While excelling at solving complex problems, Large Reasoning Models (LRMs) are still constrained by the overthinking issue. Most current studies rely on reward shaping in Reinforcement Learning (RL) to shorten the Chain-of-Thought (CoT) of LRMs, remaining sample-inefficient and non-robust due to the absence of guided exploration and prioritized exploitation. To address these issues, we propose a novel policy optimization framework with Self-Imitation and self-Guidance MechAnisms (SIGMA), which reshapes the exploration and exploitation through two core components: (i) self-imitation exploitation, which enables the prioritized exploitation of high-value prompts and rollouts by introducing a self-imitated loss and a dynamic sampling strategy based on compression rate; (ii) self-guidance exploration, which provides a preference-aware exploration guidance through diverse and pluggable self-rewriting strategies. Experiments across various datasets indicate that our method achieves superior reasoning efficiency without compromising, and even facilitating, the overall accuracy. Furthermore, ablation studies show that the proposed mechanisms can provide flexible control interfaces for the tradeoff between the reasoning accuracy and efficiency of LRMs.

Scaling test-time computation in Multimodal Large Language Models (MLLMs) enforces heavy visual analysis and long logical reasoning uniformly across queries, leading to severe token waste and overthinking. We propose **Ada-DeVL**, a framework that decouples visual perception depth from logical reasoning intensity, allowing the model to adaptively select one of four visuo-reasoning pathways according to problem complexity. To learn this adaptive behavior, we first apply a Self-Distillation Supervised Fine-Tuning (SFT) stage that generates difficulty-aligned, hallucination-reduced data while avoiding the exposure bias of cross-model distillation. We then use Group Relative Policy Optimization (GRPO) with a composite reward, incorporating diversity and format-adherence constraints to prevent mode collapse and shortcut exploitation. Notably, on the MMMU benchmark, the Qwen2.5-VL-7B based Ada-DeVL achieves **54.0%** accuracy, surpassing Vision-SR1 while saving **39.6%** tokens; the Qwen3-VL-8B based model reaches **64.0%**, outperforming its vanilla counterpart and reducing token consumption by **44.6%**, and achieves competitive accuracy against the strong Qwen3-VL-8B-Thinking baseline with significantly lower token costs (averaging 1560.8 vs. 4889.0 tokens).