GradCuit: Credit-Assigned Gradient Flow for Robust Test-Time Latent Reasoning
Zhaoxin Yu · hf · 2026-08-04
The paper proposes GradCuit (gradient through circuit), an optimization method to enhance LLM test-time latent reasoning.
- Pain Point: Existing optimization-based latent reasoning methods typically connect states to the reasoning trajectory via decoded tokens, making sequence-level credit assignment indirect.
- Technical Innovation: GradCuit inserts optimizable latent states at a selected Transformer layer. Causal self-attention provides a differentiable path from subsequent token log-probabilities directly to the latent states, enabling reward-weighted gradients to be assigned directly.
- Results: Across multiple reasoning benchmarks, GradCuit achieves an average accuracy of 64.5%, outperforming chain-of-thought prompting by 6.6 points and the strongest competing method by 2.4 points. It also shows greater stability across learning rates, and token-level gradient attribution reveals that latent influence concentrates on reasoning-connector tokens.
More from Research
- Text Diffusion Models Show Latency and Throughput Gains for Longer Contexts — bodonoghue85 · 2026-08-04
- Text Diffusion Models Offer Latency and Throughput Gains at Medium Batch Sizes — bodonoghue85 · 2026-08-04
- DiffusionGemma Tech Report: Parallel Decoding Breaks LLM Inference Speed Limits — bodonoghue85 · 2026-08-04
- Peking University & Xiaohongshu Introduce BigMac: 1.9x Faster Multimodal Training with Stable Memory — jiqizhixin · 2026-08-04
- Multimodal Model Topo-Omni Precisely Identifies Functionally Selective Brain Regions — martin_schrimpf · 2026-08-04
- NeurIPS 2026 Workshop to Explore Dynamic Alignment in Human-AI Coupled Systems — StanfordAILab · 2026-08-04