, long text divided with
/
3. Lists with
- , quotes with
, code with
, important sentences can use
4. Preserve all original info, no additions/removals/modifications
5. No // document-level tags
The passage is a single academic abstract about Graph retrieval-augmented generation vs. LoRA adapters for knowledge graphs. Let me structure it:
The passage discusses:
1. Graph retrieval-augmented generation approach (problem statement)
2. Alternative: compiling knowledge graph into LoRA adapters
3. Results on MetaQA dataset
4. Findings about knowledge storage not being recoverable by similarity
5. Weight geometry correlation
6. Cost quantification and deployment implications
7. Conclusion
I could structure this with:
- An intro for the background
- An
for the alternative approach
- An
for results
- An
for the retrieval findings
- An
for conclusion
Actually, since this is relatively short, I might use
sections to break it up. Let me think about natural sections:
1. First part: description of graph RAG and the problem (context window cost, data exposure)
2. The alternative: offline compilation into LoRA adapters
3. Results on MetaQA
4. The non-recoverability finding
5. Weight geometry findings
6. Conclusion
Let me draft:
Graph retrieval-augmented generation places retrieved subgraphs into the model's context window at query time, paying a recurring token cost and exposing source data on every call.
We study an alternative: compiling a knowledge graph offline into a bank of LoRA adapters, one per entity, that serve as a parametric knowledge layer queried by injecting weights rather than text, at zero query-time context cost.
Results on MetaQA
On the MetaQA dataset, we find that subgraph-trained adapters encode context-free factual knowledge that generalizes to unseen questions: on single-valued relations the adapter gains +0.243 exact-match score over a base model that is nearly blind closed-book (0.007), and only the correct adapter recovers this knowledge (an oracle gap of +0.283 over the base model).
Knowledge Is Stored Locally, Not Recoverable by Similarity
However, the stored knowledge is not recoverable by similarity: given a query with no subgraph, embedding-based and weight-space geometry retrieval both perform at chance, because a semantically neighbouring entity's adapter does not contain the answer — knowledge is stored locally and does not transfer.
Weight geometry correlates with subgraph semantics ($ρ$ = +0.329) but not with functional retrievability.
Conclusion
We quantify the byte and context-token costs against graph retrieval-augmented generation and discuss deployment implications. Our results establish that parametric knowledge graph memory is feasible for storing knowledge, and identify selecting and composing the right adapters by a mechanism other than semantic similarity as the central open problem — motivating a learned, query-conditioned composition mechanism.
Wait, the rule says "不得增删改内容" (no additions/removals/modifications). The section headings I add are structural elements, not content modifications, so that's fine — they help divide the long text. But I should be careful the headings don't alter meaning.
Let me keep the original text exactly. I need to make sure I preserve all the text exactly as given:
"Graph retrieval-augmented generation places retrieved subgraphs into the model's context window at query time, paying a recurring token cost and exposing source data on every call."
"We study an alternative: compiling a knowledge graph offline into a bank of LoRA adapters, one per entity, that serve as a parametric knowledge layer queried by injecting weights rather than text, at zero query-time context cost."
"On the MetaQA dataset, we find that subgraph-trained adapters encode context-free factual knowledge that generalizes to unseen questions: on single-valued relations the adapter gains +0.243 exact-match score over a base model that is nearly blind closed-book (0.007), and only the correct adapter recovers this knowledge (an oracle gap of +0.283 over the base model)."
"However, the stored knowledge is not recoverable by similarity: given a query with no subgraph, embedding-based and weight-space geometry retrieval both perform at chance, because a semantically neighbouring entity's adapter does not contain the answer - knowledge is stored locally and does not transfer."
"Weight geometry correlates with subgraph semantics ($ρ$ = +0.329) but not with functional retrievability."
"We quantify the byte and context-token costs against graph retrieval-augmented generation and discuss deployment implications."
"Our results establish that parametric knowledge graph memory is feasible for storing knowledge, and identify selecting and composing the right adapters by a mechanism other than semantic similarity as the central open problem - motivating a learned, query-conditioned composition mechanism."
I'll use for important sentences/metrics. I could bold the key findings like "+0.243 exact-match score" and "+0.283" and the conclusion sentence.
There's no list or blockquote or code in this text, so those elements won't appear. That's fine — the rules say to use them when applicable.
Let me write the final HTML fragment. Should I split into more paragraphs? The original is a single block of text. Splitting sentences into paragraphs would technically modify the structure but the rule says "每个自然段落用
包裹" — each natural paragraph wrapped in
. Since the original is one continuous block, I could either keep it as one
or break it into logical paragraphs. Breaking into logical paragraphs is a common interpretation for "natural paragraphs." I'll break it into logical paragraphs since the original text flows as multiple ideas.
Actually, to be safe about "保留原文所有信息，不得增删改内容" — splitting paragraphs doesn't add/remove/modify content, it just formats. I'll split into logical paragraphs.
Let me finalize:
Graph retrieval-augmented generation places retrieved subgraphs into the model's context window at query time, paying a recurring token cost and exposing source data on every call.
We study an alternative: compiling a knowledge graph offline into a bank of LoRA adapters, one per entity, that serve as a parametric knowledge layer queried by injecting weights rather than text, at zero query-time context cost.
Experimental Results
On the MetaQA dataset, we find that subgraph-trained adapters encode context-free factual knowledge that generalizes to unseen questions: on single-valued relations the adapter gains +0.243 exact-match score over a base model that is nearly blind closed-book (0.007), and only the correct adapter recovers this knowledge (an oracle gap of +0.283 over the base model).
Knowledge Storage Is Local
However, the stored knowledge is not recoverable by similarity: given a query with no subgraph, embedding-based and weight-space geometry retrieval both perform at chance, because a semantically neighbouring entity's adapter does not contain the answer — knowledge is stored locally and does not transfer.
Weight geometry correlates with subgraph semantics ($ρ$ = +0.329) but not with functional retrievability.
Conclusion
We quantify the byte and context-token costs against graph retrieval-augmented generation and discuss deployment implications.
, code with, important sentences can use 4. Preserve all original info, no additions/removals/modifications 5. No // document-level tags The passage is a single academic abstract about Graph retrieval-augmented generation vs. LoRA adapters for knowledge graphs. Let me structure it: The passage discusses: 1. Graph retrieval-augmented generation approach (problem statement) 2. Alternative: compiling knowledge graph into LoRA adapters 3. Results on MetaQA dataset 4. Findings about knowledge storage not being recoverable by similarity 5. Weight geometry correlation 6. Cost quantification and deployment implications 7. Conclusion I could structure this with: - An introfor the background - An
for the alternative approach - An
for results - An
for the retrieval findings - An
for conclusion Actually, since this is relatively short, I might use
sections to break it up. Let me think about natural sections: 1. First part: description of graph RAG and the problem (context window cost, data exposure) 2. The alternative: offline compilation into LoRA adapters 3. Results on MetaQA 4. The non-recoverability finding 5. Weight geometry findings 6. Conclusion Let me draft:
Graph retrieval-augmented generation places retrieved subgraphs into the model's context window at query time, paying a recurring token cost and exposing source data on every call.
We study an alternative: compiling a knowledge graph offline into a bank of LoRA adapters, one per entity, that serve as a parametric knowledge layer queried by injecting weights rather than text, at zero query-time context cost.
Results on MetaQA
On the MetaQA dataset, we find that subgraph-trained adapters encode context-free factual knowledge that generalizes to unseen questions: on single-valued relations the adapter gains +0.243 exact-match score over a base model that is nearly blind closed-book (0.007), and only the correct adapter recovers this knowledge (an oracle gap of +0.283 over the base model).
Knowledge Is Stored Locally, Not Recoverable by Similarity
However, the stored knowledge is not recoverable by similarity: given a query with no subgraph, embedding-based and weight-space geometry retrieval both perform at chance, because a semantically neighbouring entity's adapter does not contain the answer — knowledge is stored locally and does not transfer.
Weight geometry correlates with subgraph semantics ($ρ$ = +0.329) but not with functional retrievability.
Conclusion
We quantify the byte and context-token costs against graph retrieval-augmented generation and discuss deployment implications. Our results establish that parametric knowledge graph memory is feasible for storing knowledge, and identify selecting and composing the right adapters by a mechanism other than semantic similarity as the central open problem — motivating a learned, query-conditioned composition mechanism.
Wait, the rule says "不得增删改内容" (no additions/removals/modifications). The section headings I add are structural elements, not content modifications, so that's fine — they help divide the long text. But I should be careful the headings don't alter meaning. Let me keep the original text exactly. I need to make sure I preserve all the text exactly as given: "Graph retrieval-augmented generation places retrieved subgraphs into the model's context window at query time, paying a recurring token cost and exposing source data on every call." "We study an alternative: compiling a knowledge graph offline into a bank of LoRA adapters, one per entity, that serve as a parametric knowledge layer queried by injecting weights rather than text, at zero query-time context cost." "On the MetaQA dataset, we find that subgraph-trained adapters encode context-free factual knowledge that generalizes to unseen questions: on single-valued relations the adapter gains +0.243 exact-match score over a base model that is nearly blind closed-book (0.007), and only the correct adapter recovers this knowledge (an oracle gap of +0.283 over the base model)." "However, the stored knowledge is not recoverable by similarity: given a query with no subgraph, embedding-based and weight-space geometry retrieval both perform at chance, because a semantically neighbouring entity's adapter does not contain the answer - knowledge is stored locally and does not transfer." "Weight geometry correlates with subgraph semantics ($ρ$ = +0.329) but not with functional retrievability." "We quantify the byte and context-token costs against graph retrieval-augmented generation and discuss deployment implications." "Our results establish that parametric knowledge graph memory is feasible for storing knowledge, and identify selecting and composing the right adapters by a mechanism other than semantic similarity as the central open problem - motivating a learned, query-conditioned composition mechanism." I'll use for important sentences/metrics. I could bold the key findings like "+0.243 exact-match score" and "+0.283" and the conclusion sentence. There's no list or blockquote or code in this text, so those elements won't appear. That's fine — the rules say to use them when applicable. Let me write the final HTML fragment. Should I split into moreparagraphs? The original is a single block of text. Splitting sentences into paragraphs would technically modify the structure but the rule says "每个自然段落用
包裹" — each natural paragraph wrapped in
. Since the original is one continuous block, I could either keep it as one
or break it into logical paragraphs. Breaking into logical paragraphs is a common interpretation for "natural paragraphs." I'll break it into logical paragraphs since the original text flows as multiple ideas. Actually, to be safe about "保留原文所有信息，不得增删改内容" — splitting paragraphs doesn't add/remove/modify content, it just formats. I'll split into logical paragraphs. Let me finalize:
Graph retrieval-augmented generation places retrieved subgraphs into the model's context window at query time, paying a recurring token cost and exposing source data on every call.
We study an alternative: compiling a knowledge graph offline into a bank of LoRA adapters, one per entity, that serve as a parametric knowledge layer queried by injecting weights rather than text, at zero query-time context cost.
Experimental Results
On the MetaQA dataset, we find that subgraph-trained adapters encode context-free factual knowledge that generalizes to unseen questions: on single-valued relations the adapter gains +0.243 exact-match score over a base model that is nearly blind closed-book (0.007), and only the correct adapter recovers this knowledge (an oracle gap of +0.283 over the base model).
Knowledge Storage Is Local
However, the stored knowledge is not recoverable by similarity: given a query with no subgraph, embedding-based and weight-space geometry retrieval both perform at chance, because a semantically neighbouring entity's adapter does not contain the answer — knowledge is stored locally and does not transfer.
Weight geometry correlates with subgraph semantics ($ρ$ = +0.329) but not with functional retrievability.
Conclusion
We quantify the byte and context-token costs against graph retrieval-augmented generation and discuss deployment implications.