AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Unlocking Speech–Text Compositional Powers: Instruction-Following Speech Language Models without Instruction Tuning

Congrui Du, Yang Zhang, Kaizhi Qian, Shiyu Chang

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Instruction data selection · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
combine speech-adapted weights from base model with instruct weight difference instead of continuing pre-training the instruct modelSpeechCombine (Qwen3-8B) · OpenbookQAaccuracy: 78.46 -> 86.59Improved
weight combination instead of continuous pre-training plus SFTSpeechCombine · EmphAssess emphasis detectionF1: 2.61 -> 60.84Improved
remove speech captions [cap] from pre-training data (ablation)SpeechCombine · EmphAssess emphasis detectionF1: 60.84 -> 0.39Hurt
add [cap] speech captions to pre-training data (ablation)SpeechCombine · GenEmotion-enscore: 27.18 -> 45.42Improved
add [text] transcription before each speech section (ablation)SpeechCombine · EmphAssess emphasis detectionF1: 0.40 -> 60.84Improved
add [text] transcription before each speech section (ablation)SpeechCombine · GenEmotion-enscore: 35.69 -> 45.42Improved
enable long thinking via answer template (ablation)SpeechCombine · OpenbookQAaccuracy: 64.83 -> 86.59Improved
enable long thinking (ablation)SpeechCombine · GenEmotion-enscore: 36.37 -> 45.42Improved
add instruction weight difference delta theta_inst instead of in-context examples (ablation)SpeechCombine · GenEmotion-enscore: 15.34 -> 45.42Improved
in-context examples replace delta theta_inst vs no-thinking SpeechCombine (ablation)SpeechCombine · OpenbookQAaccuracy: 64.83 -> 71.68Improved
increase combination weight lambda (ablation)SpeechCombine · QA vs speech taskstask performanceMixed
force 'Based on the prosody' at start of thinking (ablation)SpeechCombine · EmphAssess emphasis detectionanswer correctnessImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy