Zhaoxin Yu, Qi Shen, Hengli Li, Zhaowei Zhang, Song-Chun Zhu, Chi Zhang, Zilong Zheng
4 min
Existing test-time reasoning methods often optimize latent variables by backpropagating through the final decoded tokens. This approach creates an information bottleneck, making it difficult to understand how specific latent updates influence the reasoning process. The authors ask: can we establish a more direct, interpretable, and robust gradient pathway between latent states and the model's internal reasoning trajectory?
GradCuit (Gradient through Circuit) modifies the inference process by inserting learnable latent variables into an intermediate Transformer layer. Unlike previous methods that treat the model as a black box, GradCuit leverages the Transformer's self-attention mechanism as a differentiable circuit. By placing latents directly into the hidden state space, the model's own attention pathways allow gradients from the entire generated sequence to flow backward directly to the latents. This enables iterative refinement of the reasoning process based on outcome feedback without updating the underlying model parameters.
GradCuit consistently outperforms standard Chain-of-Thought (CoT) prompting and existing latent-reasoning baselines across five different LLM backbones and three reasoning benchmarks (GPQA-Diamond, GSM8K, and MATH-500). On average, it achieves a 6.6 percentage point improvement over CoT and a 2.4 point lead over the strongest competing method. Beyond accuracy, GradCuit demonstrates superior robustness to variations in learning rates and optimization settings. Furthermore, gradient attribution analysis reveals that the model's latent updates are most sensitive to reasoning-connector tokens (e.g., "therefore," "because"), providing a window into how the model adapts its internal logic during inference.
This work shifts the paradigm of test-time scaling from simple output reranking or sampling to active adaptation of the model's internal reasoning process. By providing a direct, interpretable link between outcome feedback and internal latent states, GradCuit offers a more stable and transparent way to improve LLM reasoning performance.
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.
Alex: Does it actually produce better results than those older approaches?
Sam: The results are notable. It outperformed standard approaches by a meaningful margin, and — perhaps more importantly — it was consistent. It didn't just get lucky on certain problems. It reliably adapted its internal logic across different tasks, which is a harder bar to clear.
Alex: You mentioned earlier that this also makes the AI's reasoning more interpretable. What does that mean in practice?
Sam: It means we can look at where the correction signal concentrates — and see which parts of the AI's thinking were actually doing the work. The researchers found it focuses on what they call "reasoning-connector tokens." Those are the words that link logical steps together: "therefore," "because," "so." They're the hinges of an argument. GRADCUIT naturally gravitates toward them, which tells us something real about how the model is building its logic.
Alex: So instead of a black box, you get something closer to a visible map of the reasoning.
Sam: Precisely. We aren't just guessing at what the model did — we can trace the path it took. That matters a great deal for reliability and for catching errors before they reach the final answer.
Alex: Are there limitations worth noting?
Sam: There are. The system depends on having what's called a "reward model" — a separate tool that judges whether the AI's reasoning path is good or bad. If you don't have a clear way to measure success, GRADCUIT has nothing to steer toward. Also, choosing which layer to insert the nudges into matters quite a bit, and getting that right requires some trial and error. It's a precise tool, not a universal fix.
Alex: So it needs the right setup to work well — it's not simply a switch you flip.
Sam: That's a fair way to put it. What it does offer is a meaningful step toward AI models that don't just guess and hope — but actively refine their own logic in a way we can observe and understand. That's a direction worth paying attention to.
Alex: Thanks for listening to ResearchPod.