arXiv:2607. 06648v1 Announce Type: new Abstract: Latent reasoning methods perform multi-step inference entirely in the model's continuous hidden states, promising more compact and efficient reasoning.
By Hengyu Jin, Shu Yang, Di Wang
arXiv:2607. 16097v1 Announce Type: cross Abstract: Reinforcement learning (RL) has become central to improving large language models (LLMs) on complex reasoning tasks, yet RL post-training is largely studied in isolation from the pretraining that precedes it.
By Jingyan Shen, Ang Li, Salman Rahman, Yifan Sun, Micah Goldblum, Matus Telgarsky, Pavel Izmailov
Latent Recurrent Thoughts (LRT) proposes a method for reasoning with frozen large language models by operating in the model’s continuous representation space. A small auxiliary network generates initial latent vectors, which a tiny recurrent reasoner refines over multiple steps, decoupling computational depth from model size. Experiments on symbolic and natural‑language reasoning tasks show that LRT outperforms prior frozen‑decoder continuous‑space methods and chain‑of‑thought prompting while using far less inference compute.
By Zhaoliang Chen, Jie Fu
LatentHarness unifies memory access and latent reasoning by treating them as sequential latent actions—THINK, RECALL, and EXIT—within a language model. It is trained via counterfactual policy distillation, which evaluates the impact of each action on the emitted token and learns when to recall evidence versus continue reasoning. On six long‑context reasoning benchmarks, a 1.4B‑parameter LatentHarness model outperforms the strongest baselines by 2.8% and 10.0% relative, while running 5.9× faster than the leading long‑context baseline.
By Xiaoqiang Wang, Suyuchen Wang, Bang Liu
The paper investigates ‘thinking leakage’ in post‑training hybrid reasoning models that operate in NoThink mode. Using a causal mediation framework and bidirectional interventions, the authors show that the performance gains attributed to NoThink training largely stem from the model drifting toward its Think mode. Across three models and three training methods on competition math benchmarks, leakage ratios between 42% and 79% were observed, indicating that much of the accuracy improvement is due to re‑invoking existing Think behavior rather than genuine NoThink capability.
By Zehao Liu, Vasant G. Honavar
arXiv:2608.29956v1 Announce Type: new
Abstract: Large language models often answer complex reasoning questions without revealing intermediate steps, raising whether they reason latently or complete p...
By Armaan Singh, Ryan Trinh Le, Jasmine Kaur, Abdullah Sultan, Edward Lue Chee Lip, Kiran Nijjer, Adnan Ahmed, Vasu Sharma