Structure–performance coupled analog circuit reasoning with Large Language Models

Figures (9) Tables (7)

Figure 1.
Overview of the key manual steps in analog circuit design flow. The larger the area of a step block, the higher its difficulty and importance.
Figure 2.
Typical failures of LLMs in analog circuit design. Red marks core failure causes.
Figure 3.
The overview of ACR-Agent. The circuit analysis task is sequentially processed by the Choosing Agent and Pre-analysis Agent. It is sent to the Analysis Agent along with the selected concept and subcircuit entries, and the generated sequence of sub-tasks and function groups of circuit devices, ultimately generating LaTeX expressions. The circuits in Task and Subcircuit are all SPICE Netlists, which is simplified by circuit diagrams in the figure.
Figure 4.
The construction process of ACR-Bench. The data undergoes selection, conversion, and human analysis, and grouping to form ACR-Bench. The figure omits the depiction of human operators in the conversion, human analysis, and grouping.
Figure 5.
Representations of the circuit. The example amplifier in diagram can be represented by the SPICE Netlist.
Figure 6.
The performance of different LLMs on ACR-Bench. LLMs are ranked by their $ Pass@1 $.
Figure 7.
Results of manual error analysis.
Figure 8.
Four representative tasks of successful reasoning and errors. All circuits in tasks are SPICE Netlist, which is simplified by circuit diagrams in the figure.
Figure 9.
Four representative design of DeepSeek-V3 and ACR-Agent based on DeepSeek-V3 in Analogcoder with utilizing subcircuit tools. All circuits in their design are PySpice code, which is simplified by circuit diagrams in the figure.

Require: Analog Circuit Reasoning Task

$ T $

Ensure: Final Solution

$ E $

1: Agent 1: Choosing Agent

$ (A_{c}) $

$ L $

$ \gets $

$ L_{C} $

$ \cup $

$ L_{SC} $

　　　▷ Combine Concept Library

$ (L_{C}) $

with Subcircuit Library

$ (L_{SC}) $

$ (L_1,L_2,...L_n) \gets $

Separate(

$ L $

)

4: for each

$ L_i $

$ L_n $

$ s_i $

$ \gets $

$ A_{c(i)}(L_i,T) $

6: end for

$ S $

$ \gets $

$ A_{c(n+1)}(L,s_1,s_2,...,s_n,T) $

8: if

$ S \not= \sum s_i $

then

$ S' $

$ \gets $

$ A_{c(n+2)}(L,s_1,s_2,...,s_n,T,S) $

10: 　 while

$ S' \not= S $

11: 　　

$ S $

$ \gets $

Debate

$ (A_{c(n+1)}(L,s_1,...,s_n,T,S'),A_{c(n+2)}(L,s_1,...,s_n,T,S)) $

12: 　 end while

13: end if

14: Agent 2: Pre-analysis Agent

$ (A_{p}) $

15:

$ L \gets A_{p(split)}(S,T,C) $

　　　　　　　　　▷

$ C $

means Constraints

16:

$ P \gets A_{p(plan)}(S,T,C) $

17: Agent 3: Analysis Agent

$ (A_{a}) $

18:

$ E \gets A_{a}(S,P,L,T) $

return

$ E $

Table 1.

ACR-Agent Workflow.

Dataset	Analog circuits	SPICE	Data type	Task type	Expression
Masala-chai^[38]	✔	✔	Real	Generate	✘
Amsnet^[39]	✔	✔	Real	Generate	✘
Amsbench^[40]	✔	✘	Real	Reason + design	✘
Eee-bench^[41]	✔¹	✘	Real	Reason	✘
Analogcoder^[8]	✔	✔	Real	Design	✘
Analoggenie^[22]	✔	✔	Real	Design	✘
Autocircuit-rl^[19]	✔	✔	Synthetic	Design²	✘
LaMAGIC^[20]	✔	✔	Synthetic	Design²	✘
Chipnemo^[5]	✔¹	✘	Real	Reason + design	✘
CIRCUIT^[26]	✔	✔	Real	Reason	✘
ACR-Bench	✔	✔	Real	Reason	✔
¹ Mix of analog circuit and other topics. ² Containing only one type of typical analog circuit. Analog circuits: dataset theme on analog circuits; SPICE: whether analog‑circuit representations are SPICE netlists; Data type: source type of the dataset; Task type: target task types of the dataset; Expression: whether the output is in expression format. The bold text indicates that ACR‑Bench possesses all the characteristics listed in the table.

Table 1.

Comparison between ACR-Bench and other datasets.

Topic	Number
Transistor fundamentals	6
Basic amplifiers	16
Differential amplifiers	10
Active loads	8
Output stages	19
Feedback networks	18
Operational amplifiers	28
Frequency response and compensation	16
Circuit noise	4
Non-linear circuits	6
Total	131

Table 2.

Number of ACR-Bench problems in each topic.

Method	ACR-Bench			Simple single-step			Simple multi-step			Complex single-step			Complex multi-step
Method	Pass@1	Pass@5	R@5	Pass@1	Pass@5	R@5	Pass@1	Pass@5	R@5	Pass@1	Pass@5	R@5	Pass@1	Pass@5	R@5
GPT4.1	0.30	0.48	0.14	0.48	0.70	0.24	0.24	0.45	0.08	0.18	0.35	0.05	0.09	0.11	0.03
Analogxpert (GPT4.1)	0.30	0.51	0.12	0.44	0.68	0.21	0.27	0.48	0.04	0.23	0.45	0.05	0.09	0.17	0.05
SizingAgent (GPT4.1)	0.29	0.50	0.11	0.41	0.66	0.19	0.26	0.49	0.07	0.24	0.44	0.06	0.07	0.15	0.01
Analogcoder (GPT4.1)	0.28	0.45	0.14	0.43	0.63	0.26	0.23	0.43	0.05	0.16	0.34	0.05	0.10	0.13	0.06
Atelier (GPT4.1)	0.23	0.41	0.08	0.33	0.54	0.14	0.22	0.40	0.09	0.14	0.36	0.01	0.07	0.13	0.01
AMS-KG (GPT4.1)	0.30	0.49	0.13	0.44	0.67	0.21	0.27	0.46	0.11	0.18	0.40	0.05	0.10	0.17	0.02
ACR-Agent (GPT4.1)	0.37	0.56	0.18	0.60	0.81	0.34	0.27	0.41	0.08	0.22	0.47	0.07	0.12	0.20	0.02
DeepSeek-V3	0.36	0.53	0.19	0.54	0.71	0.34	0.24	0.44	0.13	0.25	0.52	0.06	0.15	0.23	0.06
Analogxpert (DeepSeek-V3)	0.32	0.51	0.18	0.47	0.69	0.28	0.22	0.36	0.13	0.27	0.53	0.11	0.11	0.21	0.05
SizingAgent (DeepSeek-V3)	0.34	0.56	0.16	0.53	0.79	0.29	0.20	0.42	0.05	0.26	0.52	0.11	0.10	0.19	0.03
Analogcoder (DeepSeek-V3)	0.34	0.52	0.19	0.52	0.73	0.34	0.27	0.47	0.13	0.20	0.39	0.07	0.11	0.20	0.02
Atelier (DeepSeek-V3)	0.31	0.50	0.18	0.46	0.67	0.30	0.25	0.44	0.14	0.23	0.45	0.10	0.07	0.18	0.00
AMS-KG (DeepSeek-V3)	0.37	0.53	0.21	0.55	0.71	0.36	0.26	0.38	0.14	0.31	0.59	0.13	0.10	0.22	0.01
ACR-Agent (DeepSeek-V3)	0.44	0.61	0.26	0.68	0.86	0.46	0.28	0.37	0.13	0.33	0.60	0.13	0.16	0.28	0.06
Pass@1: probability of success in one trial. Pass@5: probability of at least one success among five repeated trials. R@5: probability of success across all five repeated trials. The bold values indicate the best‑performance metrics of different methods built upon the same base model on ACR‑Bench and various tasks.

Table 3.

Main results. The performance of ACR-Agent on ACR-Bench.

Method	ACR-Bench			Simple Single-step			Simple Multi-step			Complex Single-step			Complex Multi-step
Method	Pass@1	Pass@5	R@5	Pass@1	Pass@5	R@5	Pass@1	Pass@5	R@5	Pass@1	Pass@5	R@5	Pass@1	Pass@5	R@5
DeepSeek-V3	0.36	0.53	0.19	0.54	0.71	0.34	0.24	0.44	0.13	0.25	0.52	0.06	0.15	0.23	0.06
Basic libraries	0.40	0.57	0.25	0.65	0.81	0.46	0.23	0.38	0.13	0.27	0.53	0.08	0.11	0.23	0.01
w/o Libraries	0.37	0.57	0.20	0.57	0.80	0.32	0.24	0.36	0.15	0.29	0.58	0.09	0.14	0.25	0.04
w/o Subcircuit Library	0.39	0.58	0.22	0.59	0.82	0.36	0.27	0.41	0.17	0.28	0.54	0.11	0.13	0.23	0.03
w/o Concept Library	0.41	0.60	0.21	0.64	0.86	0.39	0.27	0.43	0.12	0.32	0.53	0.10	0.10	0.22	0.00
w/o Choosing Agent	0.41	0.57	0.24	0.64	0.80	0.43	0.25	0.35	0.14	0.30	0.60	0.10	0.13	0.23	0.03
w/o Constraints	0.41	0.61	0.21	0.65	0.86	0.39	0.27	0.45	0.12	0.33	0.56	0.10	0.10	0.24	0.00
w/o Pre-analysis Agent (split)	0.42	0.58	0.24	0.65	0.84	0.41	0.30	0.45	0.16	0.25	0.48	0.11	0.14	0.23	0.03
w/o Pre-analysis Agent (plan)	0.42	0.60	0.24	0.66	0.86	0.43	0.25	0.35	0.13	0.32	0.56	0.17	0.15	0.28	0.02
ACR-Agent	0.44	0.61	0.26	0.68	0.86	0.46	0.28	0.37	0.13	0.33	0.60	0.13	0.16	0.28	0.06
Pass@1: probability of success in one trial. Pass@5: probability of at least one success among five repeated trials. R@5: probability of success across all five repeated trials. The bold values indicate the best‑performance metrics of different methods built upon the same base model on ACR‑Bench and various tasks.

Table 4.

Results of ablation study.

Method Input tokens Cache hit tokens Output tokens Cost ($ \text{\$} $)

DeepSeek-V3 614 158 530 0.0009
ACR-Agent 8,174 5,072 1,513 0.0033

Table 5.
Token consumption of ACR-Agent on ACR-Bench.

Method	Input tokens	Cache hit tokens	Output tokens	Cost ($ \text{\$} $)
DeepSeek-V3	614	158	530	0.0009
ACR-Agent	8,174	5,072	1,513	0.0033

Model	Metrics	Analogcoder (w/o tools)		Analogcoder
Model	Metrics	Base	ACR-Agent	Base	ACR-Agent
GPT4.1	Pass@1	0.57	0.59	0.69	0.72
GPT4.1	Pass@5	0.68	0.70	0.90	0.96
DeepSeek-V3	Pass@1	0.54	0.60	0.63	0.69
DeepSeek-V3	Pass@5	0.62	0.71	0.82	0.88

Table 6.

ACR-Agent application in circuit designing on the Analogcoder benchmark.