Research map / Large language models / Language, safety and interpretability
Semantic Voting: A Self-Evaluation-Free Approach for Efficient LLM Self-Improvement on Unverifiable Open-ended Tasks
Chunyang Jiang, Yonggang Zhang, Yiyang Cai, Chi-Min Chan, Yulong Liu, Mingming Chen, Wei Xue, Yike Guo
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 |
|---|---|---|---|
| use SVSI instead of entropy minimization | Llama-8B · WMT24++ de-en | BLEU: 0 -> 20.76 | Improved |
| use SVSI instead of entropy minimization | Llama-3B · WMT24++ de-en | BLEU: 10.52 -> 8.71 | Hurt |
| use SVSI instead of self-judging | Qwen-7B · CNN/DailyMail | ROUGE-L: 7.49 -> 20.69 | Improved |
| train with SVSI vs base model | Qwen-7B · WMT24++ de-en | BLEU: 16.42 -> 20.19 | Improved |
| train with SVSI vs base model | Llama-8B · WMT24++ de-en | BLEU: 20.22 -> 20.76 | No clear effect |
| self-judging pseudo-labels | Llama-8B · CNN/DailyMail | ROUGE-L: 20.68 -> 0.08 | Hurt |
| add self-generation clustering filter (ablation) | SVSI · Llama-1B | performance | Improved |
| add self-generation clustering filter (ablation) | SVSI · Qwen-1.5B | performance | No clear effect |
| add semantic voting on top of clustering-only pairs (ablation) | SVSI · translation and summarization | performance | Improved |
| flip SV preference pairs (ablation) | SVSI · translation and summarization | performance | Hurt |
| use DeBERTaV3 embeddings instead of SimCSE (ablation) | SVSI · WMT24++ avg BLEU, Llama-1B | BLEU: 5.13 -> 3.54 | Hurt |
| use Mean Shift clustering instead of HDBSCAN (ablation) | SVSI · WMT24++ avg BLEU, Llama-1B | BLEU: 5.13 -> 3.85 | Hurt |
Closest papers by meaning
- Importance Weighting Can Help Large Language Models Self-Improve (AAAI 2025 · similarity 0.88)
- SelfIE: Self-Interpretation of Large Language Model Embeddings (ICML 2024 · similarity 0.87)
- Self-Supervised Alignment with Mutual Information: Learning to Follow Principles without Preference Labels (NeurIPS 2024 · similarity 0.87)
- Scalable Best-of-N Selection for Large Language Models via Self-Certainty (NeurIPS 2025 · similarity 0.87)
- Self Iterative Label Refinement via Robust Unlabeled Learning (NeurIPS 2025 · similarity 0.87)
- Self-correcting for Debiasing Large Language Models (ICML 2026 · similarity 0.87)
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.
