Escape Meridian

Zhang, Zhishuai, Toomey, Oscar, Jang, Taekwang · Nov 16, 2026

Zhishuai Zhang, Oscar Toomey and Taekwang Jang: Razavi-Bench

Markdown · JSON · Original source

Zhishuai Zhang, Oscar Toomey and Taekwang Jang propose Razavi-Bench to track how AI models answer questions about analog circuits. They built it from 50 questions in Razavi’s two Analog Design Experiments With AI articles and graded answers with against reference answers and rubrics. Across 73 released between May 2024 and November 2026, the top score rose from 28.83% to 98.33%. The authors say most of that increase came before the reference answers were published. They report that leading models now answer almost every textbook question correctly. The hardest questions, they say, involve circuits that resemble textbook examples but work differently, or behavior an obvious analysis misses. Their central point is that, within two and a half years, questions that exposed model limits became ones models answer almost perfectly. These results describe progress on this ; the abstract does not establish how well they reflect performance on other analog circuit problems.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

arXiv cs.AIArticle · arxiv.orgOpen original source

Original source

cs.AI · Article

Source details

Original source date: Nov 16, 2026 (2026-11-16).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-17T12:16:06.835Z. First observed: 2026-11-17T08:28:34.121Z. Retrieved: 2026-11-17T12:04:50.447Z.

Referenced by

Recorded citations in public readings. A reference alone does not establish agreement or influence.

No public references recorded here yet.

Reference data (JSON)

Limits

The is built from 50 questions drawn from two articles; the abstract does not establish how well its scores reflect performance on analog circuit problems outside that set. The reported answers were graded by against reference answers and rubrics; the abstract does not describe further grading validation.

Open question

How would the leading models perform on analog circuit questions outside the 50 questions in Razavi-Bench?

Source-stated

The authors propose Razavi-Bench, using 50 questions from two Analog Design Experiments With AI articles to track AI models’ analog circuit analysis.

We propose Razavi-Bench, a benchmark built from the 50 questions in Razavi's two Analog Design Experiments With AI articles

Source-stated

The benchmark answers were graded by GPT-6 Luna against reference answers and rubrics.

graded by GPT-6 Luna against reference answers and rubrics

Source-stated

Across 73 multimodal large language models released between May 2024 and November 2026, the top score rose from 28.83% to 98.33%.

Across 73 multimodal large language models released between May 2024 and November 2026, the top score rose from 28.83% to 98.33%

Source-stated

The authors say most of the score increase happened before reference answers were published.

mostly before the reference answers were published.

Source-stated

The authors describe the hardest questions as involving circuits that resemble textbook examples but work differently, or behavior missed by an obvious analysis.

the hardest questions show a circuit that resembles a textbook one but works differently, or ask about behavior the obvious analysis misses.

Source-stated

The authors characterize the change as a set of questions that once exposed AI model limits becoming questions models answer almost perfectly within two and a half years.

Within two and a half years, a set of questions that exposed the limits of AI models has become one they answer almost perfectly.

Source and contributor context

Zhang, Zhishuai, Toomey, Oscar, Jang, Taekwang · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Multimodal model

A model that works with more than one kind of input or output, such as text, images and audio.

Large language model

A language model with many learned numerical settings, used to process and generate language.

Benchmark

A defined set of tests used to compare systems. Results depend on the tasks and measures selected.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

GPT-6 Luna

GPT-6 Luna is an OpenAI model announced on September 22, 2026 as a cheaper GPT-6 tier.

Editorial provenance

Version 1. Publication ID: 82526d87-869e-4df3-88ac-558ecc9612e5.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-17T12:16:06.835Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 41c3c6b000cd4ad02ff6672d3e0b9b36310b8fdb67fa913a12069c2586ddf7b1. Source text hash: a27a1c6619bca99cf1399031adb017538d9096a9e8bc5bf0e7b971e64f209fad.

Revision and source-check history

Version 1 · 2026-11-17T12:16:06.835Z

Zhishuai Zhang, Oscar Toomey and Taekwang Jang: Razavi-Bench

Zhishuai Zhang, Oscar Toomey and Taekwang Jang propose Razavi-Bench to track how AI models answer questions about analog circuits. They built it from 50 questions in Razavi’s two Analog Design Experiments With AI articles and graded answers with against reference answers and rubrics. Across 73 released between May 2024 and November 2026, the top score rose from 28.83% to 98.33%. The authors say most of that increase came before the reference answers were published. They report that leading models now answer almost every textbook question correctly. The hardest questions, they say, involve circuits that resemble textbook examples but work differently, or behavior an obvious analysis misses. Their central point is that, within two and a half years, questions that exposed model limits became ones models answer almost perfectly. These results describe progress on this ; the abstract does not establish how well they reflect performance on other analog circuit problems.

Original source

cs.AI · Article

Source details

Original source date: Nov 16, 2026 (2026-11-16).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-17T12:16:06.835Z. First observed: 2026-11-17T08:28:34.121Z. Retrieved: 2026-11-17T12:04:50.447Z.

Limits

The is built from 50 questions drawn from two articles; the abstract does not establish how well its scores reflect performance on analog circuit problems outside that set. The reported answers were graded by against reference answers and rubrics; the abstract does not describe further grading validation.

Open question

How would the leading models perform on analog circuit questions outside the 50 questions in Razavi-Bench?

Source-stated

The authors propose Razavi-Bench, using 50 questions from two Analog Design Experiments With AI articles to track AI models’ analog circuit analysis.

We propose Razavi-Bench, a benchmark built from the 50 questions in Razavi's two Analog Design Experiments With AI articles

Source-stated

The benchmark answers were graded by GPT-6 Luna against reference answers and rubrics.

graded by GPT-6 Luna against reference answers and rubrics

Source-stated

Across 73 multimodal large language models released between May 2024 and November 2026, the top score rose from 28.83% to 98.33%.

Across 73 multimodal large language models released between May 2024 and November 2026, the top score rose from 28.83% to 98.33%

Source-stated

The authors say most of the score increase happened before reference answers were published.

mostly before the reference answers were published.

Source-stated

The authors describe the hardest questions as involving circuits that resemble textbook examples but work differently, or behavior missed by an obvious analysis.

the hardest questions show a circuit that resembles a textbook one but works differently, or ask about behavior the obvious analysis misses.

Source-stated

The authors characterize the change as a set of questions that once exposed AI model limits becoming questions models answer almost perfectly within two and a half years.

Within two and a half years, a set of questions that exposed the limits of AI models has become one they answer almost perfectly.

Source and contributor context

Zhang, Zhishuai, Toomey, Oscar, Jang, Taekwang · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Multimodal model

A model that works with more than one kind of input or output, such as text, images and audio.

Large language model

A language model with many learned numerical settings, used to process and generate language.

Benchmark

A defined set of tests used to compare systems. Results depend on the tasks and measures selected.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

GPT-6 Luna

GPT-6 Luna is an OpenAI model announced on September 22, 2026 as a cheaper GPT-6 tier.

Editorial provenance

Version 1. Publication ID: 82526d87-869e-4df3-88ac-558ecc9612e5.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-17T12:16:06.835Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 41c3c6b000cd4ad02ff6672d3e0b9b36310b8fdb67fa913a12069c2586ddf7b1. Source text hash: a27a1c6619bca99cf1399031adb017538d9096a9e8bc5bf0e7b971e64f209fad.

passed · 2026-11-17T12:16:06.835Z · anthropic/claude-opus-5.5

Publication: 82526d87-869e-4df3-88ac-558ecc9612e5

attribution — pass: The reading attributes the benchmark and its findings to the listed authors, Zhishuai Zhang, Oscar Toomey and Taekwang Jang. It keeps source claims attributed with phrasing such as 'they report' and 'the authors say'.

Authors:Zhishuai Zhang, Oscar Toomey, Taekwang Jang

summary — pass: The summary leads with the actors and the development, then gives the method, the score trend, the authors' note on timing and their central thesis. It keeps the authors' force without inflating it. A general reader can follow the main point: on these 50 questions, top AI scores rose from about 29% to 98%. Terms such as multimodal large language model and rubric are kept exact for tooltips. The question follows directly from the benchmark's limited scope and does not lead the reader.

Within two and a half years, a set of questions that exposed the limits of AI models has become one they answer almost perfectly.

date — pass: The displayed original publication date, 2026-11-16, matches the arXiv v1 submission date. The feed date of 2026-11-17 is later and was correctly not used.

[Submitted on 16 Nov 2026]

support — pass: Every claim matches the abstract verbatim in substance: 50 questions, grading by GPT-6 Luna, 73 models, the 28.83% to 98.33% rise, most gains before the reference answers were published, the nature of the hardest questions, and the two-and-a-half-year thesis. The limitations are framed as gaps in the abstract rather than as invented facts. The reading does not claim real-world reliability.

the top score rose from 28.83% to 98.33% in 30 months, mostly before the reference answers were published

media — pass: The selected media is the verified direct link to the original arXiv abstract page. That link is the most direct way for readers to reach the source.

Challenge or correct this reading

Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.