![[2602.14077] GTS: Inference-Time Scaling of Latent Reasoning with a Learnable Gaussian Thought Sampler](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.14077] GTS: Inference-Time Scaling of Latent Reasoning with a Learnable Gaussian Thought Sampler
Summary
The paper introduces the Gaussian Thought Sampler (GTS), a novel approach to inference-time scaling in latent reasoning models, enhancing trajectory diversity through structured perturbations rather than random noise.
Why It Matters
This research addresses inefficiencies in current latent reasoning models by proposing a learnable method for perturbation distribution, potentially leading to more effective AI systems. The findings could influence future developments in machine learning, particularly in enhancing model performance under constrained sampling conditions.
Key Takeaways
- GTS provides a structured alternative to traditional stochastic perturbation methods.
- The model predicts context-dependent perturbation distributions for improved reasoning.
- Experiments show GTS outperforms heuristic baselines in inference-time scaling.
- Improving latent reasoning requires optimizable exploration mechanisms.
- The research emphasizes the need for structured exploration over increased stochasticity.
Computer Science > Computation and Language arXiv:2602.14077 (cs) [Submitted on 15 Feb 2026] Title:GTS: Inference-Time Scaling of Latent Reasoning with a Learnable Gaussian Thought Sampler Authors:Minghan Wang, Ye Bai, Thuy-Trang Vu, Ehsan Shareghi, Gholamreza Haffari View a PDF of the paper titled GTS: Inference-Time Scaling of Latent Reasoning with a Learnable Gaussian Thought Sampler, by Minghan Wang and 4 other authors View PDF HTML (experimental) Abstract:Inference-time scaling (ITS) in latent reasoning models typically introduces stochasticity through heuristic perturbations, such as dropout or fixed Gaussian noise. While these methods increase trajectory diversity, their exploration behavior is not explicitly modeled and can be inefficient under finite sampling budgets. We observe that stronger perturbations do not necessarily translate into more effective candidate trajectories, as unguided noise may disrupt internal decision structure rather than steer it. To provide a more structured alternative, we model latent thought exploration as conditional sampling from learnable densities and instantiate this idea as a Gaussian Thought Sampler (GTS). GTS predicts context-dependent perturbation distributions over continuous reasoning states and is trained with GRPO-style policy optimization while keeping the backbone frozen. Experiments on GSM8K with two latent reasoning architectures show that GTS achieves more reliable inference-time scaling than heuristic baselines. The...
