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

2026-07-22 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Efficient Chain-of-Modality Reasoning (ECoM Reasoning) is a novel framework designed to enhance the reasoning capabilities of Spoken Language Models (SLMs), particularly in spoken mathematical question answering. SLMs typically struggle with verbalized mathematical expressions compared to text-based large language models. ECoM Reasoning addresses this by introducing compressed reasoning, where a textual component is jointly used for speech guidance and reasoning representation. This approach significantly improves reasoning accuracy while reducing the token budget compared to the standard Chain-of-Modality (CoM) architecture. To facilitate this, the authors developed Progressive Compression, a curriculum-based training strategy that transitions models from full-form to compressed reasoning. Benchmarking on spoken mathematical question answering tasks demonstrated ECoM Reasoning's effectiveness, showing a 21% accuracy improvement over standard CoM without explicit reasoning and a 3% gain over CoM with full reasoning traces, all while utilizing only 40% of the text tokens.

Key takeaway

For NLP Engineers developing Spoken Language Models (SLMs) for complex tasks like mathematical question answering, you should consider implementing compressed reasoning techniques. ECoM Reasoning demonstrates that by integrating compressed textual components for both speech guidance and reasoning, you can achieve significant accuracy gains (e.g., 21% over standard CoM) while substantially reducing token usage (40% of text tokens). This approach offers a path to more inference-efficient and capable SLMs, allowing your models to handle intricate verbalized expressions more effectively.

Key insights

ECoM Reasoning improves SLM accuracy and efficiency by compressing textual reasoning traces for joint speech guidance.

Principles

• Compressing reasoning traces enhances SLM efficiency.
• Jointly using text for guidance and reasoning is effective.
• Curriculum learning aids complex model training.

Method

ECoM Reasoning compresses textual components to serve as speech guidance and reasoning representation. Progressive Compression trains this via a curriculum from full-form to compressed reasoning.

In practice

• Apply compressed reasoning to SLMs.
• Use curriculum learning for complex model training.
• Optimize token usage in multimodal models.

Topics

• Spoken Language Models
• Chain-of-Modality
• Progressive Compression
• Mathematical Reasoning
• Model Efficiency
• Multimodal AI

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

Related on AIssential

• 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.
• Efficient Chain-of-Modality Reasoning via Progressive Compression for Spoken Language Models — Compressed textual reasoning enhances SLM accuracy and efficiency by dual-purposing intermediate text for speech guidance and core problem-solving.
• SLMJury: Can Small Language Models Judge as Well as Large Ones? — Small Language Models can effectively judge model outputs, but their performance is highly dependent on the task domain and judging paradigm.
• LoRi: Low-Rank Distillation for Implicit Reasoning — Reasoning trajectories in LLMs exhibit low-rank structure, enabling efficient distillation into compact latent processes.
• 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.
• Multilingual Chain-of-Thought Compression via Cross-Lingual Distillation — MCD compresses multilingual Chain-of-Thought reasoning using cross-lingual data alignment and reinforcement learning, improving efficiency and low-resource language robustness.

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 Computation and Language.