GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning
GradCuit routes outcome feedback directly into test-time latent states inside a frozen Transformer.
The paper says those latents sit between the prompt representation and generated continuation, giving continuation-token probabilities a differentiable path back through the model. Across five instruction-tuned backbones and three reasoning benchmarks, it reports 64.5% average accuracy, 6.6 points above chain-of-thought prompting and 2.4 above the strongest compared method. It also claims lower sensitivity across learning-rate settings than LatentSeek. Its attribution analysis finds the strongest latent influence around reasoning-connector tokens and early-to-middle layers. ArXiv · AI/CL/LG's note
The paper says those latents sit between the prompt representation and generated continuation, giving continuation-token probabilities a differentiable path back through the model. Across five instruction-tuned backbones and three reasoning benchmarks, it reports 64.5% average accuracy, 6.6 points above chain-of-thought prompting and 2.4 above the strongest compared method. It also claims lower sensitivity across learning-rate settings than LatentSeek. Its attribution analysis finds the strongest latent influence around reasoning-connector tokens and early-to-middle layers. ArXiv · AI/CL/LG's note
score 5