PACT - Steerability: leaderboard
PACT enterprise rule compliance under pressure, using three repeated trials at temperature 1. PACTScore weights initial compliance 0.75 and follow-up compliance 0.25; six diagnostic axes remain separate. Explicit instruct variants are marked as such and Inkling is included.
Metric: Score (0-1). Source: arxiv.org. Saturation forecast: Not forecast. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4.3 | 0.56 | #204 |
| 2 | Qwen 3.6 27B | 0.55 | #182 |
| 3 | DeepSeek V4 Pro | 0.55 | #112 |
| 4 | Claude Haiku 4.5 | 0.53 | #287 |
| 5 | Qwen 3.5 35B A3B | 0.52 | #268 |
| 6 | Gemini 3 Flash | 0.51 | #103 |
| 7 | Nemotron 3 Super | 0.49 | #344 |
| 8 | Kimi K2.7 Code | 0.45 | #95 |
| 9 | Gemma 4 26B | 0.45 | #322 |
| 10 | MiniMax-M2.5 | 0.45 | #316 |
| 11 | Kimi K2.6 | 0.44 | #113 |
| 12 | Llama 3.1 8B Instruct | 0.4 | #1017 |
| 13 | GPT-OSS-120B | 0.4 | #346 |
| 14 | Seed-OSS-36B-Instruct | 0.39 | #384 |
| 15 | GLM-4.7 | 0.38 | #202 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=pact-steerability · How It Works · Data refreshed daily, snapshot 2026-11-05.