# Instruction-Following Eval (IFEval)

> A benchmark of 541 prompts built from 25 verifiable instruction types. It tests whether a model follows checkable constraints such as keyword, length, casing, and response-format requirements.

Canonical page: https://benchlm.ai/benchmarks/ifeval

- Category: [Instruction Following](/instruction-following)
- Last updated: September 15, 2026

## About IFEval

- Year: 2023
- Tasks: 541 prompts across 25 instruction types
- Format: Constrained generation
- Difficulty: Instruction precision
- Paper: [Instruction-Following Evaluation for Large Language Models](https://arxiv.org/abs/2311.07911)

IFEval applies deterministic checkers and reports four views of the same responses: prompt-level strict, instruction-level strict, prompt-level loose, and instruction-level loose. Prompt-level scoring requires every instruction in a prompt to pass; loose scoring applies limited normalization before checking.

IFEval is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (32 models)

| Rank | Model | Creator | Score |
|------|-------|---------|-------|
| 1 | [Qwen3.5-27B](/models/qwen3-5-27b) | Alibaba | 95% |
| 2 | [Agents-A1](/models/agents-a1) | InternScience | 94.8% |
| 3 | [Agents-A1-4B](/models/agents-a1-4b) | InternScience | 94.8% |
| 4 | [Qwen3.7 Plus](/models/qwen3-7-plus) | Alibaba | 94.6% |
| 5 | [Qwen3.7 Max](/models/qwen3-7-max) | Alibaba | 94.3% |
| 6 | [Qwen3.6 Plus](/models/qwen3-6-plus) | Alibaba | 94.3% |
| 7 | [dots3-note Preview](/models/dots3-note-preview) | Dots Studio | 93.9% |
| 8 | [Kimi K2.5](/models/kimi-k2-5) | Moonshot AI | 93.9% |
| 9 | [o3-mini](/models/o3-mini) | OpenAI | 93.9% |
| 10 | [Qwen3.5-122B-A10B](/models/qwen3-5-122b-a10b) | Alibaba | 93.4% |
| 11 | [GLM-5](/models/glm-5) | Z.AI | 92.6% |
| 12 | [Qwen3.5 397B](/models/qwen3-5-397b) | Alibaba | 92.6% |
| 13 | [K-EXAONE 2.0](/models/k-exaone-2-0) | LG AI Research | 92.4% |
| 14 | [o1](/models/o1) | OpenAI | 92.2% |
| 15 | [Qwen3.5-35B-A3B](/models/qwen3-5-35b-a3b) | Alibaba | 91.9% |
| 16 | [LFM2.5-8B-A1B](/models/lfm2-5-8b-a1b) | LiquidAI | 91.8% |
| 17 | [Claude Opus 4.5](/models/claude-opus-4-5) | Anthropic | 90.9% |
| 18 | [GPT-4.1 mini](/models/gpt-4-1-mini) | OpenAI | 88.5% |
| 19 | [GPT-4.1](/models/gpt-4-1) | OpenAI | 87.4% |
| 20 | [MiniCPM5-2B](/models/minicpm5-2b) | OpenBMB | 86.7% |
| 21 | [DeepSeek V3](/models/deepseek-v3) | DeepSeek | 86.1% |
| 22 | [ZAYA1-8B](/models/zaya1-8b) | Zyphra | 85.6% |
| 23 | [GPT-4.1 nano](/models/gpt-4-1-nano) | OpenAI | 83.2% |
| 24 | [LFM2.5-VL-3B](/models/lfm2-5-vl-3b) | LiquidAI | 82.3% |
| 25 | [Kanana-2 3B Instruct](/models/kanana-2-3b-instruct) | Kakao | 81.0% |
| 26 | [Celeris-1](/models/celeris-1) | Celeris | 80.8% |
| 27 | [MiniCPM5-1B](/models/minicpm5-1b) | OpenBMB | 80.4% |
| 28 | [Kanana-2 1.3B Instruct](/models/kanana-2-1-3b-instruct) | Kakao | 77.6% |
| 29 | [Mellum2-12B-A2.5B-Thinking](/models/mellum2-12b-a2-5b-thinking) | JetBrains | 76.5% |
| 30 | [Mellum2-12B-A2.5B-Instruct](/models/mellum2-12b-a2-5b-instruct) | JetBrains | 75.8% |
| 31 | [LFM2.5-230M](/models/lfm2-5-230m) | LiquidAI | 71.7% |
| 32 | [LFM2.5-VL-450M](/models/lfm2-5-vl-450m) | LiquidAI | 61.2% |

## FAQ

### What does IFEval measure?

IFEval measures whether a model follows checkable constraints in 541 prompts built from 25 instruction types. Examples include using required keywords, avoiding specified terms, meeting length limits, or producing a requested structure. Automatic checkers score each response, which makes the result reproducible for the constraints they can verify.

### Which IFEval score should I compare?

Match all four metric choices before comparing models: prompt-level or instruction-level, and strict or loose. Prompt-level scoring requires every instruction in a prompt to pass. Loose scoring applies limited normalization to reduce formatting-related failures. An unlabeled IFEval number is not enough evidence for a precise comparison.

### Does a high IFEval score mean a model is a good assistant?

No. A high IFEval score means the model obeyed more objectively checkable constraints under the selected metric. It does not show that the answer was factual, useful, well judged, stylistically appropriate, or reliable across a long conversation. Pair it with broader instruction tests and a workflow-specific prompt set.

## Compare Top Models on IFEval

- [Qwen3.5-27B vs Agents-A1](/compare/agents-a1-vs-qwen3-5-27b)
- [Agents-A1 vs Agents-A1-4B](/compare/agents-a1-vs-agents-a1-4b)
- [Agents-A1-4B vs Qwen3.7 Plus](/compare/agents-a1-4b-vs-qwen3-7-plus)
- [Qwen3.7 Plus vs Qwen3.7 Max](/compare/qwen3-7-max-vs-qwen3-7-plus)