Instruction Stacking Collapse - Math: leaderboard
The Stacking-Collapse benchmark (2026) places a random stack of instructions, such as valid JSON, a word limit, named sources or a fixed opening, in the system prompt and checks each one with a deterministic verifier. Score is the mean share of 20 stacked instructions satisfied while answering GSM8K math problems, in percent; higher is better.
Metric: Mean follow rate (%; share of the 20 stacked instructions a response satisfies on GSM8K math prompts, judged by per-instruction verifiers; stacks drawn uniformly at random from 22 deterministic instructions in six categories, 30 stacks x 10 items per size, temperature 0). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 59.1 | #161 |
| 2 | Gemini 2.5 Flash | 43.1 | #453 |
| 3 | GPT-5 Mini | 20.6 | #348 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=instruction-stacking-collapse-math · How It Works · Data refreshed daily, snapshot 2026-11-13.