Reasoning Is Learning to Budget: From Explicit Chain-of-Thought to Latent Reasoning and Adaptive Compute Allocation (Deep Dive)

Summary

Over the past year, “let the model think a little longer” has become almost the default way to improve reasoning. But a batch of papers that appeared in October 2026 paints a soberer picture: across 6 reasoning models and 4 benchmarks, on average 64.3% of thinking tokens come after accuracy has already peaked, and “thinking to the end” scores on average 9.03 percentage points lower than “thinking just enough” [1]. The question has therefore shifted from “can the model think” to “is the thinking worth it”.

This article follows the technical line through three questions: how much to think (how to measure the waste, and where the ceiling lies), when to think (how to hand the decision to think deeply to a cheap routing decision), and how to think (whether every step has to be written out as text). We also break down 5 core algorithms, with step-by-step procedures, notation and small worked examples, and close by discussing what this means for cost-sensitive, auditable settings such as financial risk control.

What the article covers

1. Background: why “thinking more” has become expensive
2. How much to think: measure the waste first
3. When to think: hand deep thinking to a cheap switch
4. How to think: move thinking into latent space
5. On the training side: making exploration itself cheaper
6. Side by side: trade-offs across eight approaches
7. Our view
8. Outlook
9. Practical implications
10. Closing remarks

References

1. [1] Weixin Guan, Liang Li, Shiyu Xie, et al. Constructing Pareto Frontiers of Optimal Reasoning Termination Methods for Overthinking in Reasoning Models. Institute of Information Engineering, Chinese Academy of Sciences, et al., 2026-10-09. https://arxiv.org/abs/2610.13570
2. [2] Taehyeon Kim, Hyunsoo Lee, Youngsoo Jang, Moontae Lee. Efficiently Learning To Reason or Not to Reason: Root-token Policy Optimization for Adaptive Thinking. LG AI Research, 2026-10-13. https://arxiv.org/abs/2610.17767
3. [3] Tien-Phat Nguyen, Truong Nguyen, Thanh-Thien Le, et al. Route-and-Generate: KL-Regularized Router Improvement for Reasoning Routing in LLMs. 2026-10-21. (abstract only) https://arxiv.org/abs/2610.30351
4. [4] Bo Yan, Qi Zhang, Zhen Tan, Tianlong Chen, Song Wang. PORA: Learning When More Reasoning Pays Off for LLM Agents. 2026-10-19. https://arxiv.org/abs/2610.27119
5. [5] Kutay Tire, Atula Tejaswi, Haris Vikalo, Sujay Sanghavi. INTACT: Inducing Latent Reasoning on a Frozen Backbone. UT Austin, 2026-10-14. https://arxiv.org/abs/2610.19568
6. [6] Gerard Grau García, Arnau Padrés Masdemont, Niccolò Grillo, et al. From Chain-of-Thought to Loops: Non-Autoregressive Latent Reasoning via Looped Transformers. Qualcomm AI Research et al., 2026-10-08. https://arxiv.org/abs/2610.11472
7. [7] Pratyay Dutta, Kowshik Thopalli, Vivek Narayanaswamy. COMPASS: Finding Where Reasoning Lives in Language Models. 2026-10-05. https://arxiv.org/abs/2610.07469
8. [8] Jingjie Qiu, Sheldon Yu, Xunyi Jiang, et al. Does Latent Chain-of-Thought Actually Reason? A Single, Unifying Taxonomy for Latent Reasoning. 2026-10-15. https://arxiv.org/abs/2610.21182
9. [9] Nghi Dao, Calder Morton-Ferguson, Noah Golowich, Adam Block. ReAPO: Retry-Aware Policy Optimization for Efficient Exploration in LLM Reasoning. 2026-10-20. https://arxiv.org/abs/2610.28770
10. [10] Wenhang Shi, Yiren Chen, Zhe Zhao, et al. Self-Evolution Reasoner: Continuous Optimization via Policy-Intrinsic Exploration and Exploitation. 2026-10-25. (abstract only) https://arxiv.org/abs/2610.34137
11. [11] Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, et al. Training Large Language Models to Reason in a Continuous Latent Space (COCONUT). 2024-12-09. (background) https://arxiv.org/abs/2412.06769
12. [12] Gongfan Fang, Xinyin Ma, Xinchao Wang. Thinkless: LLM Learns When to Think. 2025-05-19. (background) https://arxiv.org/abs/2505.13379

HX Research · Frontier AI Deep Dive · Great Unicorn Intelligence