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:

ARC Competition First Place Team Revealed: Mostik Bridges the Gap for Small Models, Saving on Large Model Training Costs

Dynamic Beating AI News: The current leader in the ARC-AGI-3 Kaggle competition, CSTL, is the recently unveiled AI startup Mostik, with a current score of 7.51%. The team consists of 15 members, including 12 PhDs and a Fields Medalist. However, the competition is not over yet. The second Milestone is set to close on September 30, with the final submission deadline on November 2.


Mostik's approach involves enabling different AI models to directly exchange internal states. While traditional multi-agent systems mainly rely on text-based communication, Mostik trains a bridge between two models to transform one model's hidden states into a representation understandable by the other model, without requiring fine-tuning on either side.


Their disclosed experiment involved a 753B GLM-5.2 and a 4B Qwen3.5, where the large model handles prefilling and the smaller model handles decoding. The GLM-5.2 is read-only and does not generate tokens; it passes its internal state to the 4B small model to generate answers. Mostik claims that this approach can reduce the performance gap between the two models by about half, with only around 40% of the computation needed compared to a medium-sized single model achieving the same accuracy. However, the final performance is still below that of the 753B large model itself.


Mostik has not yet revealed which models and harnesses were used to achieve the 7.51% score. The combination of 753B+4B mentioned above is just another technical demonstration and should not be considered the configuration for winning the competition. The ARC-AGI-3 competition itself focuses on the entire Agent system, with previous stage winners relying only on the Qwen 3.6 27B model but leveraging harnesses, tools, and context management to secure the first position.


On the other hand, this approach requires access to the models' hidden states, rendering cloud-based models like Claude and GPT currently unusable. Google DeepMind engineer Susan Zhang has also raised doubts, questioning whether freezing two models and separately training a "bridge" once internal state control is achieved is indeed the most cost-effective engineering solution.

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