Startup With Fields Medalist Claims Phone-Size Qwen and Cloud GLM Topped ARC-AGI 3 via Learned 'Bridge'

Quantum Bit noted that GPT-6 Astra had already been reported to score near-perfect on the same benchmark. The team's reported results show the bridged 4B model recovering about 50 percent of the performance gap with the 753B model and improving its own accuracy by 25 percent; on a subset of harder problems, the improvement reached two times. Mostik says the cost of running the large model fell to roughly one-twentieth of what it would normally be. To obtain the same marks without the bridge would require an intermediate-size model, and the bridge route uses about two-fifths as much compute as that model.

The company compared its method with standard ways of combining models, including text relay. Across all tested compute levels for the large model, the bridge performed better. Its advantage was biggest when the handoff had to be very early, because a text relay would have no message to pass; hidden states already contain information from the processing done so far. With more compute allocated to the large model, both methods eventually approached the large model's solo performance, but the bridge retained a better performance-compute trade-off throughout.

Mostik's argument is that current multi-model systems, including programming subagents, model committees and router systems, communicate only through text, throwing away most of what a model computes. For every generated token, a model builds more than 100 hidden vectors with around one million values, about two megabytes of state, and then picks one token from a vocabulary of roughly 150,000 terms. The output is about 17 bits; the other two megabytes are discarded.

Quantum Bit cites recent interpretability work to explain why this matters. An ICLR 2026 paper found that Qwen-3 carried a representation of the word 'accountant' several tokens before it wrote that word, and the representation is what let the model choose 'an' instead of 'a' in advance. Anthropic's interpretability team observed a similar effect in Claude 3.5 Haiku: the model settled on rhymes before writing poetic lines, and the team described a structure called 'J-space' in which a model holds a set of unexpressed concepts that are not echoes of the input or guesses about the next token.

Mostik is Russian for 'little bridge.' The company was founded four months ago and has 15 people, 12 of them PhDs. Its chief scientist, Stanislav Smirnov, won the Fields Medal in 2010 for work on conformal invariance in percolation and the planar Ising model. Smirnov said in Quantum Bit's report that finding mathematical common ground between two AI models is very difficult, and no suitable mathematical language has yet emerged to describe how representations correspond. CEO Sasha Malysheva is one of the developers of the core method.

The startup is now exploring distillation and specialization. In distillation, a bridge could give a student model access to the teacher's internal states, not just its final output layer, so supervision could reach the student earlier in generation. Early results, the team says, show a student model getting closer to the teacher with the same training budget. In specialization, a small model with a bridge could learn one professional area more efficiently while keeping general ability; if that works, adding a skill to a large model would not necessarily require retraining the entire system.

Mostik also mentions a possible safety angle. Chain-of-thought provides some monitoring ability, but some abnormal behavior can get around it. A bridge would add an observable channel, since the sender's hidden state, the transformed signal and the receiver's actions can all be logged. The company acknowledges these directions are early and that questions remain about whether the bridge works across more models, tasks and real workloads.