llm-circuit-finder compares duplicated layers and reports BBH logical deduction gains

• The new llm-circuit-finder toolkit packages an inference-time experiment for GGUF models in llama.cpp-style setups: sweep contiguous layer ranges, duplicate selected blocks, and measure whether reasoning improves without retraining.
• In the repo summary, duplicating layers 12-14 in Devstral-24B raised BBH logical deduction from 0.22 to 0.76, while duplicating layers 7-9 in Qwen2.5-32B improved reasoning by 17% repo summary.
• The project is framed as a reproducible workflow, not just a claim: it includes sweep.py for circuit discovery, layer_path.py for path modification, and evaluation scripts, according to the toolkit page.
• The HN discussion pushes back on the explanation, with commenters arguing the effect may reflect near-identity layers, routing or looping behavior, or training artifacts rather than a settled “reasoning circuit” mechanism.

The release is a small research toolkit around inference-time model surgery. The toolkit page says it replicates Ng's RYS method by duplicating specific contiguous layer blocks during inference, working with llama.cpp and GGUF models and tested on Mistral- and Qwen-family architectures.

That makes the engineering contribution concrete: instead of retraining or merging checkpoints, the workflow searches for layer spans that are worth repeating at runtime. The bundled tools cover three steps — search, path modification, and eval — via sweep.py, layer_path.py, and evaluation scripts, as described on the repo page and summarized in the Show HN post.

The headline result is large enough to get attention. In the HN summary, the author reports that duplicating three layers in a 24B model pushed logical deduction from 0.22 to 0.76, and the linked repo page gives a second example where duplicating layers 7-9 in Qwen2.5-32B improved reasoning by 17%.

But the mechanism is very much in dispute. One commenter says the proposed explanation “does not pass the smell test,” arguing duplicated layers may be “near-identity blocks” or may undo reasoning damage introduced during training or RLHF skeptical thread. Another suggests the pattern looks more like a routing or looping effect — “a higher-level MoE-style routing problem” with paths such as 13,13,14,14,15,15,16 — rather than evidence of a clean circuit interpretation routing interpretation. A third comparison points to Solar 10.7B depth up-scaling, where repeated layers appeared during continued training, which makes this look more like an inference-time variant of a known idea than a wholly new primitive prior-art note.