header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Ideal has disclosed the Mach-Mind-4-Flash training method: synthesizing a model from over ten experts

According to VisionOne Beating monitoring, Ideal Cars has released the Mach-Mind-4-Flash technical report. The model had been previously unveiled, with this release focusing on additional post-training details.

The team first individually trained over a dozen expert models in fields such as mathematics, coding, writing, security, search, and tool invocation. They then used MOPD (Multi-teacher Online Policy Distillation) to merge them into a single universal model. Each category was guided by the respective experts to reduce interference between multiple abilities during mixed training.

The report also explains how the Agent was trained. The model not only learned tool names and invocation formats but also actually read files, modified code, ran tests, and invoked tools in an isolated sandbox. Upon operation failure, it could observe errors and environmental changes, then continue to make modifications. Tool training covered over 190 stateful domains and 3500+ tool interfaces; a single training trajectory for programming tasks could last up to 300 rounds.

This fusion was not entirely lossless. Instruction-following abilities were largely preserved, with ClawBench and ClawEval even outperforming individual Agent experts. However, SWE-bench Verified decreased from 73.80 in the expert model to 71.10. The team acknowledged that highly specialized abilities like long-range programming would still suffer losses during the merge.

The report also revealed how the model's tendency to "overthink" was reduced. HMPO dynamically set the reasoning budget based on the median length of correct answers, rewarding only shorter answers that were correct. This led to a length reduction of 19% to 46%, with a maximum accuracy drop of 0.7 percentage points. Currently, this method is only used for single-round reasoning and has not yet covered multi-round Agent tasks.

举报 Correction/Report
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish