Research map / Large language models / Language, safety and interpretability
SageLM: A Multi-aspect and Explainable Large Language Model for Speech Judgement
Yuan Ge, Junxiang Zhang, Xiaoqian Liu, Bei Li, Xiangnan Ma, Chenglong Wang, Kaiyang Ye, Yangfan Du, Linfeng Zhang, Yuxin Huang, Tong Xiao, Zhengtao Yu and 1 more
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 |
|---|---|---|---|
| fine-tune with rationale-augmented SFT instead of label-only SFT or GRPO (ablation) | speech judge (Qwen2.5-omni) · SpeechFeedback preliminary | compare accuracy / agreement | Improved |
| use GRPO rule-based RL instead of label-only SFT (ablation) | speech judge · SpeechFeedback preliminary, scoring accuracy | scoring accuracy | Improved |
| use GRPO instead of label-only SFT (ablation) | speech judge · SpeechFeedback preliminary, pairwise compare | compare accuracy | No clear effect |
| train with GRPO without supervising rationales (ablation) | speech judge · manual analysis of 100 sampled cases | rationale-score consistency: explanation contradicted final score in 39% of cases | Hurt |
| add stage 1 semantic training only (GRPO or SFT) (ablation) | speech judge · acoustic evaluation | acoustic agreement | Hurt |
| add stage 2 acoustic preference data after stage 1 (ablation) | SageLM · acoustic and semantic evaluation | agreement | Improved |
| end-to-end SageLM instead of Whisper + GPT-4o cascade | speech judge · AlpacaEval real S2S outputs | agreement: 69.30% -> 87.97% | Improved |
| two-stage SFT on Qwen2.5-omni-7B vs direct prompting | Qwen2.5-omni-7B · SpeechFeedback semantic test, Helpfulness | agreement (%): 56.50 -> 83.61 | Improved |
| use task-specific PandaLM-7B vs general Qwen2.5-32B in cascade (ablation) | Whisper cascade judge · SpeechFeedback semantic test | agreement (%): 67.58 vs 67.40 (Helpfulness) | No clear effect |
| scale base model from 3B to 7B for SageLM (ablation) | SageLM · SpeechFeedback acoustic test, Mixed | agreement (%): 94.44 -> 91.36 (Mixed); 97.47 -> 98.65 (Implicit) | Mixed |
| increase combined response pair length beyond 60 s (ablation) | SageLM · AlpacaEval long responses | agreement with humans | Hurt |
Closest papers by meaning
- Paralinguistics-Aware Speech-Empowered Large Language Models for Natural Conversation (NeurIPS 2024 · similarity 0.88)
- SD-Eval: A Benchmark Dataset for Spoken Dialogue Understanding Beyond Words (NeurIPS 2024 · similarity 0.88)
- JudgeLM: Fine-tuned Large Language Models are Scalable Judges (ICLR 2025 · similarity 0.87)
- Enhancing Speech Large Language Models through Reinforced Behavior Alignment (NeurIPS 2026 · similarity 0.87)
- MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark (ICLR 2026 · similarity 0.87)
- Audio Large Language Models Can Be Descriptive Speech Quality Evaluators (ICLR 2025 · 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.
