Researchers propose Graph of Concept Predictors (GCP), a reasoning-aware active distillation framework for Large Language Models (LLMs) that improves performance under limited annotation budgets while yielding more interpretable and controllable training dynamics.

MetaCtrl is a lightweight controller that adaptively regulates large language models (LLMs) to improve their reasoning accuracy while reducing inference-time generation. It observes the evolving reasoning trace, decides whether to continue, simplify, or conclude reasoning, and is trained using reinforcement learning. MetaCtrl improves the accuracy of LLMs on various benchmarks, including mathematics, science, and code, and can transfer to unseen reasoners without further training.

Researchers propose ACTR, a framework to improve multilingual safety alignment in reasoning large language models by strengthening safety reasoning through neuron-selective consistency optimization. ACTR achieves lower attack success rates on jailbreak queries and preserves or improves performance on multilingual knowledge and mathematical reasoning tasks.

A new framework, Self-Play Search Distillation (SPSD), generates high-quality synthetic data for Large Language Models (LLMs) using self-play of MuZero-like networks trained on board games. This technique improves LLM performance in reasoning tasks, such as mathematics, with minimal human annotation.