AbstentionBench - false premise - FalseQA - Recall: leaderboard
AbstentionBench evaluates whether a model appropriately declines unanswerable or underspecified questions. This slice reports false premise - FalseQA - Recall as Abstention Recall (%); values converted from 0-1 to percent. Precision and recall must be read together: always abstaining is not necessarily useful. Scenario/dataset pairs and SFT/DPO/RL checkpoints remain separate; no synthetic overall average. Source: facebookresearch/AbstentionBench analysis/abstention_performance.csv.
Metric: Abstention Recall (%). Source: github.com. Saturation forecast: Estimated already saturated. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 1.5 Pro | 69 | #454 |
| 2 | O1 | 65.94 | #190 |
| 3 | Qwen 2.5 32B Instruct | 64.05 | #509 |
| 4 | Llama 3.3 70B Instruct | 60.7 | #541 |
| 5 | GPT-4o | 60.26 | #357 |
| 6 | Llama 3.1 405B Instruct | 56.62 | #470 |
| 7 | OLMo-7B-0724-Instruct-hf | 55.17 | #1441 |
| 8 | Llama 3.1 70B Instruct | 53.57 | #567 |
| 9 | Llama-3.1-Tulu-3-70B-DPO | 51.67 | #732 |
| 10 | Mistral 7B Instruct (v0.3) | 51.53 | #1255 |
| 11 | Llama-3.1-Tulu-3-70B | 51.24 | #499 |
| 12 | Llama 3.1 8B Instruct | 47.74 | #1017 |
| 13 | Llama-3.1-Tulu-3-8B-SFT | 43.23 | #977 |
| 14 | DeepSeek R1 Distill Llama 70B | 42.79 | #614 |
| 15 | Llama-3.1-Tulu-3-8B-DPO | 41.34 | #992 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=abstentionbench-false-premise-falseqa-recall · How It Works · Data refreshed daily, snapshot 2026-11-05.