comparisons and targeted ablation studies would help establish the novelty and incremental value of the proposed techniques.

Questions for authors
  1. How does the learned timing prediction guide ECO action selection? Section 3.2.3 introduces action-impact regression to predict delay changes, while Section 3.4.2 selects replacement actions based on model output probabilities. Could the authors clarify how the predicted delay changes are incorporated into action selection? Are these probabilities directly generated by the LLM or derived from the predicted timing improvements?
  2. What are the individual contributions of ATR and action-impact fine-tuning to downstream ECO performance? The current ablation studies evaluate the graph encoder and cell feature database, but do not isolate the effects of ATR pretraining and action-impact fine-tuning. Could the authors provide ablation results that remove these two training components separately while keeping the remaining optimization pipeline unchanged? This would help establish whether the learned timing capabilities are responsible for the reported ECO improvements.
  3. What are the key methodological advances over closely related work? Prior studies such as BRIDGES and LEGO-Size have explored graph-to-LLM alignment, timing-oriented pretraining, and CNN-based cell timing representations. Could the authors clarify the specific methodological differences between ECOpilot and these approaches, particularly regarding multimodal alignment and standard-cell feature learning? What evidence demonstrates the additional value of the proposed techniques?
  4. How are the baseline methods adapted and configured to ensure a fair comparison? Section 4.3 compares ECOpilot with ECO-GNN and Cell Sensitivity. Given that ECO-GNN originally focuses on power-related optimization, could the authors clarify how it was adapted or retrained for timing ECO? Additionally, were the training conditions, optimization objectives, number of evaluated candidate actions, and computational budgets comparable across the methods?
Reviewer_ukMpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ECOpilot combines circuit graph representations, physical cell features, and logical effort with a frozen LLM for timing prediction and tool-guided circuit optimization. Post-implementation results demonstrate some timing improvements, but the paper does not sufficiently establish the necessity of the LLM, the justification for its core decision mechanism, or whether the gains over simpler methods warrant the additional complexity.

Critical strengths

Evaluating physically implemented circuits across multiple operating corners provides stronger evidence of optimization effectiveness than prediction accuracy or local timing estimates alone.

Critical weaknesses
  1. The necessity of the LLM is not established. The GNN encodes structure, cell features supply physical knowledge, and STA tools validate outcomes; the paper does not demonstrate capabilities that a simpler predictor or search strategy could not provide.
  2. The training objective and decision rule are disconnected. Training regresses delay changes, whereas inference selects replacements using action generation probabilities, without a clear mathematical or implementation link; a higher generation probability does not imply a greater timing benefit.
  3. Multi-corner aggregation is not aligned with the optimization objective. Averaging action probabilities across corners is neither equivalent to optimizing worst-case timing nor sufficient to ensure that constraints are satisfied at every corner, leaving the core decision mechanism an insufficiently validated heuristic.
  4. A faster cell does not necessarily make the entire path faster. Increasing drive strength can also increase input capacitance, slowing upstream logic and affecting other fanout branches; the paper does not establish that its selection score, based on local delay contribution and cell optimization potential, reliably reflects the net path-level benefit.
  5. It is unclear whether cell-level representations preserve critical timing-arc information. Different input pins, rising and falling transitions, and conditional timing arcs of the same cell can have different delays; the paper does not explain how these distinctions are retained, making it difficult to assess whether the compressed features support accurate replacement decisions.
  6. The experiments do not establish comparative advantage or cost-effectiveness. In Table 3, the simple Cell Sensitivity baseline matches or outperforms ECOpilot on WNS in eight of nine design–corner pairs, while ECO-GNN is trained only at TT; without comparisons that match tool-call budgets and account for end-to-end costs, it remains unclear whether gains on other metrics justify the complex LLM workflow.
Questions for authors
  1. Why is an LLM necessary instead of a simpler predictor or search algorithm?
  2. How does delay regression translate into action probabilities, and why should averaging these probabilities across corners improve worst-case timing?
  3. How do cell-level features and local candidate selection preserve timing-arc distinctions and capture the path-level effects of input capacitance changes?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.