DelistBench (Closed-Book): leaderboard
DelistBench (Yao et al., 2026) asks LLMs to reconstruct security-level delisting announcement records for 1,200 securities as of a historical cutoff, a database-assurance task, here without web access. A positive record counts as correct only with the right status, an announcement date within seven days and the broad reason, and a negative record needs a correct status. Overall joint accuracy in percent; higher is better.
Metric: Overall joint accuracy within 7 days (%; all 1,200 security-level delisting records: a Category A record needs the correct positive status, the announcement date within seven days and the broad reason, Category B and C records a correct negative status; invalid outputs and abstentions count as errors; closed-book, no web access). Source: arxiv.org. Saturation forecast: Around April 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.6 Flash | 70.6 | #72 |
| 2 | Gemini 3.5 Flash Lite (Medium) | 66.5 | #300 (Gemini 3.5 Flash Lite) |
| 3 | DeepSeek V4 Flash (Non-reasoning) | 62.6 | #254 (DeepSeek V4 Flash) |
| 4 | GPT-5.6 Luna (Low) | 57.2 | #117 (GPT-5.6 Luna) |
| 5 | DeepSeek V4 Pro (High) | 56.7 | #181 (DeepSeek V4 Pro) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=delistbench-closed-book · How It Works · Data refreshed daily, snapshot 2026-11-12.