Search papers, labs, and topics across Lattice.
This paper introduces GradCuit, a novel approach that enhances latent reasoning in large language models by optimizing instance-specific continuous states at test time while keeping model parameters fixed. By inserting optimizable latent states within a Transformer layer and leveraging causal self-attention, GradCuit enables direct credit assignment from the output to the latent states, leading to improved accuracy and robustness. The method outperforms existing techniques, achieving an average accuracy of 64.5% across multiple benchmarks and demonstrating significant interpretability through token-level gradient attribution.
Directly optimizing internal reasoning from outcome feedback, GradCuit achieves a 6.6% accuracy boost over chain-of-thought prompting while enhancing robustness and interpretability.
Optimization-based latent reasoning improves large language model outputs by optimizing instance-specific continuous states at test time while keeping model parameters frozen. Existing methods, however, typically connect these states to the reasoning trajectory through decoded tokens, making sequence-level credit assignment indirect and obscuring how latent updates shape subsequent reasoning. We introduce GradCuit (gradient through circuit), which inserts optimizable latent states at a selected Transformer layer between the hidden representations of the prompt and the generated continuation. Causal self-attention provides every continuation-token log-probability with a differentiable path to every preceding latent state through the remaining Transformer blocks, enabling reward-weighted gradients from the entire continuation to be assigned directly to the latents. Across five instruction-tuned backbones, three reasoning benchmarks, and two answer formats, GradCuit achieves 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. GradCuit also demonstrates greater robustness: across seven learning-rate settings, it consistently outperforms LatentSeek while reducing the standard deviation of accuracy from 1.53 to 0.82, and even its random-walk variant remains competitive with LatentSeek. For interpretability, token-level gradient attribution reveals that latent influence concentrates on reasoning-connector tokens, while layer analysis identifies early-to-middle Transformer layers as the most effective optimization space. By directly optimizing internal reasoning from outcome feedback, 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.
