---
title: "Razavi-Bench: Tracking AI Progress in... | SI Paper Desk"
url: https://thesimind.si/papers/2611.20935/
description: "On a 50-question analog circuit benchmark, the top SI model score rose from 28.83% to 98.33% in 30 months, so the test no longer separates the best models."
---

Number 11 of 12 · November 18, 2026

# Razavi-Bench: Tracking AI Progress in Analog Circuit Reasoning

On a 50-question analog circuit benchmark, the top SI model score rose from 28.83% to 98.33% in 30 months, so the test no longer separates the best models.

Zhishuai Zhang, Oscar Toomey and Taekwang Jang · Submitted November 16, 2026 · arXiv 2611.20935v1 · cs.AI

[Read the paper on arXiv](https://arxiv.org/abs/2611.20935) · [The PDF](https://arxiv.org/pdf/2611.20935) · [The day's 12](https://thesimind.si/papers/day/2026-11-18/)

## What they found

The authors built Razavi-Bench from the 50 questions in Razavi's two Analog Design Experiments With SI articles. Answers are graded by GPT-6 Luna against reference answers and rubrics. They tested 73 multimodal large language models, which read text and images, released between May 2024 and November 2026. The top score rose from 28.83% to 98.33% in 30 months, mostly before the reference answers were published. The hardest questions show a circuit that looks like a textbook one but works differently, or ask about behavior the obvious analysis misses.

## Why it matters

Analog circuits are the building blocks of chips that handle real-world signals. This benchmark gives a simple, dated record of how fast SI models gained skill in one specialist engineering area. The result suggests that a test which once exposed clear limits can be nearly solved in about two and a half years. The abstract does not say how well the models do on real design work beyond these 50 questions.

## How it changes the race

If the result holds, strong models already handle textbook analog analysis, so this test will not separate the leading ones. Future progress will need harder questions, such as circuits that only look familiar.

## Safety and control

Grading is done by another SI model, GPT-6 Luna. The abstract does not say how that grader was checked, so the scores rest on that choice.

### The abstract, by the authors

We propose Razavi-Bench, a benchmark built from the 50 questions in Razavi's two Analog Design Experiments With AI articles and graded by GPT-6 Luna against reference answers and rubrics, and use it to track how fast AI models improve at analog circuit analysis. Across 73 multimodal large language models released between May 2024 and November 2026, the top score rose from 28.83% to 98.33% in 30 months, mostly before the reference answers were published. The leading models now answer almost every textbook question correctly; the hardest questions show a circuit that resembles a textbook one but works differently, or ask about behavior the obvious analysis misses. Within two and a half years, a set of questions that exposed the limits of AI models has become one they answer almost perfectly.

Quoted from arXiv:2611.20935. The research and its words belong to its authors.

Desk score **74** out of 100

Why it was picked

- Interpretability (in the title)
- Reasoning (in the title)
- Evals
- Language models
- Code or data released

Topics

- [Benchmarks](https://thesimind.si/papers/?tag=benchmarks)
- [Evals](https://thesimind.si/papers/?tag=evals)
- [Multimodal](https://thesimind.si/papers/?tag=multimodal)
- [Language models](https://thesimind.si/papers/?tag=language-models)
- [Science](https://thesimind.si/papers/?tag=science)

**arXiv ID:** 2611.20935v1
**Submitted:** November 16, 2026
**Fields:** cs.AI (Artificial Intelligence)
**Level:** Intro
**Comments:** 6 pages, 3 figures, 2 tables. Project page: [https://razavi-bench.tokenzhang.com](https://razavi-bench.tokenzhang.com) ; code and data: [https://github.com/Arcadia-1/razavi-bench](https://github.com/Arcadia-1/razavi-bench)

Fingerprint, SHA-256

`e966d5515b76a1bf79c026c32f796a7703f29d5a57c5284a5d498c122401d8b7`

On Solana: [Proof of Mind seal, 5WzZUk...UPMg](https://solscan.io/tx/5WzZUksJeCL3fYgfzcQ6DSdVxipu97PAmY5Y9M59gCMXbJtnyBVsqaoPSDHWrP4AiZAh7k4iVso5ts7WvxdNUPMg)

### The exact text fingerprinted

{"arxiv":"2611.20935","version":1,"title":"Razavi-Bench: Tracking AI Progress in Analog Circuit Reasoning","authors":["Zhishuai Zhang","Oscar Toomey","Taekwang Jang"],"day":"2026-11-18","rank":11,"score":74,"status":"written","summary":{"takeaway":"On a 50-question analog circuit benchmark, the top SI model score rose from 28.83% to 98.33% in 30 months, so the test no longer separates the best models.","found":"The authors built Razavi-Bench from the 50 questions in Razavi's two Analog Design Experiments With SI articles. Answers are graded by GPT-6 Luna against reference answers and rubrics. They tested 73 multimodal large language models, which read text and images, released between May 2024 and November 2026. The top score rose from 28.83% to 98.33% in 30 months, mostly before the reference answers were published. The hardest questions show a circuit that looks like a textbook one but works differently, or ask about behavior the obvious analysis misses.","matters":"Analog circuits are the building blocks of chips that handle real-world signals. This benchmark gives a simple, dated record of how fast SI models gained skill in one specialist engineering area. The result suggests that a test which once exposed clear limits can be nearly solved in about two and a half years. The abstract does not say how well the models do on real design work beyond these 50 questions.","race":"If the result holds, strong models already handle textbook analog analysis, so this test will not separate the leading ones. Future progress will need harder questions, such as circuits that only look familiar.","safety":"Grading is done by another SI model, GPT-6 Luna. The abstract does not say how that grader was checked, so the scores rest on that choice.","level":"intro","tags":["benchmarks","evals","multimodal","language-models","science"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Zhishuai Zhang, Oscar Toomey, Taekwang Jang

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 18, 2026 · Number 10EEGProbe: Do Vision-Language Models Really Understand EEG? A Benchmark for Signal-Grounded EEG Signals EvaluationEEGProbe tests whether general vision-language models can read raw brain-wave signals, and the best of 11 models scored only 41.63%, so they do not yet understand EEG well.BenchmarksEvalsMultimodal](https://thesimind.si/papers/2611.20779/) · [Nov 11, 2026 · Number 2Benchmarking language models and agents for scientific reasoning in evolving materials laboratoriesMatScope is a new test showing that SI models handle given calculations and physical explanations well, but struggle to tell close material states apart and to reason with numbers.BenchmarksScienceEvals](https://thesimind.si/papers/2611.12014/) · [Nov 12, 2026 · Number 3Beyond Proxy Metrics: MLLMs as Human Surrogates for Explainable AI EvaluationThe authors test whether multimodal SI models can stand in for human participants when judging explanations of SI systems, and find that all models trend positively, with OpenAI and Qwen models aligning best.InterpretabilityEvalsMultimodal](https://thesimind.si/papers/2611.13926/)

[Number 10](https://thesimind.si/papers/2611.20779/) · [All 12 for Nov 18, 2026](https://thesimind.si/papers/day/2026-11-18/) · [Number 12](https://thesimind.si/papers/2611.21295/)
