Loopie looped MoE claims 2025 IMO and IPhO gold without tools

For a while now the field has treated depth via looping as an interesting curiosity that rarely paid off. The rule of thumb was that if you had N times the pre-training compute, you were better off spending it on N times the parameters than on running the same weights N times. A new preprint on Hugging Face Papers, titled Loop the Loopies!, argues that rule is not as fixed as people assumed.

The authors describe Loopie, a looped Transformer built as a Mixture-of-Experts. There are two variants, a 20 billion parameter model with 2 billion active and a 6 billion parameter model with 0.6 billion active. They compare against a vanilla 30B-A3B baseline trained on the same compute budget, and report that the looped MoE comes out ahead. The eye-catching headline is the reasoning result: gold-medal performance at the 2025 IMO and IPhO, without tools, after a post-training pipeline the paper says equips Loopie with strong reasoning abilities.

If it holds up, the implication for anyone running inference at scale is that active parameter count and total parameter count can be pulled apart more aggressively than the current MoE recipes assume. Sharing weights across loop iterations means you get depth without paying for extra parameters, which changes the math on memory-limited hardware and on serving cost per reasoning trace. It also opens a door for teams that cannot outspend the frontier labs on raw parameter count.

The honest caveat is that this is one preprint, self-reported, and the abstract itself frames the win as ablations against a single vanilla baseline. Take the gold-medal claim as reported, not settled, until independent groups run the models against IMO and IPhO problems themselves. The reporting also does not tell you what the post-training pipeline actually is, how the olympiad answers were graded, or what inference latency looks like once you unroll the loop at serving time.

Still, the direction is the interesting part. If depth via weight sharing genuinely competes with parameter scaling at matched compute, the cost curve for small-active-parameter reasoning models tightens up, and the 'just make it bigger' default loses another chunk of its power.