Small language models outperform 763B model at analog circuit design

A 4-billion-parameter model trained with reinforcement fine-tuning achieved 87% success on circuit specification tasks, surpassing a 763-billion-parameter model by 48.5 percentage points.

Big Tech

Yan Xu · Zhengqi Gao · Ching-Yun Ko · Zhang-Wei Hong · Wenjie Lu · Tao Yu · +3 more

MIT · Analog Devices, Inc. · IBM Research

Research Digest·12 November 2026·2 min read

Illustration: Small language models outperform 763B model at analog circuit design

In one line

A small locally-deployed model fine-tuned with reinforcement learning outperforms a massive cloud API model on multi-turn analog circuit optimization.

Before you read this

Three ideas this paper builds on

reinforcement learning with verifiable rewardsReinforcement learning that uses rewards from objectively checkable signals like correct answers or verification feedback.159 papers →agentic reinforcement learningTraining approach that lets LLM-based agents learn multi-step tool use and planning via reinforcement learning, often with dense or per-tool rewards.120 papers →agent-driven simulationAutonomous agents automatically generate, control, or extend simulations to improve accuracy, interactivity, or performance.39 papers →

Read it as

Designing analog circuits is a multi-step process relying on secret manufacturing data, which prevents using hosted AI tools that must access that data.

The SLIC framework trains a compact model to iteratively improve designs using feedback from a circuit simulator and reinforcement learning fine-tuning on open data.

The model solved 87.1 percent of in-distribution design tasks, surpassing a 763-billion-parameter cloud model while generalizing to new circuits and real secret data, all through entirely local inference.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

reinforcement learning with verifiable rewards agentic reinforcement learning agent-driven simulation on-policy distillation rollout verification test-time scaling