Can smaller language models outperform larger ones at graph extraction?

RAGU's central claim is about what the model inside a retrieval pipeline is for. The paper says practitioners default to large API models because "extraction quality determines graph quality," and calls that "a false premise": the capabilities an in-pipeline LLM needs, "comprehension, extraction, reasoning over context," are "language skills, not factual recall." Language skills "grow only weakly with model size," while world knowledge "scales steeply." The evidence in the excerpt is Meno-Lite-0.1, a 7B model "optimized for language skills," which outperforms Qwen2.5-32B on knowledge-graph construction (+12.5% relative harmonic mean) and matches it on English GraphRAG tasks.

The argument is paired with a structural change. Existing GraphRAG systems treat graph construction as "a single LLM extraction pass," producing "noisy, duplicated entities with no mechanism to consolidate information across chunks." RAGU separates extraction from consolidation: two-stage typed extraction, DBSCAN-backed deduplication, LLM summarization, then Leiden community detection. The two moves reinforce each other. A pipeline that consolidates after extraction asks the extractor for a narrow, repeatable skill and leaves cleanup to later stages, which is the kind of job a small specialized model can plausibly do. The paper's practical framing is a "single-GPU budget."

Against the neighbors, this note adds an axis they leave out: which model builds the graph, not only which graph is built. Can knowledge graphs enable multi-hop reasoning in one retrieval step? describes HippoRAG's LLM-extracted schemaless graph as an offline indexing step; RAGU treats that step as the weak point and contests its cost. Its Leiden stage is the one described in Can community detection enable RAG systems to answer global corpus questions?, now preceded by deduplication. Where Can query-time graph construction replace pre-built knowledge graphs? escapes the cost of a pre-built graph by skipping it, RAGU keeps the pre-built graph and lowers its price. The excerpt also reports a division of labor with HippoRAG 2. RAGU retrieves the most complete context (evidence recall up to 0.84 versus at most 0.76) and leads on synthesis tasks. HippoRAG 2 leads on retrieval precision and on chain-following multi-hop reasoning on MuSiQue. The paper says the wider multi-hop gap under verbose prompts is "largely an answer-format artifact."

The excerpt does not establish the scaling claim itself. It says "as we show next," but the scaling evidence is not in the passages provided, so the weak growth of language skills with size rests on the paper's assertion here. The excerpt also gives no training data, no definition of the harmonic mean, and no comparison against larger models than Qwen2.5-32B. The English-task match says nothing about other languages. The benchmark results cover GraphRAG-Bench (Medical) and MuSiQue only. What follows at this strength is narrow: for the extraction stage of a multi-step GraphRAG pipeline, a compact domain-adapted model is a credible alternative to a larger general model. Whether the same holds for the answering stage, or in other domains, is not addressed.