id
gotchas/meta-llama-cpp-has-no-paper-do-not-invent-one
Concepts
llama.cpparxivprovenancecitationggml
Symptom
A reader auditing this repo for inference coverage will notice that llama.cpp - by most measures the most widely deployed LLM inference engine in the world - has engine, kernel and gotcha records but NO record in data/papers/. The tempting response is to go and find an arXiv ID for it, and that is the trap. Searching for one produces papers that merely MENTION llama.cpp, and a plausible-looking identifier attached to 'llama.cpp' would be a fabricated citation for one of the most-cited projects in the field.
Root cause
None. llama.cpp has never published a paper. Verified 2026-10-04 against export.arxiv.org/api/query: search_query=all:"llama.cpp" returns 5 entries and every one is a third-party study that uses llama.cpp rather than describing it (2601.14277 a quantization evaluation of llama.cpp, 2406.10816 an Armv9 inference-optimization study, 2511.05502 a comparative Apple Silicon serving study, 2603.23996 a forensic analysis of Ollama). No title search for llama.cpp as a system returns anything either.
Workaround
Cite the repository, which is what llama.cpp itself does. The repo already represents llama.cpp correctly through engine/kernel source records rather than a paper record; a future pass finding llama.cpp work worth recording should add an engine or kernel source, never a paper.
notes
Recorded as an honest NEGATIVE because the repo's own contract (AGENTS.md) says every claim must be traceable and contradicted later, and a fabricated identifier is the failure mode this repo has already shipped twice (FlashDecoding, and Marlin before its real preprint 2408.11743 was established). The same discipline applies to absence: stating 'there is no paper, here is the search that proves it' is more useful than leaving a silent hole that invites the next agent to fill it with a guess. Note the contrast with the other famous engine work in this repo: vLLM/PagedAttention, Orca, SGLang, FlashInfer, TensorRT-LLM and DeepSpeed all DID publish, which is exactly why they have paper records - the absence here is specific to llama.cpp, not a general gap in engine coverage.