{"id":"arxiv/2602.01148","url":"https://arxiv.org/abs/2602.01148","tags":["cs.AI","cs.IT","cs.LG","math.OC"],"score":0,"title":"Capabilities and Fundamental Limits of Latent Chain-of-Thought","domain":"cs.AI","source":"arxiv","authors":["Jiaxuan Zou","Yaozhong Xiong","Yong Liu"],"reviews":0,"subject":"cs.AI","summary":"Latent Chain-of-Thought (Latent CoT) models promise efficient reasoning via continuous representations, yet exhibit puzzling performance inconsistencies: excelling at exploration (ProsQA: 97.0%) but failing at computation (GSM8K: 34.1%). We reveal that this trade-off is governed by decisional certainty. Our contributions are threefold: (1) We theoretically characterize the fundamental Exploration-Execution Trade-off, proving that high certainty enables precise execution but inhibits exploration, while low certainty facilitates search but causes error accumulation. (2) We introduce the…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-02-01","reactions":{}}