GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning
GradCuit improves reasoning by optimizing per-query latent states while model weights stay frozen. Its reported gains suggest a test-time scaling path beyond sampling and reranking.
GradCuit inserts optimizable latent states inside a frozen Transformer and routes continuation-wide reward gradients back to them. Across **five backbones**, three benchmarks, and two answer formats, it averaged **64.5% accuracy**.
Builders evaluating reasoning systems should compare latent optimization with extra sampling or reranking. The paper reports a **6.6-point gain over chain-of-thought prompting** and finds early-to-middle layers most effective for updates.
GradCuit inserts optimizable latent states inside a frozen Transformer and routes continuation-wide reward gradients back to them. Across **five backbones**, three benchmarks, and two answer formats, it averaged **64.5% accuracy**. Builders evaluating reasoning systems should compare latent optimization with extra sampling or reranking. The paper reports a **6.6-point gain over chain-of-thought prompting** and finds early-to-middle layers most effective for updates. These are research results, not evidence of production economics or broad task coverage. Optimization still requires outcome feedback, and the supplied material does not report latency, compute cost, or behavior outside the tested benchmarks.
This adds test-time latent-state optimization as a distinct reasoning lever: improve a frozen model using outcome gradients rather than prompts, extra samples, or weight training. The reported gain across several backbones supports testing where internal updates occur, especially early-to-middle layers, but missing latency and compute data prevents judging whether it is preferable to cheaper inference-time baselines.