When the Edge Is Missing: HyGRL Keeps the Text in the Graph

TL;DR for operators

When a question requires several facts to be connected, improving entity coverage is not enough if the graph still lacks the relations needed to move between those facts. In a 200-query structural-connectivity pilot, pure Freebase connected the required endpoints within three hops in 24.1% of cases. Wikidata raised entity linking from 77.0% to 85.6%, but connectivity reached only 37.6%. Adding document text into the Freebase-based graph raised connectivity to 69.7%.

HyGRL turns that observation into an architecture: retain raw text chunks as graph nodes, connect them to mentioned entities, and train a small MLP to decide which neighboring nodes deserve exploration. The LLM helps train and evaluate that router, but does not make the online routing decisions.

For enterprise QA teams, the relevant trade is therefore not simply “Graph-RAG versus text RAG.” It is whether to spend more on structured graph construction and online LLM reasoning, or preserve document evidence inside the graph and shift navigation intelligence into a learned retrieval component. HyGRL’s benchmark results support the latter design within its tested setting, but its 35.95 GB HotpotQA RAM footprint and entity-linking filter make direct production generalization premature.

More linked entities do not guarantee a usable reasoning path

A company can have the relevant information in two places and still fail to retrieve the connection between them. One document mentions an acquisition. Another record identifies a subsidiary. A structured graph contains both entities, yet no explicit relation joins the pieces required by the question.

That is the problem behind HyGRL: Adaptive Hybrid Graph Reasoning for Multi-Entity Questions.¹ The paper’s Appendix A pilot is particularly revealing because it separates entity presence from answer connectivity. Wikidata links more entities than Freebase in the 200 sampled HotpotQA and MuSiQue queries, but still connects the required endpoints within three hops in only 37.6% of cases. The HyGRL configuration retains Freebase’s 77.0% linking rate yet reaches 69.7% connectivity.

This is an exploratory structural pilot, not the paper’s main benchmark result. Its role is diagnostic: it shows why enriching the graph solely by finding more entities may leave the retrieval problem unsolved.

HyGRL instead keeps raw document chunks inside the graph and links them bidirectionally to the entities they mention. These chunks become text relay nodes: retrieval can move from a structured entity into original text and back into another entity even when the knowledge graph lacks a direct relation.

The important architectural choice is what the system refuses to discard. Rather than reducing the corpus entirely to extracted triples, it preserves source text as active retrieval infrastructure.

The graph substrate matters, but so does deciding where to traverse

Adding more possible paths creates another problem: the retriever now needs to decide which paths deserve attention.

HyGRL represents retrieval as a sequential graph-search problem. The current reasoning state contains recently selected nodes scored for question relevance. Candidate neighbors receive features based on their relationship to that state plus a Personalized PageRank signal. A lightweight MLP then scores each candidate, and adaptive beam search expands only sufficiently promising, unvisited nodes.

The ablation results distinguish the contribution of the graph representation from the contribution of routing:

The first two drops are larger than the BFS replacement. That pattern suggests the substrate—having both textual and structured pathways—is doing substantial work. The learned router then improves how that substrate is used.

Across the filtered 2WikiMultiHopQA, HotpotQA, and MuSiQue subsets, full HyGRL reaches 56.24 average EM and 63.87 average F1, the strongest averages among the systems evaluated. It also records the highest EM on each of the three datasets.

These are comparative benchmark results, not evidence that the architecture dominates every retrieval workload. The evaluated questions were filtered to require at least two successfully linked seed entities.

LLM judgment trains the router rather than operating it query by query

A natural reading of “learned graph reasoning” is that an LLM agent repeatedly decides which edge to follow. HyGRL uses a different division of labor.

The router first imitates a teacher assembled from four signals: a context-aware cross-encoder, Personalized PageRank, path-aware Adamic-Adar, and Jaccard overlap. Their weights are calibrated and fused, giving the MLP a denser initialization than a sparse end-task reward alone would provide.

The policy is then refined with two forms of LLM supervision. Pairwise preference training asks which of two candidate expansions is more logically useful. Self-Critical Sequence Training evaluates whether an entire sampled retrieval trajectory provides sufficient evidence relative to a greedy trajectory. Local preference and trajectory-level sufficiency therefore train different aspects of routing.

This is where the system’s economics differ from agentic retrieval. LLM judgment is used during offline preparation and training, while the online graph router itself uses the learned MLP without LLM calls. The paper also reports roughly a 90% reduction in construction-time LLM token cost relative to LLM-heavy graph-building approaches.

That does not mean HyGRL is LLM-free. LLMs participate in entity extraction, calibration verification, preference labeling, trajectory sufficiency judgments, and downstream generation. “Zero-token retrieval” applies specifically to the online routing step.

Better retrieval is visible in the evidence, not only the generated response

Answer metrics can conceal weak retrieval: a generator may recover despite incomplete support, or produce the right string for the wrong evidence.

The paper therefore measures evidence fidelity on 2Wiki. HyGRL reaches 44.7% SF-F1 for textual support and 39.9% UE-F1 for the unified text-and-graph evidence set, both highest among the compared retrieval systems. It does not win every component: HippoRAG2-hybrid records 21.9% E-F1 for structured triples versus HyGRL’s 17.1%.

For retrieval-product owners, Cognaptus infers a useful evaluation pattern here. If the system is expected to support auditability, investigation, or evidence-backed enterprise QA, teams can score whether required passages and structured facts were actually recovered instead of relying exclusively on answer accuracy.

The paper does not establish that its specific fidelity metrics transfer unchanged to enterprise schemas. The broader measurement principle is the transferable part.

Lower online LLM cost is exchanged for graph infrastructure

HyGRL’s deployment results prevent an overly simple efficiency claim. On HotpotQA, average end-to-end latency is 1.82 seconds compared with 1.15 seconds for the LLM-only measurement, implying roughly 0.67 seconds of retrieval overhead. Reported inference VRAM is approximately 475 MB.

System RAM is 35.95 GB—the highest among the retrieval systems listed in the paper’s efficiency table. Keeping the heterogeneous graph resident makes routing inexpensive in tokens and GPU memory, but moves part of the burden into conventional memory infrastructure.

For a high-volume enterprise QA service, Cognaptus therefore sees a concrete decision boundary. HyGRL-like routing is more attractive when repeated online LLM search calls are expensive, query volume is high enough to amortize offline graph and policy preparation, entity linking is reliable, and sufficient RAM can be provisioned. A sparse-query workload with unstable entities or rapidly changing corpora may produce a different cost balance.

The evidence stops at linked, benchmark-style questions

Two boundaries matter most.

First, benchmark evaluation excludes questions without at least two successfully linked seed entities. HyGRL therefore demonstrates what its router can do after a minimum graph-entry condition has been satisfied. It does not establish robustness to poor entity linking across arbitrary enterprise documents.

Second, the learned policy inherits judgments from its LLM mentor. The paper finds 88% agreement between LLM pairwise preferences and expert consensus on 100 reviewed HotpotQA pairs. Its robustness test also shows gradual rather than catastrophic degradation when preference labels are deliberately flipped: F1 falls by 0.90 points at 15% noise, 2.06 at 30%, and 4.09 at 50%.

That experiment is a robustness test, not proof that mentor quality is unimportant. The router is still learning an approximation of the mentor’s preferences.

HyGRL’s strongest contribution is therefore architectural rather than a universal performance claim. It demonstrates that the missing relation in a knowledge graph does not have to be reconstructed as another structured edge before reasoning can proceed. Source text itself can remain part of the traversable graph, while learned routing decides which mixed evidence paths deserve computation.

For enterprise retrieval systems built over documents plus imperfect structured data, that changes a practical design choice: before investing in more exhaustive graph extraction or more online LLM navigation, test whether preserving the original text inside the graph creates the connections the workload actually needs.

Cognaptus: Automate the Present, Incubate the Future.