💻 computer science

Verified Self-Improving Learning of Large Language Models for Multi-Objective Point-Spec Circuit Design

This paper introduces \methodfull, a simulation-driven self-improving framework that iteratively generates, verifies, and repairs circuit netlists using Pareto-Verified Preference Optimization and Safe Proximal Policy Optimization to achieve high-precision, multi-objective point-spec circuit design across diverse analog families.

Original authors: Juzheng Zhang, Kehao Zhang, Jinwen Liu, Yangfan Tang, Lijuan Li, Hongliang Liu, You Chen

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Juzheng Zhang, Kehao Zhang, Jinwen Liu, Yangfan Tang, Lijuan Li, Hongliang Liu, You Chen

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a brilliant but inexperienced apprentice how to design complex electronic circuits (like the power converters in your phone charger or the amplifiers in a speaker). The goal isn't just to make any circuit work; the goal is to hit a very specific target (e.g., "exactly 5 volts output") while also meeting strict engineering rules (like "must not overheat" and "must be efficient").

This paper introduces a new training method called MO-SIMI (Multi-Objective Self-Improving Model Iteration) that uses a Large Language Model (LLM) as that apprentice. Here is how it works, broken down into simple concepts:

1. The Problem: The "Guess and Check" Trap

Traditionally, designing these circuits is like trying to find a needle in a haystack by throwing random needles at the wall.

Old AI methods were like a student who reads a textbook (training data) and then guesses the answer. They often get the general idea but fail to hit the exact numbers required.
Other AI methods try to tweak the answer after guessing, but they often break the circuit in the process or get stuck in a loop of small, unhelpful changes.

2. The Solution: A "Coach, Simulator, and Safety Net" System

The authors created a training loop that acts like a rigorous coaching session with three distinct phases:

Phase A: The Simulation Lab (The "Test Drive")

The AI generates a circuit design (a "netlist," which is just a recipe for connecting electronic parts). Instead of just hoping it works, the system immediately runs it through a circuit simulator (a digital test drive).

The Result: The simulator says, "Pass," "Fail," or "Almost."
The Twist: If the design is almost right (e.g., it outputs 4.9V instead of 5.0V), the system doesn't throw it away. It applies a "Restricted Repair." Think of this as a mechanic who is only allowed to turn a few specific knobs (like adjusting a resistor value) but is forbidden from changing the engine type or the wiring layout. If the knob-turning fixes the problem, the "almost" becomes a "pass."

Phase B: The Preference Coach (Learning from Winners)

Once the system has a batch of designs, it compares them.

It doesn't just look at the "best" one. It uses a method called Pareto Ranking. Imagine a race where you care about both speed and fuel efficiency. A car that is slightly slower but much more efficient might be better than a fast, gas-guzzling one.
The system creates "preference pairs": "This design is better than that one because it met the target and was more efficient."
The AI is then retrained on these comparisons, learning to prefer the "winning" style of design over the "losing" style. This is the PVPO step.

Phase C: The Safety Net (The "Rollback" Button)

This is the most critical innovation. The AI tries to learn from these preferences and update itself. But, what if the update makes the AI worse?

In many AI systems, a bad update sticks, and the model forgets what it knew.
MO-SIMI uses a "Guard." After the AI tries to update, the system tests it again. If the new version fails to meet the standards (or if the "hit rate" drops), the system hits the Rollback button. It throws away the new update and goes back to the previous, safe version.
This ensures the AI never gets worse. It only moves forward, step by step, with every change verified to be an improvement.

3. The Results: From "Maybe" to "Master"

The paper tested this on three types of circuits:

DC-DC Converters (Power converters, like in chargers).
Amplifiers (For boosting signals).
Oscillators (For generating waves).

The Achievements:

High Precision: On the power converter tasks, the system hit the exact target specifications 97.8% of the time.
Complexity: It successfully designed circuits with 20+ components (which is considered very complex for AI).
Versatility: Once they taught the system how to design power converters, they simply swapped the "rulebook" (the verifier) and the same system could design amplifiers and oscillators with similar success.

The Big Picture Analogy

Imagine a chef trying to bake a cake that must taste exactly like a specific recipe, using only specific ingredients.

Old AI: The chef reads the recipe, guesses the amounts, bakes it, and if it's slightly off, they just try again from scratch.
MO-SIMI:
1. The chef bakes a cake.
2. A taste-tester (Simulator) says, "It's 90% right, but too sweet."
3. A "Restricted Repair" rule says, "You can only add a pinch of salt, you cannot change the flour." The chef adds salt, and it's perfect.
4. The chef compares this perfect cake to a bad one and learns why the first one was better.
5. Before the chef tries a new technique, a "Safety Manager" checks: "Did this new technique ruin your ability to make the perfect cake?" If yes, the chef goes back to the old technique. If no, they keep the new one.

Conclusion:
The paper claims that by combining simulation verification, smart repairs, preference learning, and a safety rollback, they created an AI that can reliably design complex, high-precision electronic circuits without human intervention, and it gets better at it every time it tries.

Technical Summary: Verified Self-Improving Learning of Large Language Models for Multi-Objective Point-Spec Circuit Design

Problem Statement
Analog circuit topology design remains a bottleneck in Electronic Design Automation (EDA) due to the exponential expansion of the design space and the sparsity of solutions that satisfy strict functional requirements alongside multiple performance constraints. While Large Language Models (LLMs) have been applied to circuit design, existing approaches face significant limitations:

Training-free methods (e.g., prompt engineering with external tools) struggle to stably hit strict point specifications under multiple hard constraints.
Training-based methods often lack scalability (limited to small circuits with <10 components), fail to generalize across diverse circuit families, and lack fine-grained multi-objective optimization mechanisms for specific point targets.
Current gaps include insufficient modeling of point specifications and a lack of mechanisms to iteratively improve generation capabilities using verified simulation data.

Methodology: MO-SIMI
The paper proposes Multi-Objective Self-Improving Model Iteration (MO-SIMI), a simulation-driven framework designed to iteratively refine an LLM's ability to generate point-spec circuit netlists. The framework operates on a supervised fine-tuned (SFT) model and proceeds through a unified, protected training loop:

Initialization: An initial reference model ( $\pi_0$ ) is obtained via supervised fine-tuning on a curated corpus of INC DSL netlists derived from Xyce open-source regression data.
Simulation-Validated Sampling: For a given task, the model generates candidate netlists which are evaluated by a unified NGSpice-based verifier. This yields binary decisions (OK, AR, CE) and continuous metrics.
Restricted Repair: To maximize the utility of near-miss candidates, the framework applies a "restricted repair" operator to simulatable failures that narrowly miss targets. This operator adjusts only pre-defined numerical knobs (e.g., passive R/L/C values or duty ratios) within a fixed scope, leaving topology and component types unchanged. Successful repairs convert failures into verified positives.
Preference Construction (PVPO): Within each task group, successful candidates are ranked using Pareto dominance and NSGA-II crowding distance. The framework constructs preference pairs $(x^+, x^-)$ where $x^+$ is the best verified candidate and $x^-$ is a weaker candidate or a repaired failure. These pairs train the model using Pareto-Verified Preference Optimization (PVPO), which distills multi-objective preferences from simulator outcomes.
Safe Reinforcement Learning (Safe-PPO): The model undergoes Proximal Policy Optimization (PPO) with a "guard rollback" mechanism. A hard-task pool is maintained for tasks with low recent attainment rates to focus simulation budget. Crucially, any policy update is only accepted if it passes an acceptance check on a fixed guard set; otherwise, the model is rolled back to the previous best version. This ensures monotonic improvement in goal attainment and engineering constraints.

Key Contributions

Verified Self-Evolution: The paper introduces a training paradigm where the model improves via its own verified design attempts. By expanding positive evidence through restricted repair and extracting supervision from task-local multi-objective ordering, the system creates a self-evolving loop for sparse point-spec analog design.
Engineering-Scale Generation: The framework successfully generates stable, multi-module DC–DC converters with 20+ components, a scale significantly larger than prior training-based methods (which typically handle <10 components).
Cross-Family Transferability: The workflow is demonstrated to transfer across distinct circuit families (DC–DC, Amplifiers, and Oscillators) by swapping only the task definitions and verifiers, maintaining strong performance without architectural changes.

Experimental Results
Evaluated on a Grid77 benchmark for DC–DC converters (1% tolerance) and tasks for amplifiers and oscillators (5% tolerance), MO-SIMI achieved the following:

DC–DC Performance: Reached an Attainment Rate (AR) of 0.9779 and an Engineering-Constraint Pass Rate (CE) of 0.8247. The model maintained an OK rate (simulation validity) of 1.0000.
Multi-Objective Capability: The system successfully generated complex multi-module circuits (voltage conversion + filtering + damping) while satisfying multiple simultaneous constraints (efficiency, ripple, overshoot).
Ablation Studies: Removing Pareto ranking reduced AR to 0.5442, and removing repair bootstrapping reduced AR to 0.6000, confirming that the combination of fine-grained preference ranking and repair is essential for the final performance.
Stability: The Safe-PPO with rollback mechanism ensured that the Attainment Rate on the acceptance set remained monotonic throughout training, preventing performance degradation from unstable updates.

Significance
The paper claims that MO-SIMI provides a reusable recipe for scaling simulation-validated self-improvement to diverse circuit families. By integrating restricted repair, Pareto-based preference learning, and acceptance-protected reinforcement learning, the framework addresses the critical gap of generating high-quality, multi-component analog circuits that meet strict point specifications. The work demonstrates that verified self-improvement loops can overcome the limitations of both prompt-based assistants and traditional search-based optimization in the context of complex analog design.