arXiv paper: GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning

A new arXiv paper introduces GradCuit, a test-time method that optimizes continuous latent states inside a frozen language model and assigns outcome-based gradients directly to those states through the model’s remaining Transformer layers. In experiments across five instruction-tuned backbones, three reasoning benchmarks, and two answer formats, the authors report 64.5% average accuracy—6.6 percentage points above chain-of-thought prompting and 2.4 points above the strongest competing method. They also report lower sensitivity to learning-rate choices than LatentSeek and use token-level gradient attribution to identify where latent states most affect reasoning. The work matters because it proposes a way to adapt a model’s internal reasoning process at inference time while making the influence of those updates more directly analyzable, though the findings are reported in a single arXiv paper and have not been independently validated here.