TASC-m - False Alarms: leaderboard
Time-sensitive safety awareness in authored conversations. TASC v1 uses one sample per authored conversation through OpenRouter, with provider-default temperature, top-p and reasoning effort. These are published observations; this intake makes no model calls. Condition A uses a duration parser. Conditions B and C use GPT-5.6 Luna as the judge of record, Claude Sonnet 5 as a second judge and Claude Opus 5 for adjudication. TASC-m Table 3 uses 3,397 shared inputs, with column denominators 234, 877, 1,347, 275 and 79. TASC-p has 33 complete scenario tuples per model, except DeepSeek-V4-Pro with 32. Non-violation catches are false alarms, so lower is better. Recovery is conditional on Condition B misses; the undefined Opus recovery remains missing. Exact API snapshots are unreported. Question access: Primary result reviewed; exact scored release availability not confirmed.
Metric: Published rate (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o mini (Benchmark Sweep Run: source=TASC; provider defaults; one sample; full authored history) | 0.07 | |
| 2 | GPT-5 mini (Benchmark Sweep Run: source=TASC; provider defaults; one sample; full authored history) | 0.37 | |
| 3 | MiniMax M3 (Benchmark Sweep Run: source=TASC; provider defaults; one sample; full authored history) | 0.67 | |
| 4 | GPT-5.5 (Benchmark Sweep Run: source=TASC; provider defaults; one sample; full authored history) | 0.97 | |
| 5 | Gemini 3.7 Flash (Benchmark Sweep Run: source=TASC; provider defaults; one sample; full authored history) | 1.11 | |
| 6 | Claude Haiku 4.5 (Benchmark Sweep Run: source=TASC; provider defaults; one sample; full authored history) | 1.41 | |
| 7 | Qwen3.8-Max (Benchmark Sweep Run: source=TASC; provider defaults; one sample; full authored history) | 1.48 | |
| 8 | DeepSeek-V4-Pro (Benchmark Sweep Run: source=TASC; provider defaults; one sample; full authored history) | 1.71 | |
| 9 | GPT-5.6 Terra (Benchmark Sweep Run: source=TASC; provider defaults; one sample; full authored history) | 2.08 | |
| 10 | GLM-5.2 (Benchmark Sweep Run: source=TASC; provider defaults; one sample; full authored history) | 2.08 |
Interactive version: theaggregate.ai/benchmark?slug=tasc-m-false-alarms · How It Works · Data refreshed daily, snapshot 2026-11-06.