Researchers have introduced GradCuit (gradient through circuit), a novel approach to test-time latent reasoning that directly optimizes intermediate hidden states in large language models (LLMs) without updating model parameters. The method, detailed in a recent paper on arXiv, addresses a key limitation of existing optimization-based latent reasoning techniques.
Traditional methods connect latent states to the reasoning trajectory through decoded tokens, making credit assignment indirect and obscuring how latent updates influence subsequent reasoning. GradCuit instead inserts optimizable latent states at a selected Transformer layer between the hidden representations of the prompt and the generated continuation. Thanks to causal self-attention, every continuation-token log-probability has a differentiable path to every preceding latent state, enabling reward-weighted gradients from the entire continuation to be assigned directly to the latents.
In experiments across five instruction-tuned backbones, three reasoning benchmarks, and two answer formats, GradCuit achieved an average accuracy of 64.5%, outperforming chain-of-thought prompting by 6.6 percentage points and the strongest competing method by 2.4 points. The method also demonstrated greater robustness: across seven learning-rate settings, it consistently outperformed LatentSeek while reducing the standard deviation of accuracy from 1.53 to 0.82. Even its random-walk variant remained competitive with LatentSeek.
For interpretability, token-level gradient attribution revealed that latent influence concentrates on reasoning-connector tokens, while layer analysis identified early-to-middle Transformer layers as the most effective optimization space. The authors argue that GradCuit opens a new axis of robust and interpretable test-time scaling, where LLMs adapt how they reason rather than merely regenerate, sample, or rerank outputs.
The code and project page are publicly available, and the paper has been recommended by the Semantic Scholar API alongside related works on latent reasoning.