Dr-CiK - Codex Evidence: leaderboard
Context-aided time-series forecasting in Dr-CiK (Tang et al., McGill University, ServiceNow Research and Mila, 2026) with the evidence held fixed: every forecaster gets the numerical history plus the same evidence that the Codex agent (GPT-5.5, high reasoning effort) synthesized from a corpus of supporting and distractor documents, and direct-prompt LLMs sample 25 forecast trajectories per task. The score is scaled CRPS (probabilistic forecast error normalized by the mean absolute target, winsorized at 5 per task) averaged over the 225 tasks every forecaster completed; lower is better.
Metric: Scaled CRPS (mean over 225 tasks, per-task values winsorized at 5; 25 sampled trajectories; forecaster given Codex-synthesized evidence). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 0.3 | #803 |
| 2 | Qwen 3.5 4B | 0.31 | #1180 |
| 3 | Qwen 3.5 27B | 0.32 | #358 |
| 4 | Gemini 3.1 Flash Lite (Preview) | 0.33 | #413 |
| 5 | Mistral Medium 3.1 | 0.38 | #852 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=dr-cik-codex-evidence · How It Works · Data refreshed daily, snapshot 2026-11-12.