Mostik Tops ARC-AGI with Latent Model Bridging

Russian mathematicians behind Mostik have reached first place on the ARC-AGI benchmark using a new technique for connecting large and small language models. Instead of the standard approach where one model writes an intermediate text response that another reads, Mostik transfers hidden internal states through a small, trainable bridge. The original models remain frozen.

In the first experiment, a 753B-parameter GLM-5.2 model passed its latent representation to a 4B-parameter Qwen-3.5 model via this bridge. The large model handled only the prompt processing (prefill), while the small model generated the final answer. The 4B model closed roughly 50% of the gap to the 753B model’s performance, boosted its own accuracy by about 25%, and achieved up to 2x improvement on harder task subsets. The combined pipeline required 2.5x less compute than a mid-sized model delivering comparable quality, and latent hand-off outperformed plain text transfer by up to 10 percentage points.

Mostik sees potential applications in multi-agent systems, knowledge distillation, model specialization, and research into internal representations.

Mostik, X post