List your company

Early-stopping for mathematical reasoning on 5K corpus 1.0 (test)

67.4Savings RatioTTT QK (dh=128)
14.56828.2844255.716Apr 1, 2026

Evaluation results

MethodLinks
TTT QK (dh=128)Label mode=Supervised labels, delta=0.2, Backbone=Qwen2.5-32B, epsilon=0.052026.0467.420.4
TTT no-QKLabel mode=Supervised labels, delta=0.2, Backbone=Qwen2.5-32B, epsilon=0.052026.0467.319.2
TTT no-QKLabel mode=Consistent labels (no ground truth), delta=0.2, Backbone=Qwen2.5-32B, epsilon=0.052026.0464.419.3
TTT QK (dh=128)Label mode=Consistent labels (no ground truth), delta=0.2, Backbone=Qwen2.5-32B, epsilon=0.052026.0462.918.7
Static ProbeLabel mode=Supervised labels, delta=0.2, Backbone=Qwen2.5-32B, epsilon=0.052026.0462.520.8
TTT no-QKLabel mode=Supervised labels, delta=0.15, Backbone=Qwen2.5-32B, epsilon=0.052026.0457.515.2
Static ProbeLabel mode=Consistent labels (no ground truth), delta=0.2, Backbone=Qwen2.5-32B, epsilon=0.052026.0457.319.7
TTT QK (dh=128)Label mode=Supervised labels, delta=0.15, Backbone=Qwen2.5-32B, epsilon=0.052026.045615
TTT no-QKLabel mode=Consistent labels (no ground truth), delta=0.15, Backbone=Qwen2.5-32B, epsilon=0.052026.0452.914.1
TTT QK (dh=128)Label mode=Consistent labels (no ground truth), delta=0.15, Backbone=Qwen2.5-32B, epsilon=0.052026.0452.415
Static ProbeLabel mode=Supervised labels, delta=0.15, Backbone=Qwen2.5-32B, epsilon=0.052026.0451.215.9
Static ProbeLabel mode=Consistent labels (no ground truth), delta=0.15, Backbone=Qwen2.5-32B, epsilon=0.052026.0448.315.6
TTT no-QKLabel mode=Supervised labels, delta=0.1, Backbone=Qwen2.5-32B, epsilon=0.052026.0447.511
TTT QK (dh=128)Label mode=Supervised labels, delta=0.1, Backbone=Qwen2.5-32B, epsilon=0.052026.0441.410.3
TTT no-QKLabel mode=Consistent labels (no ground truth), delta=0.1, Backbone=Qwen2.5-32B, epsilon=0.052026.0440.79.6
TTT QK (dh=128)Label mode=Consistent labels (no ground truth), delta=0.1, Backbone=Qwen2.5-32B, epsilon=0.052026.0439.711.3
Static ProbeLabel mode=Supervised labels, delta=0.1, Backbone=Qwen2.5-32B, epsilon=0.052026.043810.5
Static ProbeLabel mode=Consistent labels (no ground truth), delta=0.1, Backbone=Qwen2.5-32B, epsilon=0.052026.0434.59.8
TTT no-QKLabel mode=Supervised labels, delta=0.05, Backbone=Qwen2.5-32B, epsilon=0.052026.0428.25.3
TTT QK (dh=128)Label mode=Supervised labels, delta=0.05, Backbone=Qwen2.5-32B, epsilon=0.052026.0423.34.6
TTT QK (dh=128)Label mode=Consistent labels (no ground truth), delta=0.05, Backbone=Qwen2.5-32B, epsilon=0.052026.0423.26.4
Static ProbeLabel mode=Supervised labels, delta=0.05, Backbone=Qwen2.5-32B, epsilon=0.052026.04225.5
TTT no-QKLabel mode=Consistent labels (no ground truth), delta=0.05, Backbone=Qwen2.5-32B, epsilon=0.052026.04224.5
Static ProbeLabel mode=Consistent labels (no ground truth), delta=0.05, Backbone=Qwen2.5-32B, epsilon=0.052026.0416.64.9