IFHierBench: leaderboard
IFHierBench (2026) tests hierarchical instruction following: one prompt asks for a layered output whose artifact, sections and nested fields each carry verifiable constraints, at tree depths 0 to 3. Prompt-level strict accuracy in percent (every constraint satisfied); higher is better.
Metric: Prompt-level strict accuracy (%; share of prompts whose response satisfies every constraint in the tree; prompts whose output constraints are nested in a tree (artifact, sections, fields) of depth 0 to 3, checked by Python verifiers as in IFEval; 10 structure-level and 25 content-level constraint types). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 (High) | 42.8 | #102 (Claude Opus 4.6) |
| 2 | Gemini 3 Flash (Preview) | 33.2 | #144 |
| 3 | DeepSeek R1 | 15.5 | #473 |
| 4 | Kimi K2.5 | 14.7 | #256 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ifhierbench · How It Works · Data refreshed daily, snapshot 2026-11-13.