SABER-Math - Combinatorics: leaderboard
SABER-Math (2026) tests retrieval models on finding mathematically related competition problems: each query problem reranks a pool of 150 candidate problems whose graded relevance comes from pairwise LLM judgments. This board covers the combinatorics queries, with the query given as a problem statement and each candidate as a problem with its solution. Scored as nDCG@10 on a 0 to 100 scale; higher is better.
Metric: nDCG@10 (0-100) with exponential gain, printed 0 to 1 and scaled by 100: each competition or olympiad problem query reranks its own pool of 150 candidate problems, graded 0 to 5 by Bradley-Terry fits of GPT-OSS-120B pairwise relevance judgments over a 20-round Swiss tournament; every system at its default configuration and task instruction, statement-full setting (the query is a problem statement, each document a problem with its solution), the queries of the combinatorics domain; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 43 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | ReasonReranker-Qwen3-32B | 77.7 | |
| 2 | ReasonEmbed-Qwen3-8B | 73.8 | |
| 3 | Diver-GroupRank-32B | 72.9 | |
| 4 | RaDeR-7B | 71.5 | |
| 5 | RaDeR-14B | 71.4 | |
| 6 | Qwen3-Reranker-4B | 71.2 | |
| 7 | Diver-4B | 70.8 | |
| 8 | Qwen3-Reranker-8B | 69.7 | |
| 9 | SPLADE-Code-8B | 69.4 | |
| 10 | RaDeR-3B | 68 |
Interactive version: theaggregate.ai/benchmark?slug=saber-math-combinatorics · How It Works · Data refreshed daily, snapshot 2026-11-13.