AnalogLayoutWorld - Eligible Engineering Quality: leaderboard

AnalogLayoutWorld (arXiv 2611.16991v1, Table 1) evaluates analog circuit layout from fixed netlists, specifications and a SkyWater process design kit. There are 40 tasks, with three independent trials per task and a three-hour agent budget. Official grading starts after agent termination, outside that budget. Each native model, harness and reasoning setting remains distinct. GPT models use their API; other models run locally with SGLang. DRC, LVS, simulation and full acceptance use all task attempts as their denominator. Simulation passes only when all prescribed post-layout cases pass. Engineering quality is the eligibility-gated E_t Q_t value. The overall score is the native weighted task score, with submission, physical, geometry, morphology and electrical checks; it is not an unweighted mean of the displayed columns. Area, tokens and tool calls remain diagnostics. The unbound in-house Novara system is retained separately for identity review. Native eligibility-gated engineering quality, on a 0–1 scale. Higher is better.

Metric: Native mean score (0-1). Source: arxiv.org. Saturation forecast: Around June 2027. 12 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-6 AstraMax0.566#6
2GPT-6.1 SolMax0.498#10
3GLM-5.3Max0.358#68
4GLM-5.3 (Max) (Harness: InternBootcamp)0.35
5Kimi K3Max0.167#56
6GLM-5.3 FlashMax0.146#125
7Qwen3.8-Flash-NextxHigh0.14#117
8GLM-5.3 Flash (Max) (Harness: InternBootcamp)0.117
9DeepSeek V4.1 FlashMax0.086#62
10DeepSeek V4 ProMax0.049#191
11Kimi K2.6Max0#181
12GLM-5.2 (FP8)Max0

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=analoglayoutworld-eligible-engineering-quality · How It Works · Data refreshed daily, snapshot 2026-11-16.