Research map / Large language models / Language, safety and interpretability
Data-Centric Lessons To Improve Speech-Language Pretraining
Vishaal Udandarao, Zhiyun Lu, Xuankai Chang, Yongqiang Wang, Albin Madappally Jose, Fartash Faghri, Joshua P Gardner, Chung-Cheng Chiu
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 fine instead of coarse speech-text interleaving (ablation) | speech-text pretraining · SQA avg (SWQ/STQ/SLQ) | accuracy (%): 37.6 -> 40.7 | Improved |
| mix in Krist synthetic data (ablation) | web-crawl pretraining · SQA avg (SWQ/STQ/SLQ) | accuracy (%): 40.7 -> 41.5 | Mixed |
| mix in Quest QA-format synthetic data (ablation) | web-crawl pretraining · SQA avg (SWQ/STQ/SLQ) | accuracy (%): 40.7 -> 47.9 | Improved |
| mix both Quest and Krist (token-proportional) (ablation) | web-crawl pretraining · SQA avg (SWQ/STQ/SLQ) | accuracy (%): 40.7 -> 42.1 | Improved |
| use deterministic instead of stochastic modality sampling (ablation) | interleaved pretraining · SQA avg (SWQ/STQ/SLQ) | accuracy (%): 41.4 -> 42.4 | Improved |
| deterministic sampling effect on text benchmarks (ablation) | interleaved pretraining · CoreEN / MMLU | accuracy (%): 60.6 -> 60.1 CoreEN; 65.7 -> 65.2 MMLU | Hurt |
| apply all data interventions with audio-loss masking | understanding-only SpeechLM · SQA avg (SWQ/STQ/SLQ) | accuracy (%): 42.5 -> 51.8 | Improved |
| mask loss on audio tokens (ablation) | SpeechLM pretraining · SQA avg (SWQ/STQ/SLQ) | accuracy (%): 42.4 -> 51.8 | Improved |
| fine interleaving after SFT (ablation) | SpeechLM · spoken-alpaca | LC win rate (%): 42.6 -> 44.3 | Improved |
| add Quest synthetic data after SFT (ablation) | SpeechLM (fine) · spoken-alpaca | LC win rate (%): 44.3 -> 47.4 | Improved |
| add synthetic data on audio-response Eval 5 (ablation) | SpeechLM (fine) · Audio Eval 5 | win rate (%): 41.9 -> 39.3 | Hurt |
| fine interleaving reduces text-audio reverse KL (ablation) | speech-text pretraining · Spoken-LLaMA-Questions | reverse-KL: 3.20 -> 2.21 | Improved |
Closest papers by meaning
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.90)
- Webscale-RL: Automated Data Pipeline for Scaling RL Data to Pretraining Levels (ICLR 2026 · similarity 0.87)
- Speech Self-Supervised Learning Using Diffusion Model Synthetic Data (ICML 2024 · similarity 0.86)
- Measuring Audio's Impact on Correctness: Audio-Contribution-Aware Post-Training of Large Audio Language Models (ICLR 2026 · similarity 0.86)
- MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark (ICLR 2026 · similarity 0.86)
- CCFQA: A Benchmark for Cross-Lingual and Cross-Modal Speech and Text Factuality Evaluation (AAAI 2026 · similarity 0.86)
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.
