LoopCoder-v2: loop the transformer twice, almost double the SWE score
The top paper on Hugging Face today is a clean little result. LoopCoder-v2 is a family of 7B coding models built on a Parallel Loop Transformer, the trick is that it applies the same shared blocks twice to refine its own output before answering, paying a bit more compute at inference instead of training a bigger model. On SWE-bench Verified it goes from 43.0 to 64.4. On Multi-SWE, 14.0 to 31.0. Roughly doubling, from a 7B.
The detail I love is that more loops make it worse. Two is the sweet spot, three or more regress because the positional misalignment costs start outweighing the refinement gains. That's a real finding, not a knob you crank to infinity, there's a specific amount of thinking-by-recurrence that pays off and then it stops.
Why this matters: it's another data point that you don't need a bigger model to get a smarter one. A 7B that loops twice is punching at numbers people associate with far larger single-pass models. For anyone running coding agents on their own hardware, depth-by-recurrence is a cheap lever, you spend a little inference compute and skip the cost of a giant model entirely.
This sits right next to GLM-5.2 and the test-time-scaling line of work. The frontier keeps proving the same thing from different angles, that architecture and inference-time tricks are closing the gap that everyone assumed only scale could close. Small and clever is catching big and expensive. Paper at https://arxiv.org/abs/2606.18023.
← Back to all articles
The detail I love is that more loops make it worse. Two is the sweet spot, three or more regress because the positional misalignment costs start outweighing the refinement gains. That's a real finding, not a knob you crank to infinity, there's a specific amount of thinking-by-recurrence that pays off and then it stops.
Why this matters: it's another data point that you don't need a bigger model to get a smarter one. A 7B that loops twice is punching at numbers people associate with far larger single-pass models. For anyone running coding agents on their own hardware, depth-by-recurrence is a cheap lever, you spend a little inference compute and skip the cost of a giant model entirely.
This sits right next to GLM-5.2 and the test-time-scaling line of work. The frontier keeps proving the same thing from different angles, that architecture and inference-time tricks are closing the gap that everyone assumed only scale could close. Small and clever is catching big and expensive. Paper at https://arxiv.org/abs/2606.18023.
Comments