KAN Author Ziming Liu Releases Architecture IQ Benchmark: LLM Training Intuition Still Lags Humans

Ziming Liu, an MIT researcher and first author of the KAN paper, released the Architecture IQ benchmark on October 23–24 and posted a thread of tweets laying out his argument: LLM "training intuition" still has fundamental limitations, and humans can win in certain scenarios — he noted that Kaiming He designed ResNet on intuition alone, without the million-token chains of thought today's LLMs rely on. He further argued that the "language" of AI research has yet to be invented, much like an LLM trained by Aristotle could never learn modern math and physics that didn't exist yet; today's training techniques resemble alchemy, and LLMs learn correlations rather than mechanisms.

Confirmed

• Benchmark design: 500 problems in total (per MetaCircle's release); each problem provides a synthetic dataset and three training recipes (model/optimizer/hyperparameter combinations), and the task is to predict which combination achieves the lowest test loss. Answers are verified by real training runs, and tasks span regression, classification, and next-token prediction.
• Key results: GPT-5.6 Sol achieved 76.4% overall accuracy (above chance); LLMs ranked optimizers fairly well (85–87%), but were only 38–61% accurate at choosing architectures and were insensitive to data characteristics. Compressing "training physics" intuitions (e.g., cumulative effective learning rate U, roughly ηT for Adam and 2η√T for Adagrad, NTK curvature, zero-momentum SGD failures, etc.) into a decision-tree strategy yielded 78.2% out-of-bag accuracy.
• Empirical knowledge base experiment: using Claude Opus 5 to build an evidence-weighted knowledge base of up to 20 distilled principles, without any weight updates, GPT-4o's accuracy on 50 held-out problems rose from 53.2% to 67.6%.
• Four key limitations: imperfect intuition (humans outperform in some scenarios), lack of structure, insufficient compression, and insensitivity to data.

Why it matters

• The work provides quantitative evidence on whether LLMs can autonomously do AI research: intuition is relatively strong at the optimizer level but clearly lacking at the architecture-selection level, and experience can be distilled and reused at low cost — suggesting human experience and structured knowledge bases remain key levers for improving models' research capabilities.

2026-10-23 ~ 2026-10-24 · 11 related posts

• [source] ArchitectureIQ Benchmarks LLMs' Training Intuition: GPT-5.6 Sol Hits 76.4% vs 33% Chance — ZimingLiu11 · 2026-10-23
• Architecture IQ benchmark shows LLMs still lag humans in training intuition, ResNet-style — ZimingLiu11 · 2026-10-24
• Architecture IQ: measuring LLMs' training intuition, top models still fall short — ZimingLiu11 · 2026-10-24
• New benchmark tests LLMs' training intuition: predict which optimizer setup wins — ZimingLiu11 · 2026-10-24
• [source] LLMs Rank Optimizers at 85% but Are Data-Blind on Architectures, New Benchmark Shows — ZimingLiu11 · 2026-10-24
• Training-Physics Intuition Compressed Into a Decision Tree Hits 78.2% Out-of-Fold — ZimingLiu11 · 2026-10-24
• [source] MIT's Ziming Liu: Evidence-Weighted Knowledge Base Lifts GPT-4o From 53% to 68% — ZimingLiu11 · 2026-10-24
• 20 text principles lift GPT-4o accuracy 53.2%→67.6% on training-intuition benchmark — ZimingLiu11 · 2026-10-24
• KAN Author Ziming Liu: LLM Training Intuition Has Four Limits, 'Science of AI' Language Doesn't Exist Yet — ZimingLiu11 · 2026-10-24
• KAN author Ziming Liu: the "language of AI science" hasn't been invented, so LLMs can't learn it — ZimingLiu11 · 2026-10-24
• KAN author Ziming Liu: LLMs learn alchemy, not the science of AI — ZimingLiu11 · 2026-10-24