Research map / Large language models / Language, safety and interpretability
Arithmetic Sampling: Parallel Diverse Decoding for Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 independent ancestral draws with randomly shifted arithmetic-code lattice samples | T5-base translation model · WMT14 EnFr and WMT16 EnRo; CNN/DailyMail | estimator standard deviation: cut in half once we reach N ≈ 16 samples | Helped |
| Replace independent ancestral draws with arithmetic sampling | T5-base translation model · WMT14 EnFr, oracle reranking | maximum BLEU: 42.74 -> 43.20 | Helped |
| Replace independent ancestral draws with arithmetic sampling | T5-base translation model · WMT14 EnFr, oracle reranking | maximum BLEU: 45.90 -> 46.62 | Helped |
| Replace independent ancestral draws with arithmetic sampling | T5-base translation model · WMT14 EnFr, oracle reranking | maximum BLEU: 48.68 -> 49.39 | Helped |
| Use arithmetic sampling instead of ancestral sampling | T5-base translation model · WMT14 EnFr, oracle reranking, beam sizes 5, 10, and 20 | gap to beam-search maximum BLEU closed: 33%, 63%, and 56% | Helped |
| Use arithmetic sampling instead of ancestral sampling | T5-base translation model · WMT14 EnFr, candidate diversity | n-gram diversity | Helped |
| Use arithmetic sampling instead of ancestral sampling | T5-base translation model · WMT14 EnFr, mean BLEU across sampled candidates | mean BLEU | No clear effect |
| Combine arithmetic sampling with top-k truncation | T5-base translation model · WMT14 EnFr, top-k values 2 and 10 | maximum BLEU | Helped |
| Apply a shared random uniform shift to the regular lattice codes (ablation) | Naive arithmetic sampling · Sequence-model expectations | estimator bias | Helped |
Closest papers by meaning
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.90)
- Dynamic-Width Speculative Beam Decoding for LLM Inference (AAAI 2025 · similarity 0.89)
- Semantic-guided Diverse Decoding for Large Language Model (NeurIPS 2025 · similarity 0.89)
- Hot or Cold? Adaptive Temperature Sampling for Code Generation with Large Language Models (AAAI 2024 · similarity 0.88)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.88)
- Entropy-informed Decoding: Adaptive Information-Driven Branching (ICML 2026 · similarity 0.88)
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.