为离散扩散语言模型设计高效推理扩展方法，提升生成质量同时减少计算量。
Prism: Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models
- 分层轨迹搜索动态分配计算资源，聚焦早期到中期去噪阶段。
- 局部分支与部分重掩码策略，在保留高置信度词元的同时探索多样解。
- 自验证反馈替代外部校验器，通过中间结果自评估提升效率。
推理时计算重新成为提升大语言模型推理能力的实用途径。大多数测试时扩展（TTS）算法依赖自回归解码，而离散扩散语言模型（dLLMs）采用全序列并行解码，此特性使其难以适配传统TTS方法。因此，如何为dLLMs设计高效且有效的测试时扩展方法，以充分释放其生成潜力，仍是未被充分探索的挑战。为此，我们提出Prism（剪枝、重掩码与集成自验证方法），一种面向dLLMs的高效测试时扩展框架：（i）执行分层轨迹搜索（HTS），在早期至中期去噪窗口动态剪枝并重分配计算资源；（ii）引入局部分支与部分重掩码机制，在保留高置信度词元的同时探索多样化生成路径；（iii）用自验证反馈（SVF）替代外部校验器，通过自评估提示对中间完成结果进行自我评价。在三个dLLM模型（包括LLaDA 8B Instruct、Dream 7B Instruct 和 LLaDA 2.0-mini）上的四个数学推理与代码生成基准测试中，Prism实现了优越的性能-效率权衡，以显著更少的函数求值次数（NFE）达到与最佳N采样相当的性能。代码已开源于 https://github.com/viiika/Prism。
原文摘要 · Abstract (English)
Inference-time compute has re-emerged as a practical way to improve LLM reasoning. Most test-time scaling (TTS) algorithms rely on autoregressive decoding, which is ill-suited to discrete diffusion language models (dLLMs) due to their parallel decoding over the entire sequence. As a result, developing effective and efficient TTS methods to unlock dLLMs' full generative potential remains an underexplored challenge. To address this, we propose Prism (Pruning, Remasking, and Integrated Self-verification Method), an efficient TTS framework for dLLMs that (i) performs Hierarchical Trajectory Search (HTS) which dynamically prunes and reallocates compute in an early-to-mid denoising window, (ii) introduces Local branching with partial remasking to explore diverse implementations while preserving high-confidence tokens, and (iii) replaces external verifiers with Self-Verified Feedback (SVF) obtained via self-evaluation prompts on intermediate completions. Across four mathematical reasoning and code generation benchmarks on three dLLMs, including LLaDA 8B Instruct, Dream 7B Instruct, and LLaDA 2.0-mini, our Prism achieves a favorable performance-efficiency trade-off, matching best-of-N performance with substantially fewer function evaluations (NFE). The code is released at https://github.com/viiika/Prism.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。