Adaptive Parallel Reasoning (UC Berkeley Technical Report EECS-2026-213)

作者： UC Berkeley (2026)

领域

推理能力推理

TLDR（中文）

用强化学习让模型在推理时动态分配计算：简单问题串行短推理，复杂问题自动展开多个并行推理线程再汇总收敛，突破串行 CoT 的长度瓶颈，是 test-time compute 自适应分配方向的代表性技术报告。

TLDR (English)

Uses RL to let models allocate compute adaptively at inference time: short serial reasoning for easy problems, automatically spawning multiple parallel reasoning threads that converge for hard ones, breaking the length bottleneck of serial chain-of-thought. A representative technical report on adaptive test-time compute allocation.

出现在这些文章里

• Sampling 与 Decoding：从概率到文字
• Sampling and Decoding: From Probabilities to Text

同被引用

这些论文与本文出现在同一篇文章中

• Language Models are Few-Shot Learners Tom Brown et al. (2020)
• Accelerating Large Language Model Decoding with Speculative Sampling Charlie Chen et al. (2023)
• Fast Inference from Transformers via Speculative Decoding Yaniv Leviathan et al. (2023)
• Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters Charlie Snell et al. (2024)

相关论文

同一领域的其他论文

• Agent World Model: Scaling Agentic RL with Synthetic Verifiable Environments Agent World Model Team (2026)
  应用推理能力
• MISA: Mixture of Indexer Sparse Attention MISA Team (2026)
  长上下文推理
• On-Policy Self-Distillation for Reasoning Models Self-Distilled Reasoner Team (2026)
  对齐推理能力
• DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning DeepSeek-AI (2025)
  推理能力对齐
• DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model DeepSeek-AI (2024)
  架构推理
• OpenAI o1 System Card OpenAI (2024)
  推理能力安全