又一推理新范式：将LLM自身视作「改进操作符」，突破长思维链极限

Core Insights - The article discusses the exploration of a new reasoning method called Parallel-Distill-Refine (PDR) for large language models (LLMs), which aims to improve accuracy while managing context length and computational costs [3][4][12]. Group 1: PDR Methodology - PDR consists of three main steps: (i) generating diverse drafts in parallel, (ii) distilling them into a compact text workspace, and (iii) refining the output based on this workspace [3][4]. - By adjusting the degree of parallelism, PDR can control context length and reduce computational costs, distinguishing it from traditional long chain reasoning methods [3][4]. - When parallelism is set to 1, it results in Sequential Refinement (SR), which outperforms long reasoning chains but incurs higher latency [3][4]. Group 2: Experimental Results - In mathematical tasks with verifiable answers, the PDR method showed significant improvements, with accuracy increasing by 11% and 9% in AIME 2024 and AIME 2025 tasks, respectively [4][12]. - The article reports that the accuracy for the o3-mini model improved from 76.9% (long chain reasoning) to 81.5% (SR) and 86.7% (PDR), marking an absolute increase of +9.8 percentage points [14]. - The gemini-2.5-flash model demonstrated a smaller change from SR to PDR, indicating stronger self-verification capabilities [14]. Group 3: Research Questions and Findings - The researchers posed several questions, including whether short context iterations can outperform long reasoning chains under matched budgets, and which distillation strategies yield the best results [16][19]. - Findings indicated that the top-k sampling and global summarization strategies were more effective than shared top-k and random-k methods, particularly as the reasoning budget increased [19]. - The study also highlighted that the verification capability of models significantly impacts performance, with o3-mini showing a more substantial decline in performance when incorrect candidates were injected compared to gemini-2.5-flash [21]. Group 4: Operator Consistency Training - The article discusses the impact of operator consistency training on shifting the Pareto frontier, with PDR reinforcement learning showing improvements of +3.34 percentage points in AIME 2024 and +1.67 percentage points in AIME 2025 compared to baseline methods [26]. - Continuous updates from baseline RL checkpoints resulted in greater gains, with PDR RL training yielding improvements of +5.00 and +4.59 percentage points in AIME 2024 and AIME 2025, respectively [26][27].