Efficient Chain-of-Modality Reasoning via Progressive Compression for Spoken Language Models

2026-07-23 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Spoken Language Processing · Depth: Expert, extended

Summary

The Efficient Chain-of-Modality Reasoning (ECoM Reasoning) framework addresses the reasoning gap in Spoken Language Models (SLMs) for tasks like mathematical question answering. SLMs typically struggle with verbalized math expressions, and direct transfer of text-based reasoning is computationally expensive for architectures like Chain-of-Modality (CoM). ECoM Reasoning introduces compressed textual reasoning, allowing the intermediate text to simultaneously guide speech generation and carry core reasoning information, reducing the token budget. A Progressive Compression training strategy, moving from full-form to compressed reasoning, enables this capability. Experiments on spoken mathematical QA benchmarks, including AddSub and MultiArith, show ECoM Reasoning improves accuracy by 21% over standard CoM. It also achieves 3% higher accuracy than CoM with full reasoning traces, while using only 40% of the text tokens. Furthermore, it reduces first-token latency from 4.90 s to 1.60 s.

Key takeaway

For AI Scientists and Machine Learning Engineers developing Spoken Language Models for mathematical reasoning or similar complex tasks, you should investigate ECoM Reasoning. This framework significantly boosts accuracy by 3% over full reasoning traces and reduces text token usage by 60%. It also cuts first-token latency from 4.90 s to 1.60 s. Implementing its progressive compression strategy can enhance your model's reasoning capabilities and inference efficiency without increasing the token budget.

Key insights

Compressed textual reasoning enhances SLM accuracy and efficiency by dual-purposing intermediate text for speech guidance and core problem-solving.

Principles

• Intermediate text in SLMs can be functionally dense, supporting both speech generation and compact reasoning.
• Curriculum learning, progressing from full to compressed reasoning, effectively trains complex multimodal capabilities.
• Moderate compression of reasoning traces can improve accuracy by reducing redundancy and error accumulation.

Method

ECoM Reasoning compresses intermediate user text (sentence-level removal) and reasoning text (token-level via LLMLingua-2 importance scores) within a CoM architecture. Progressive Compression trains in three stages: Standard CoM, CoM Reasoning (full-form), then ECoM Reasoning (compressed).

In practice

• Utilize token-level importance scoring (e.g., LLMLingua-2) for reasoning trace compression.
• Employ multi-stage curriculum learning for training models with compressed intermediate representations.
• Target a 40% reasoning text compression ratio for optimal accuracy-efficiency balance.

Topics

• Spoken Language Models
• Chain-of-Modality
• Reasoning Compression
• Curriculum Learning
• Mathematical Reasoning
• Inference Efficiency

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Efficient Chain-of-Modality Reasoning via Progressive Compression for Spoken Language Models — ECoM Reasoning improves SLM accuracy and efficiency by compressing textual reasoning traces for joint speech guidance.
• Dynamic Large Concept Models: Latent Reasoning in an Adaptive Semantic Space — DLCMs improve LLM efficiency by shifting computation from tokens to a compressed, concept-based reasoning space.
• Better Starts, Better Ends: Bootstrapped Iterative Self-Reasoning Distillation for Compressed Reasoning — Improving initial rollout distribution is crucial for efficient self-reasoning distillation in large language models.
• Efficient Chain-of-Modality Reasoning via Progressive Compression for Spoken Language Models — ECoM Reasoning introduces compressed reasoning to Spoken Language Models, significantly improving accuracy and efficiency for mathematical question answering.
• Progressive Cramming: Reliable Token Compression and What It Reveals — Perfect token reconstruction through cramming does not guarantee semantic utility or preserved downstream model capabilities.
• Dynamic Thinking-Token Selection for Efficient Reasoning in Large Reasoning Models — DynTS efficiently prunes redundant reasoning tokens in LRMs by identifying decision-critical KV cache states.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.