Research map / Large language models / Reasoning and chain of thought

STITCH: Simultaneous Thinking and Talking with Chunked Reasoning for Spoken Language Models

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Search and self-correction in reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes STITCH, which interleaves unspoken reasoning chunks with spoken response chunks in a speech language model so thinking overlaps with audio playback, reaching baseline latency while improving math QA accuracy.

Problem

Spoken language models respond without an internal unspoken reasoning step, and naively generating full chain-of-thought before speaking adds uncontrolled latency to the spoken response.

Approach

Alternate fixed-length unspoken reasoning chunks with text and speech token chunks, using the playback duration of each audio chunk to generate the next reasoning chunk; the Stitch-S variant starts with speech so first-packet latency matches non-reasoning baselines.

Key results

• Stitch-S averages 78.04% on five math QA datasets versus 62.98% for the no-reasoning baseline (about 15% gain), with the same first-packet latency as that baseline.
• TBS (full reasoning before speaking) reaches 79.12% average math accuracy, versus 53.08% for GLM-4-Voice.
• Stitch-R averages 78.70% on math QA, and Stitch-S averages 57.03% on non-reasoning factual QA and AlpacaEval, matching or exceeding the no-reasoning baseline's 55.19%.

Surprising findings

• The first text chunk mostly rephrases the question without new calculation, so Stitch-S can avoid reasoning delay at little accuracy cost.
• LoRA fine-tuning failed to teach useful reasoning to GLM-4-Voice (GSM8K around 35%, near the baseline).

Limitations

• Evaluated only on relatively easy math word problems and factual QA; harder reasoning is not tested.
• Relies on GPT-4o-generated reasoning and synthetic TTS audio for training and evaluation, with GPT-4o as judge.
• Latency gains depend on hardware token throughput; reasoning chunk size must be tuned to audio chunk duration.

Experimental setup

Datasets: VoiceAssistant-400K Tulu-3-Persona-Math Tulu-3-Persona-Math-Grade Tulu-3-Persona-Algebra Natural Questions TriviaQA AddSub SinglEq MultiArith SVAMP GSM8K Llama Questions

Benchmarks: AddSub MultiArith SinglEq SVAMP GSM8K Llama Questions TriviaQA WebQuestions AlpacaEval

Metrics: accuracy GPT-4o score UTMOSv2 human responsiveness score token count

Baselines: GLM-4-Voice No reasoning Mix reasoning TBS

Builds on: GLM-4-Voice chain-of-thought prompting LlamaFactory vLLM

Closest papers by meaning

• Can Speech LLMs Think while Listening? (ICLR 2026 · similarity 0.89)
• Mellow: a small audio language model for reasoning (NeurIPS 2025 · similarity 0.88)
• StreamingThinker: Large Language Models Can Think While Reading (ICLR 2026 · similarity 0.87)
• Audio-Thinker: Guiding Large Audio Language Model When and How to Think via Reinforcement Learning (AAAI 2026 · similarity 0.86)
• ThinkSound: Chain-of-Thought Reasoning in Multimodal LLMs for Audio Generation and Editing (NeurIPS 2025 · similarity 0.86)
• MMAR: A Challenging Benchmark for Deep Reasoning in Speech, Audio, Music, and Their Mix (NeurIPS 2025 · similarity 0.86)