TriggerBench - Negative Clean PM Accuracy (40K Context): leaderboard
TriggerBench (2026) tests prospective memory in conversation: a user states a constraint early in a dialogue, and a later turn creates a situation where a good assistant should act on it unprompted; a judge model scores whether each reply does. Dialogues are stretched to 40K tokens with unrelated turns between the constraint and the trigger. The board is accuracy in percent on control tasks where the context makes the reminder unnecessary, so staying silent is correct; higher is better.
Metric: PM accuracy (%) on the 289 negative controls, where the context makes the reminder unnecessary and restraint is correct; dialogues stretched to 40K tokens with unrelated turns injected between constraint and trigger; TriggerBench: 1,265 prospective-memory tasks from 488 dialogue blueprints in five dimensions (state tracking, temporal grounding, logical adherence, attention recovery, safe coding); a constraint is stated early in a dialogue and a later trigger turn calls for acting on it without being asked; GPT-4o (T=0) judges whether the reply fulfils the proactive intent; temperature 0.6; higher is better. Source: arxiv.org. Saturation forecast: Around August 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B 2507 Instruct FP8 | 82.7 | #379 |
| 2 | GPT-4.1 | 72.32 | #495 |
| 3 | Qwen 3 235B A22B 2507 FP8 (Thinking) | 67.48 | #434 |
| 4 | GPT-5.2 (High) | 39.8 | #176 (GPT-5.2) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=triggerbench-negative-clean-pm-accuracy-40k-context · How It Works · Data refreshed daily, snapshot 2026-11-13.