ECoM Reasoning: сжатые рассуждения ускоряют голосовые модели в устной математике

What problem does ECoMReasoning solve

Voice language models still reason worse than text LLMs, and, as the authors of the paper note (arXiv, July 2026), the gap is most visible in spoken math tasks. The reason is that SLMs operate on fully "spoken out" mathematical expressions: a phrase like "three times five plus two" is harder for the model to interpret than the symbolic notation 3×5+2. At the same time, simply transferring text-based reasoning chains into a voice model isn't possible — architectural constraints and extra computational costs get in the way.

How the method works

ECoMReasoning compresses the text component so that it simultaneously serves as a prompt for speech generation and as a representation of the reasoning itself. This is the key difference from the standard Chain-of-Modality (CoM) architecture, which first generates an expanded intermediate text and only then speech — spending noticeably more tokens on this.

In standard CoM, the intermediate text is needed only as a reasoning draft before speech synthesis, and the model pays for it at its full length. ECoMReasoning makes the same text perform a dual role — driving speech generation and storing the solution logic — so it can be made shorter without hurting the answer. It is precisely this merging that cuts the token budget.

Key results from experiments on spoken math QA benchmarks:

• +21% accuracy compared to standard CoM without explicit reasoning
• +3% accuracy compared to CoM with full reasoning traces
• text token usage — only 40% of the baseline
• ECoMReasoning is the first framework to introduce compressed reasoning into SLMs

What is Progressive Compression

Progressive Compression is a training strategy based on curriculum learning that gradually guides the model from reasoning in full form to compressed form. First the model learns to reason in an expanded, step-by-step way, and then — increasingly compactly — until it shifts to a compressed format. This staged transition, according to the authors' design, is what allows accuracy to be preserved while cutting text length to 40% of the original volume: the model doesn't lose the reasoning skill, it simply learns to express it more concisely.

Why this matters for voice assistants

ECoMReasoning shows that the quality of voice models' reasoning can be improved without increasing inference cost. Usually, better reasoning means more intermediate tokens — and therefore higher latency and computational load. Here the trade-off is removed: a 3% accuracy gain over a strong baseline CoM is achieved with 40% of the text tokens, meaning the result is simultaneously more accurate, faster, and cheaper. For voice assistants, which must respond aloud almost instantly, this is fundamental.

"ECoMReasoning improves SLM reasoning while remaining efficient at

inference time," the paper's abstract on arXiv states.

What this means

Voice interfaces increasingly need to not just recognize speech but solve tasks aloud — from math to step-by-step instructions. ECoMReasoning is a step toward voice AI models reasoning more accurately (21% above the baseline in tests) without becoming slower or more expensive.

Frequently asked questions

What is Chain-of-Modality?

Chain-of-Modality (CoM) is a voice model architecture in which the model first generates an intermediate reasoning text and then converts it into speech. ECoMReasoning compresses this text step, combining within it both the reasoning itself and the speech prompt.

How much more efficient is ECoMReasoning than regular CoM?

According to the experiments, ECoMReasoning is 21% more accurate than standard CoM without reasoning and 3% more accurate than CoM with full reasoning traces, while using only 40% of the text tokens.