ARCHER Baselines - CoT-FSP - MATH: leaderboard

Direct runs on identical instances; shared normalization/scoring; one response/example; AIME complete set single-run exactmatch. CoT-FSP uses5released standardbackbones; Vanilla uses2released reasoningmodels. Fine-tuned CoT-FT/Toolformer/CoA/ARCHER rows require distinct trainedcheckpoint identity and are withheld. Qwen3-8B has thinking enabled. Metric: Final-answer accuracy (%). Source: https://arxiv.org/html/2611.18494v1

Metric: Final-answer accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 5 models tracked. Updated 17 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3 8BThinking71.07#1868
2Falcon-H1R-7B70.53#1672
3Qwen 3.5 4B67.88#1227
4Gemma 4 E2B62.82#4291
5Qwen 3 1.7B57.41#5807

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=archer-baselines-cot-fsp-math · How It Works · Data refreshed daily, snapshot 2026-11-17.