A November 19 preprint by Tanmay Sah proposes circuit grafting: analytically inserting deterministic computational structures into pretrained language models. Its title, A CPU Inside an LLM, describes the approach metaphorically rather than announcing a physical processor placed inside a model.

The author reports 99.17% for Granite 4.2 3B and 100% for Qwen 3.5 4B on 1,200 six-digit arithmetic examples in BF16 using a specific ALU-prefix format. This does not establish universal mathematical accuracy. SIExpo has not independently reproduced the results or verified the construction through its own implementation.

Why task formulation matters

Our interpretation is that a bounded computational operation and a freely expressed mathematical question are not identical tasks. The former can have strictly defined inputs and expected outputs. The latter also requires understanding the request, selecting an operation and transferring the answer into an appropriate context. Success in one part does not establish success in all the others.

A proposed assessment example would present the same arithmetic in structured and natural-language forms. Its report could separate correct operation selection from correct calculation. This is our acceptance scenario, not an additional experiment described in the preprint or an already measured property of the method. Keeping these stages separate would make failures easier to interpret.

The modification has costs and scope

The full paper describes compilation mechanisms and auxiliary variants with different numerical conditions. No gradient training consequently does not mean no preparation, resources or restrictions. The abstract and detailed text are the same underlying source; they do not independently corroborate one another or establish that the reported result has been replicated elsewhere.

For a separate study, we would retain the exact base-model version, transformation and runtime environment. Results should be compared with a clear reference on a predefined dataset. A further check could examine whether the change affects tasks outside the inserted operation. These are evaluation questions, not a promise that the proposal has resolved every one of them.

A research idea with testable boundaries

The approach is interesting because it attempts to combine a learned language model with a stricter computational structure. Its significance depends on repeatability, compatible architectures and actual execution cost. A strong table in a preprint cannot replace those checks or determine operational suitability for an unrelated application.

For now, the supported conclusion is limited: a proposal and reported results exist under specified conditions. They do not establish general error-free arithmetic in language models or make external verification tools unnecessary. The next step is reproducible investigation that keeps format and resource conditions attached to the accuracy figures rather than separating an impressive number from the test that produced it.