GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning
GradCuit reports a 64.5% average accuracy by optimizing frozen-model latent states at test time.
The method inserts optimizable latent states inside a selected Transformer layer, giving continuation-token probabilities a differentiable path back to those states. The authors say this improves over chain-of-thought prompting by 6.6 points and the strongest competing method by 2.4 points across their tested backbones, benchmarks, and answer formats. They also report steadier performance across learning-rate settings than LatentSeek. Their attribution analysis finds the latent updates concentrate on reasoning-connector tokens, with early-to-middle layers working best. HF Daily Papers' note
The method inserts optimizable latent states inside a selected Transformer layer, giving continuation-token probabilities a differentiable path back to those states. The authors say this improves over chain-of-thought prompting by 6.6 points and the strongest competing method by 2.4 points across their tested backbones, benchmarks, and answer formats. They also report steadier performance across learning-rate settings than LatentSeek. Their attribution analysis finds the latent updates concentrate on reasoning-connector tokens, with early-to-middle layers working best. HF Daily Papers' note
score 5