Label-free Bias-only TTRL for Test-Time Adaptation
September 27, 2026
Label-free bias-only TTRL achieves 76.67% accuracy on MATH-500 with Qwen2.5-7B by optimizing only ~100K bias parameters while keeping the backbone frozen. This approach uses majority-vote pseudolabels as rewards, optimizing 76,000x fewer parameters than full-parameter TTRL.
HOW THIS AFFECTS YOU
●
builderYou can implement extremely efficient test-time reasoning improvements with minimal compute overhead.
●
researcherThis demonstrates that effective test-time adaptation can emerge from very restricted optimization spaces.