AsymLoop - Five-shot - SIQA: leaderboard

AsymLoop (2026): research-trained Dense,Full LoopK2 andAsymLoopK3 at matched effective depth; scale labels follow paper, with1.68B/3B/7B/20Btraining tokens. LM Evaluation Harness Five-shot downstream accuracy; unrelated public checkpoints are not aliases. Accuracy (%).

Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 12 models tracked.

Top models

#ModelScoreOverall rank
1AsymLoop Study 1B AsymLoop K3 Base43.65
2AsymLoop Study 1B Full Loop K2 Base43.09
3AsymLoop Study 1B Dense Base42.48
4AsymLoop Study 350M AsymLoop K3 Base40.69
5AsymLoop Study 350M Full Loop K2 Base39.15
6AsymLoop Study 350M Dense Base39
7AsymLoop Study 135M Full Loop K2 Base38.33
8AsymLoop Study 135M AsymLoop K3 Base37.56
9AsymLoop Study 135M Dense Base36.44
10AsymLoop Study 60M AsymLoop K3 Base36.13

Interactive version: theaggregate.ai/benchmark?slug=asymloop-five-shot-siqa · How It Works · Data refreshed daily, snapshot 2026-10-22.