Analog integrated circuit (IC) design is a multi-turn, simulator-grounded engineering task that demands iterative reasoning over coupled performance tradeoffs, yet proprietary process design kits (PDKs) make hosted large language model workflows difficult to deploy in practical design settings. We present SLIC (Small LM Iterating on Circuits), a reinforcement fine-tuning framework that trains a 4B-scale, locally deployable language model for closed-loop analog circuit optimization. SLIC pairs structured circuit representations with discretized relative-update actions, domain- and success-gated multiobjective rewards, and Reward-Guided Rollout Steering (RGRS), which turns per-step simulator feedback into signals for both policy learning and online rollout construction in this sparse, highly constrained domain. We further release an open, goal-conditioned benchmark of 41 amplifier topologies with standardized protocols for in-distribution validation, held-out topology evaluation, and cross-technology-node transfer. Trained solely with open source 65 nm CMOS predictive technology models (PTM), SLIC-Random achieves 87.1% in-distribution four-specification Pass@1 within eight turns, surpassing the 763B DeepSeek-V4.1-Flash Think by 48.5 percentage points. Beyond the training distribution, SLIC-Softmax reaches 72.6% Pass@1 on held-out topologies, and the trained policies consistently transfer across five unseen technology nodes without additional fine-tuning. This transfer extends to a commercial TSMC 65 nm CMOS PDK, where SLIC-Random achieves 49.9% held-out-topology Pass@1, demonstrating practical deployment through local inference without transmitting proprietary data to external LLM services. Performance further scales with test-time simulator interactions, reaching up to 83.2% held-out-topology Pass@1 at a 32-turn budget.
Figures & tables
Figure 1: Overview of SLIC . Given a goal-conditioned analog circuit optimization task, SLIC builds a structured prompt from the circuit representation and target specifications, then trains a language-model agent through SFT and simulator-interactive RL. The upper path shows SFT data construction from filtered teacher rollouts. The lower loop shows multi-turn GRPO with Reward-Guided Rollout Steering (RGRS), where multiple candidate responses propose relative parameter updates that are evaluated through simulation, one eligible continuation is selected to advance the trajectory, and all candidates contribute to group-relative policy updates. Subscripts denote interaction steps, and superscripts index candidates within each group.
Figure 2: Technology transfer under four-specification Pass@1 evaluation within eight turns. Policies are trained on PTM 65 nm and evaluated across PTM nodes from 22 to 130 nm and a TSMC CMOS 65 nm PDK without additional fine-tuning. Initial parameters are adjusted to the valid ranges of each technology, while the PTM 65 nm target specifications are reused. Left: a 420-task subset of the full 840 in-distribution tasks. Right: 240 held-out topology tasks.
Figure 3: Test-time scaling of cumulative Pass@1 with up to 32 interaction turns on a 420-task subset of the full 840 ID tasks (left) and 240 OOD tasks (right). Qwen3.5-4B is evaluated in non-thinking mode; we also tested thinking mode with an 8k-token generation budget, but reasoning did not terminate within this limit. Numerical annotations highlight selected values along the best-performing curve in each panel.
Figure 4: Test-time scaling of Pass@ k with the number of sampled rollouts on a 420-task subset of the full 840 ID tasks (left) and 240 OOD tasks (right) with eight-turn simulator budget per rollout. Numerical annotations highlight selected values along the best-performing curve in each panel.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Technology
Channel length L
Channel width W
PTM 22 nm
22 nm–100 μ m
45 nm–500 μ m
PTM 32 nm
45 nm–100 μ m
45 nm–500 μ m
PTM 45 nm
45 nm–100 μ m
45 nm–500 μ m
PTM 65 nm (training) †
45 nm–100 μ m
45 nm–500 μ m
PTM 90 nm
56 nm–100 μ m
45 nm–500 μ m
PTM 130 nm
85 nm–100 μ m
45 nm–500 μ m
Appendix
Table 5: Technology-specific valid ranges for editable transistor parameters used in the parameter range check (Appendix B.3 , item (3)).
Table 6
Figure 5: Tolerance sweep of four-specification Pass@1 within eight turns on 840 in-distribution tasks (left) and 240 out-of-distribution tasks (right), with evaluation tolerances ranging from 0% to 15%. The vertical dotted line marks the default 5% tolerance. Trajectories are generated with strict targets and reevaluated at each tolerance. Numerical annotations highlight selected values along the best-performing curve in each panel.
Normalization
In-distribution
Out-of-distribution
Joint
32.2±0.8
24.9±1.7
Separate
41.8±0.5
28.7±1.1
Appendix
Table 12: Qwen2.5-7B-Instruct four-specification Pass@1 (%) at Turn 8.
Figure 6: Training dynamics of SLIC under different continuation strategies. (a) Four-specification Pass@1 on the validation set used for checkpoint selection; open circles mark the best validation checkpoint for each variant. (b) Mean generated response length per turn during training, including both reasoning and actions.
Figure 7: Illustrative response excerpts from two different training tasks. Softmax organizes its response around diagnosis, proposed edits, and expected effects. Greedy repeatedly revisits the decision to revert earlier changes. Bold text highlights these patterns, and brackets mark omissions.
Figure 8: Topology-level performance and task characteristics on Standard (left) and Hard (right). (a,b) Four-specification Pass@1 (%) on each held-out topology. Language models use up to eight turns, and BO uses 192 simulator calls per topology. SLIC checkpoints are trained on Standard unless labeled otherwise. Dashes denote unavailable results. (c,d) Circuit characteristics and task difficulty. For each specification gap, the main value is the median gap among tasks that initially miss the target. The smaller percentage gives the fraction of initially valid designs that miss it. Color scales are shared across datasets: one for Pass@1 and one per factor, with darker cells indicating larger values. Topology abbreviations: CG-CSL, common-gate amplifier with a current-source load; SE-2S, single-ended two-stage amplifier; CS-PCSL, common-source amplifier with a PMOS cascode current-source load; SE-TP, single-ended telescopic PMOS amplifier; SE-CCM, single-ended cascode current-mirror amplifier; FD-CM, fully differential current-mirror amplifier.
Figure 9: Spearman rank correlations between topology-level factors and mean four-specification Pass@1 within eight turns on Standard (left) and Hard (right). Each coefficient is computed across the six held-out topologies for the indicated SLIC variant. Gain, bandwidth, and power factors describe required improvements rather than raw performance values; larger values denote larger required changes. Blue and orange indicate positive and negative associations with success, respectively, on a shared scale centered at zero. Dashes denote unavailable results.
Figure 10: Feature-wise distribution shifts for Standard and Hard . Each column compares the indicated dataset with its reference training set using Wasserstein-1 distances computed after standardization using the same reference-training statistics for both compared sets. Power and bandwidth are log-transformed, and distances are expressed in reference-training standard-deviation units. Features describe initial conditions, specification targets, initial target gaps, circuit-size counts, and 15 expert-defined binary design descriptors. The descriptors encode architectural properties using fixed rules applied to circuit metadata and canonical topology names; multiple descriptors may apply to the same topology. For binary descriptors, distances reflect differences in attribute prevalence after normalization. Larger distances indicate stronger distribution shifts.
Category
Setting
Value
Hardware
Nodes / GPUs
1 node, 8 × H100 (80 GB)
Model
Base model
Qwen3.5-4B
Training strategy
FSDP2
Model load dtype
bfloat16
Gradient checkpointing
Enabled
Data
Dataset
Single-turn step samples
Appendix
Table 16: Main supervised fine-tuning (SFT) configuration.
Figure 11: Fully differential folded-cascode amplifier used in Case 1. X0 is the amplifier core and X1–X3 generate its bias voltages. Device identifiers and connections follow the supplied netlist. In the top-level instance, Vbiasn1 is tied to Vbiasn . Identical bias labels denote the same electrical net.
Figure 12: Fully differential current-mirror amplifier used in Case 2. X0 contains the amplifier and X1 supplies its shared bias voltage. Device identifiers and connections follow the supplied netlist. Identical bias labels denote the same electrical net. In both schematics, Vin+ / Vin− and Vout+ / Vout− correspond to Vinp / Vinn and Voutp / Voutn , respectively.
Circuit design for analog integrated circuits (ICs) is challenging because circuit topology, device sizing, and circuit performance are tightly coupled. Existing large language model (LLM)-based methods either rely on inference-time correction around frozen models or train on limited circuit types without directly using simulation outcomes as training signals. A key bottleneck is that existing datasets do not provide sufficient task information for training LLMs on end-to-end specification-conditioned design. To fill this gap, we first construct a large-scale simulation-grounded dataset with 8,626 design tasks across 10 different types of analog circuits, where each task includes a textual description, labeled I/O ports, loading condition, target specifications, a testbench, and a reference netlist. Using this dataset, we present SAFTAC, a simulation-augmented fine-tuning framework that trains LLMs to generate and revise analog circuit design. SAFTAC combines three-stage supervised fine-tuning, which progressively builds the capabilities required for analog circuit design and feedback-guided correction, with simulation feedback-augmented reinforcement fine-tuning, which further optimizes models using simulation-grounded rewards. We evaluate SAFTAC under single-pass generation and two-pass generation with simulation feedback, showing improvements over the strongest closed-source LLM by 15.9% and 7.8%, respectively.
Analog IC layout design remains a labor-intensive iterative process dominated by simulation-driven refinement. Although end-to-end layout generators accelerate initial placement and routing, they still require experts to manually tune layout optimization parameters with repeated post-layout simulations for stringent design specifications. While Bayesian Optimization (BO) is widely adopted for parameter tuning in analog IC design, at the layout level it typically requires hundreds to thousands of evaluations, each involving costly parasitic extraction and post-layout simulation, which makes it impractical. Recently, Large Language Models (LLMs) have demonstrated potential in improving the sample efficiency of such simulation-driven tuning. However, their restricted access to geometric layout context and design-specific heuristics limits their ability to manipulate the layout optimization process. In this paper, we propose a simulation-aware LLM multi-agent framework that performs in-context policy improvement (ICPI) by iteratively updating layout optimization parameters exposed by an analog layout generator through an act-observe-reflect loop on compact structured layout representations. Experiments on real-world analog circuits show that, with only tens of post-layout simulations, our approach improves post-layout performance over the generator's built-in heuristics and BO-based tuning method.
Bingyang Liu, Ziming Wei, Xiaohan Gao +1
Department of Electrical & Computer Engineering, The University of Texas at Austin, TX, USA
Recent advances in large language models (LLMs) suggest strong potential for automating analog circuit design. Yet most LLM-based approaches rely on a single-model loop of generation, diagnosis, and correction, which favors succinct summaries over domain-specific insight and suffers from context attrition that erases critical technical details. To address these limitations, we propose AnalogAgent, a training-free agentic framework that integrates an LLM-based multi-agent system (MAS) with self-evolving memory (SEM) for analog circuit design automation. AnalogAgent coordinates a Code Generator, Design Optimizer, and Knowledge Curator to distill execution feedback into an adaptive playbook in SEM and retrieve targeted guidance for subsequent generation, enabling cross-task transfer without additional expert feedback, databases, or libraries. Across established benchmarks, AnalogAgent achieves 92% Pass@1 with Gemini and 97.4% Pass@1 with GPT-5. Moreover, with compact models (e.g., Qwen-8B), it yields a +48.8% average Pass@1 gain across tasks and reaches 72.1% Pass@1 overall, indicating that AnalogAgent substantially strengthens open-weight models for high-quality analog circuit design automation.
Zhixuan Bao, Zhuoyi Lin, Jiageng Wang +5
Nanyang Technological University Singapore · Institute for Infocomm Research Singapore · Institute of Microelectronics Singapore +2