Research map / Large language models / Language, safety and interpretability
Search or Accelerate: Confidence-Switched Position Beam Search for Diffusion Language Models
Mingyu Cao, Alvaro Correia, Christos Louizos, Shiwei Liu, Lu Yin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace greedy with position beam search (single token) | Dream-7B-Base · avg of HumanEval/MBPP/GSM8K | accuracy: 60.1 -> 63.0 (speedup x0.54) | Improved |
| replace greedy with position beam search (single token) | LLaDA-8B-Base · avg of HumanEval/MBPP/GSM8K | accuracy: 47.8 -> 48.9 (speedup x0.48) | Improved |
| force 2 tokens per step in position beam search (ablation) | LLaDA-8B-Base · avg | accuracy: 48.9 -> 44.9 | Hurt |
| force 2 tokens per step in position beam search (ablation) | Dream-7B-Base · avg | accuracy: 63.0 -> 60.7 | Hurt |
| add confidence-switched search/parallel (SOAR) vs greedy | Dream-7B-Base · avg | accuracy: 60.1 -> 62.4 (speedup x1.43) | Improved |
| add confidence-switched search/parallel (SOAR) vs greedy | LLaDA-8B-Base · avg | accuracy: 47.8 -> 49.2 (speedup x1.62) | Improved |
| raise confidence threshold tau from 0.8 to 0.98 (ablation) | SOAR on Dream-7B · MBPP and GSM8K | accuracy and speedup | Mixed |
| increase beam size from 1 to 4 (ablation) | SOAR on Dream-7B · HumanEval | accuracy and speedup | Mixed |
| use margin metric instead of max probability with SOAR | Dream-7B-Base · HumanEval (256) | pass@1: 47.0 -> 51.2 | Improved |
| use negative-entropy metric with SOAR | Dream-7B-Base · HumanEval (512) | pass@1: 56.1 -> 57.9 | Improved |
| apply SOAR to instruction-tuned model (pure diffusion) | LLaDA-8B-Instruct · avg | accuracy: 49.1 -> 50.5 | Improved |
| apply SOAR to instruction-tuned model (block diffusion) | LLaDA-8B-Instruct · HumanEval (512) | pass@1: 48.8 -> 46.9 | Hurt |
Closest papers by meaning
- Confidence-Based Decoding is Provably Efficient for Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- PACE-dLLM: Elastic Block Decoding via Confidence Cliff Estimation for Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata (NeurIPS 2026 · similarity 0.90)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.90)
- Accelerating Diffusion Language Models via Structured Suffix Modeling (NeurIPS 2026 · similarity 0.90)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
