LLM-to-GNN Projection

Updated 9 November 2025

LLM-to-GNN Projection is a computational mechanism that maps the outputs and reasoning traces of large language models to graph neural network inputs for structured data tasks.
It leverages a structure-enhanced retriever combining text encoding and GraphSAGE to select informative subgraphs and convert them into effective in-context learning prompts for LLMs.
Empirical findings, exemplified by AskGNN, demonstrate improved performance in node classification, link prediction, and text generation without any LLM fine-tuning.

LLM to Graph Neural Network (LLM-to-GNN) projection describes any computational mechanism that translates the outputs, features, or reasoning traces of a LLM into forms that can be directly integrated, optimized, or interpreted within a graph neural network (GNN) framework. This projection is a cornerstone for connecting large-scale LLMs—originally designed for sequential data or unstructured text—to the relational, highly structured world of graphs. As graph-structured tasks pervade information retrieval, molecular modeling, structured document understanding, knowledge graph reasoning, and scientific literature analysis, LLM-to-GNN projections provide a principled mechanism for unleashing the semantic capacity of LLMs within the inductive bias of GNNs. The following sections synthesize the dominant projection architectures, mathematical mappings, training algorithms, empirical effects, and open challenges according to recent literature, with a focus on the “AskGNN” method (Hu et al., 2024).

1. Motivation and Landscape

The use of LLM-to-GNN projection arises from the structural misalignment between transformer-based LLMs—whose token-wise contextual representations are designed for text sequences—and GNNs, which operate on graph-structured data (node, edge, subgraph, or global tasks). For text-attributed graphs (TAGs), each node may be associated with a document or rich description, yet incorporating graph topology into LLM workflows, or semantic priors into GNNs, is nontrivial. Prior approaches fall broadly into:

Feature-level alignment: e.g., projecting LLM-derived embeddings into the GNN’s feature space or vice versa (Shi et al., 11 Feb 2025, Liu et al., 2024).
Message-passing in text: simulating GNN propagation in language space using prompt engineering (Zhu et al., 5 Mar 2025, 2505.20742).
In-context learning: constructing examples, prompts, or subgraphs as LLM input via GNN-powered selection, enabling the LLM to perform graph-reasoning tasks without model fine-tuning (Hu et al., 2024).
Supervision or alignment losses: directly regularizing one model with the features or predictions of another (“distillation”) (Xu et al., 2024).

AskGNN epitomizes the graph-to-LLM in-context projection: it uses a GNN to select, assemble, and present the most informative labeled subgraphs as textual prompt context for the LLM. This permits LLMs to harness graph structure implicitly, without alteration of their architectures.

2. Structure-Enhanced Retriever as LLM-to-GNN Interface

A defining mechanism in LLM-to-GNN projection is the formulation of a structure-aware retriever comprised of a text encoder and GNN message-passing operator. The canonical pipeline is as follows (Hu et al., 2024):

Let $G = (A, X)$ denote a textual attributed graph. Each node document $x_i$ is encoded as $h_i^0 \in \mathbb{R}^d$ using an off-the-shelf text encoder (e.g., Llama tokenizer + embedding layer).
The initial representations undergo $L$ layers of a structure-propagating GNN, typically GraphSAGE:

$h_i^\ell = \sigma \left( W_\ell \cdot \text{CONCAT}\left( h_i^{(\ell-1)}, \text{MEAN}_{j \in N(i)} h_j^{(\ell-1)} \right) + b_\ell \right)$

with $\sigma =$ ReLU, $W_\ell \in \mathbb{R}^{d \times 2d}$ , $b_\ell \in \mathbb{R}^d$ , $N(i)$ the neighbors of $i$ .

Post-GNN, each node $i$ yields an embedding $h_i = h_i^L$ .
For any query node $q$ , the similarity scores $\text{sim}(h_q, h_i) = \frac{h_q \cdot h_i}{\|h_q\| \|h_i\|}$ are computed between $q$ and all labeled candidates, and the top- $K$ examples form the support set $D_q$ .

The retriever thus projects from (graph, text) to a ranked set of exemplar nodes, fusing semantic and structural cues.

3. Projection Pipeline: Graph-to-Text Context Construction

The core “projection” as per AskGNN is the mapping:

$\Pi(G, x_q) = T(\{(x_i, y_i)\}_{i=1}^K, x_q)$

where $T$ is a deterministic function—implemented as a prompt template—that flattens $K$ labeled examples plus a query node into a single input context string for the LLM. For node classification:

A fixed instruction is provided.

Each retrieved example is serialized:

1
2
3

Example #:
  Paper: <x_i>
  Label: <y_i>

The query node appears as:
1 2 3
Now classify the following paper: Paper: <x_q> Answer:
This pipeline encodes both supervision and the local structural context of $q$ $q$ into a sequential input, enabling the LLM to perform in-context learning with graph-structured priors. No special tokens or model-specific markers are used; the approach remains universal.

4. Learning-to-Retrieve Algorithm & Feedback from LLM

A defining advance is the closed-loop alignment of the retriever to the LLM’s task utility. For each candidate example $e = (x_e, y_e)$ :

The LLM’s utility is quantified using a perplexity-based metric:

$\text{PPL}(y_c | x_q, e) = \exp\left( -\frac{1}{|y_c|} \sum_{v=1}^{|y_c|} \log p(w_v | w_{<v}, x_q, e) \right)$

The utility score is:

$s(e) = \frac{1/\text{PPL}(y_q | x_q, e)}{\sum_{c \in \text{Classes}} 1/\text{PPL}(y_c | x_q, e)}$

A “pseudo-oracle” ranking $\hat{D}_q$ orders the $K$ candidates by $s(e)$ .
The retriever’s similarity scores are then optimized using a contrastive softmax loss:

$L_{\text{feedback}} = -\sum_{q \in \text{train}} \sum_{k=1}^{K} \log \left[ \frac{\exp(\text{sim}(h_q, h_{k^+}))}{\sum_{j=1}^{K} \exp(\text{sim}(h_q, h_j))} \right]$

where $h_{k^+}$ refers to the embedding in position $k$ of the ranked support set.

An auxiliary graph-based classification loss $L_{\text{clf}}$ is optionally included and the full loss is a weighted sum: $L = \beta L_{\text{feedback}} + (1-\beta) L_{\text{clf}}$ with $\beta \in [0.2, 0.3]$ empirically optimal.

This algorithm ensures that the GNN retriever selects examples that demonstrably improve the LLM’s in-context prediction accuracy.

5. Integration and Inference Protocol

After training the structure-enhanced retriever:

For test queries, one computes the top- $K$ support set $D_q = R(x_q, G)$ , constructs the input prompt $\Pi(G, x_q)$ , and feeds it to the LLM.
The LLM’s native self-attention mechanism suffices to process the entire prompt; no adapters or new tokens are required.
Prediction is terminated upon emission of a label token, which serves as the system's final output.

This pipeline generalizes to various tasks: node classification, link prediction (via binary prompts), and conditional text generation (prompting the LLM to “complete” node documents). No fine-tuning of the LLM is involved; the adaptation flows solely through the context.

6. Empirical Effects, Ablations, and Impact

Experimental validation on ogbn-arxiv, ogbn-products, and arxiv23 established:

AskGNN outperforms all competing baselines: averaged node classification accuracy (Qwen1.5-72B) reaches 71.65%, surpassing k-NN ICL (69.30%), zero-shot (64.08%), and GCN (65.22%).
For link prediction, AskGNN achieves 89.06% (vs. GCN 85.12%); for conditional text generation, Rouge-L rises to 22.15 (vs. few-shot 19.80).
Performance scales with $K$ (number of in-context examples) up to $K \approx 20$ , after which gains plateau.
Selection of support via the retriever and feedback loss yields large improvements over random/k-NN selection; naive neighbor-based pseudo-labeling methods are 3–5 points worse.
LLM-selection for noise reduction offers a further +1% gain.
The methodology maintains gains in both low-label and high-label regimes and is robust across datasets.

Ablation results confirm that:

Optimal $\beta \approx 0.2$ –$0.3$; improper weighting degrades performance.
Support set curation and integration of LLM feedback are critical; removal of either diminishes accuracy (3–5 points below full AskGNN).
Minor modifications to prompt construction (e.g., minority class removal) have negligible effect.

Significance: This class of LLM-to-GNN projection, exemplified by AskGNN, enables general-purpose, structure-aware in-context learning on graphs via off-the-shelf LLMs. It unlocks high-accuracy graph reasoning without LLM fine-tuning, is compatible with arbitrary graph domains, and isolates all adaptation in a lightweight, interpretable retriever.

7. Limitations and Open Directions

Current LLM-to-GNN projections, as implemented in AskGNN, do not alter the LLM’s architecture; they leverage prompt-based conditioning and support selection. This framework does not inject GNN representations into the LLM’s hidden layers nor does it fine-tune the LLM. Structurally, it is pull-based: the GNN projects curated subgraphs/examples into the LLM’s prompt space, but there is no bidirectional gradient flow between the LLM and GNN. Potential future work includes architectures that enable tighter coupling or mutual adaptation between the GNN and LLM (e.g., as in certain distillation or graph vocabulary learning models (Xu et al., 2024, Zhu et al., 5 Mar 2025)), direct injection of GNN embeddings into LLM token representations, and application to more complex graph tasks (e.g., subgraph classification, hierarchy reasoning).

Summary Table: (LLM-to-GNN Projection via Structure-Enhanced Retriever in AskGNN)

Component	Mechanism	Role
Text Encoder (LLM upstream)	Tokenizer + embedding layer	Encodes node documents
Structure-propagation (GNN)	Multi-layer GraphSAGE aggregation	Infuses structural information
Retriever	Cosine-similarity; top- $K$ selection	Chooses in-context examples for the LLM
Projection function $\Pi$	Prompt template maps (graph, query, support) $\rightarrow$ text string	Converts structure + supervision into LLM context
Learning-to-retrieve	LLM perplexity-based utility, contrastive ranking loss	Makes GNN support maximally helpful for LLM
Downstream task	Prompted to off-the-shelf LLM	Enables in-context learning for graph tasks

This approach demonstrates that graph structure and supervision signals can be projected into the in-context learning mechanism of LLMs via optimized prompt construction, selecting examples using a GNN retriever, without modifying or fine-tuning the LLM itself (Hu et al., 2024).