Dynamic Global-Recent Adaptive Semantic Processing

• Dynamic Global-Recent Adaptive Semantic Processing (DyGRASP) is a framework that leverages large language models and temporal GNNs to extract both recent semantic dependencies and global semantic evolution from dynamic text-attributed graphs.
• It employs a node-centric sliding window and segmented global reasoning approach to efficiently process interaction histories, reducing token redundancy while maintaining temporal context.
• Empirical evaluations demonstrate up to a 34% improvement in destination node retrieval, underscoring its robust performance across various datasets and platforms.

Searching arXiv for the specified paper and closely related context papers. Dynamic Global-Recent Adaptive Semantic Processing (DyGRASP) is a framework for reasoning over Dynamic Text-Attributed Graphs (DyTAGs), a class of graphs in which both interactions and associated text evolve over time. The method is introduced in "Global-Recent Semantic Reasoning on Dynamic Text-Attributed Graphs with LLMs" (Wang et al., 23 Sep 2025). Its central premise is that dynamic graph reasoning must account for two distinct temporal semantic regimes: recent semantic dependency among nearby interactions and global semantic dynamics of nodes across longer horizons. DyGRASP uses LLMs to extract both forms of temporal semantics and then fuses them with a temporal graph neural network (GNN), so that the final representation jointly reflects text, time, and dynamic graph structure (Wang et al., 23 Sep 2025).

1. Problem formulation in dynamic text-attributed graphs

A DyTAG is defined as

$$\mathcal{G}=\{\mathcal{V},\mathcal{E},\mathcal{T},\mathcal{D},\mathcal{R}\},$$

where $\mathcal{V}$ is the node set, $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$ is the edge set, $\mathcal{T}$ is the set of timestamps, $\mathcal{D}$ is the set of node textual attributes, and $\mathcal{R}$ is the set of edge textual attributes. A node $u$ has text attribute $D_u\in\mathcal{D}$, and an interaction is written as

$I=(u,\mathbf{r},v,t),$

where $u$ interacts with $\mathcal{V}$0 at time $\mathcal{V}$1, and $\mathcal{V}$2 is the interaction text (Wang et al., 23 Sep 2025).

The formulation emphasizes that DyTAGs differ from static text-attributed graphs because the text on nodes and edges is not merely descriptive. It evolves with time and is tied to a sequence of interactions. The paper identifies two semantic patterns induced by this setting. The first is recent semantic dependency of temporal interactions, meaning that the interpretation of an interaction text depends on nearby past interactions. The second is global semantic dynamics of nodes, meaning that the underlying meaning of a node can drift over long time horizons, reflecting gradual changes in behavior, interests, or roles (Wang et al., 23 Sep 2025).

This distinction is fundamental to DyGRASP’s design. A plausible implication is that temporal text in graph data cannot be adequately treated as either static metadata or an unstructured event log. Instead, the semantic state of the graph must be modeled as both locally context-sensitive and globally nonstationary.

2. Motivation and limitations of standard approaches

The paper positions DyGRASP against two dominant methodological families: temporal GNNs and LLM-centric semantic models. Standard temporal GNNs are described as effective at topology and time-aware aggregation, but they usually rely on shallow text features or fixed node attributes. As a result, they do not deeply reason over the language of interactions and therefore miss semantics carried by evolving text itself (Wang et al., 23 Sep 2025).

LLM-only approaches, by contrast, are strong at semantic understanding but encounter two specific difficulties on DyTAGs. The first is scale and efficiency: there are far more edges than nodes, and each edge may carry text, so naively feeding long historical interaction sequences into an LLM is too expensive. The second is temporal structure: an LLM alone does not model graph structure or dynamic message passing, and even if it summarizes text effectively, it does not explicitly learn from temporal neighborhoods (Wang et al., 23 Sep 2025).

The paper further argues that recent and global temporal semantics are both necessary. Recent interactions can alter the meaning of a current interaction, while long-term history captures deeper drift in a node’s state. Using only one granularity is therefore incomplete: recent context alone does not capture long-term evolution, and global summaries alone can miss immediate semantic dependencies (Wang et al., 23 Sep 2025). This is also where a common misconception is addressed. DyGRASP does not present LLM reasoning as a replacement for temporal GNNs; rather, it treats semantic reasoning and temporal message passing as complementary components.

3. Core architecture of DyGRASP

DyGRASP contains three main components: implicit reasoning for recent temporal semantics, explicit reasoning for global semantic dynamics, and integration with dynamic graph structure through temporal GNN layers (Wang et al., 23 Sep 2025).

For recent semantics, the paper first notes that a straightforward edge-centric strategy would feed each interaction together with all relevant past interactions into the LLM. If average degree is $\mathcal{V}$3, this naive approach has complexity $\mathcal{V}$4. DyGRASP instead uses node-centric implicit reasoning. For a node $\mathcal{V}$5, it collects all interactions with neighbors, denoted $\mathcal{V}$6, sorts them chronologically, and processes them as one causal sequence. The rationale is that LLMs are causal sequence models, so left-to-right generation aligns naturally with temporal order and prevents future leakage (Wang et al., 23 Sep 2025).

Because LLM context length is limited and long sequences degrade performance, DyGRASP introduces a sliding window mechanism. The sequence $\mathcal{V}$7 is divided into overlapping batches: $\mathcal{V}$8 where $\mathcal{V}$9 is the window length. The first half of each window provides context for the second half. Each batch is linearized with a dataset-specific prompt and passed through the LLM, and the hidden representations for each interaction are mean-pooled to yield a recent temporal semantic feature (Wang et al., 23 Sep 2025). In the paper’s complexity analysis, this reduces token use from $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$0 to $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$1.

For global semantics, DyGRASP uses explicit reasoning. For a node $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$2, its historical interactions $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$3 are evenly partitioned into $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$4 segments $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$5, with partition timestamps

$\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$6

Each segment contains interactions whose timestamps fall between $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$7 and $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$8. DyGRASP then constructs a reasoning chain reminiscent of an RNN: $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$9 where $\mathcal{T}$0 is the original node text attribute and $\mathcal{T}$1 is the LLM-generated textual description of the node after incorporating the interactions in period $\mathcal{T}$2 (Wang et al., 23 Sep 2025). Rather than summarizing the entire history in one pass, the method repeatedly updates the node’s textual description period by period, propagating long-term semantic drift forward. The paper states that this helps preserve background knowledge and carry global semantics through time.

4. Representation update and fusion with temporal graph structure

After extracting recent and global semantics, DyGRASP integrates them with graph-structural information through temporal GNN layers. At layer $\mathcal{T}$3, the model maintains recent semantic features $\mathcal{T}$4, global semantic features $\mathcal{T}$5, graph-structural features $\mathcal{T}$6, and an integrated node representation $\mathcal{T}$7 (Wang et al., 23 Sep 2025).

For a node $\mathcal{T}$8 at prediction time $\mathcal{T}$9, let

$\mathcal{D}$0

Each interaction feature in the recent stream is formed by projecting the LLM-derived feature and concatenating it with a time encoding: $\mathcal{D}$1 The recent semantic sequence is then updated by a Transformer Encoder: $\mathcal{D}$2

For the global stream, the LLM-generated description $\mathcal{D}$3 for each temporal segment $\mathcal{D}$4 is embedded by a text encoder such as BERT to produce $\mathcal{D}$5. The initial global representation is

$\mathcal{D}$6

A second Transformer Encoder updates this sequence: $\mathcal{D}$7 where

$\mathcal{D}$8

which ensures no future leakage (Wang et al., 23 Sep 2025).

The graph structure layer uses a temporal GNN: $\mathcal{D}$9 where $\mathcal{R}$0 aggregates features from the temporal neighborhood of $\mathcal{R}$1 up to time $\mathcal{R}$2. The paper explicitly demonstrates compatibility with TGAT and DyGFormer (Wang et al., 23 Sep 2025).

Finally, DyGRASP fuses the three streams using a merge layer: $\mathcal{R}$3 where $\mathcal{R}$4 is mean pooling over recent features, $\mathcal{R}$5 takes the final global state, and $\mathcal{R}$6 is implemented as an MLP. The output after $\mathcal{R}$7 layers, $\mathcal{R}$8, is the final node representation used for downstream prediction (Wang et al., 23 Sep 2025).

5. Data flow and computational interpretation

The processing pipeline in DyGRASP is explicitly staged. First, node interactions are collected in chronological order. Second, recent reasoning is performed by segmenting interactions into overlapping windows, feeding each batch to the LLM with a prompt, and extracting hidden states for each interaction. Third, global reasoning partitions the history into $\mathcal{R}$9 time segments, repeatedly prompts the LLM with the previous segment summary and current segment, and obtains evolving descriptions $u$0. Fourth, the semantic outputs are encoded by projection and time encoding. Fifth, the recent and global streams are processed by separate Transformer Encoders. Sixth, a temporal GNN performs graph propagation over the dynamic neighborhood. Seventh, pooled recent, pooled global, and structural features are merged. Eighth, the final node embedding is used with an MLP for the downstream task (Wang et al., 23 Sep 2025).

The efficiency analysis is integral to this workflow. The paper states that node-centric reasoning drastically reduces token use compared with the naive approach. On GDELT, recent reasoning decreases from 1.68B tokens without node-centric reasoning to 0.10B tokens with it. The same section reports that each interaction is processed only twice in total under the combined recent-global design, so the cost remains $u$1 (Wang et al., 23 Sep 2025).

This suggests that DyGRASP is not only a semantic augmentation strategy but also a cost-aware restructuring of how LLMs are applied to dynamic graph data. The method does not merely add language modeling to temporal GNNs; it constrains semantic inference so that it remains tractable under abundant and evolving edge text.

6. Evaluation, ablations, and practical significance

The empirical evaluation is conducted on four DTGB/DyTAG benchmarks: GDELT for political events/relations, Enron for email communication, Googlemap for user-business reviews, and Stack_elec for Stack Exchange electronics dialogues (Wang et al., 23 Sep 2025). The main task is destination node retrieval in both transductive and inductive settings, with Hit@10 as the primary metric and Hit@1 and Hit@3 shown in some figures. The paper also reports future link prediction using AP and ROC-AUC (Wang et al., 23 Sep 2025).

The main result is that DyGRASP, particularly with DyGFormer or TGAT as the temporal GNN backbone, consistently outperforms all baselines. The headline figure is up to 34% improvement in Hit@10 for destination node retrieval. The paper specifies that the largest gain appears on the Googlemap dataset in the transductive setting for DyGRASP(DyGFormer), where Hit@10 increases from 51.32 to 85.88, a gain of +34.56 points (Wang et al., 23 Sep 2025). The results also indicate that DyGRASP beats both classical temporal GNNs and the strong LLM baseline Llama-3.1-8B-Instruct, that gains are especially strong in the inductive setting, and that larger gains occur on datasets with richer textual content.

The ablation studies remove the global semantic layer (-Global), the recent semantic layer (-Recent), or both (-Recent-Global). The reported findings are that either recent or global reasoning alone improves over the base model, while combining both performs best; the two semantic types are therefore complementary (Wang et al., 23 Sep 2025). Additional studies vary the number of global segments $u$2 and the window length $u$3. Increasing $u$4 generally improves performance but increases inference time. Increasing $u$5 helps until the window becomes too large, after which LLM performance drops because long inputs hurt reasoning quality.

The paper describes DyGRASP as showing strong generalization across different temporal GNNs and LLMs. Concretely, it is tested with TGAT, DyGFormer, and multiple LLM families including Llama-3.1, Qwen2.5, and Mistral, and it consistently improves the underlying temporal GNN regardless of the specific LLM or GNN used (Wang et al., 23 Sep 2025). This supports the interpretation of DyGRASP as a modular semantic reasoning layer. Its stated practical takeaways are to model temporal text at both local/recent and long-term/global scales, avoid naive edge-wise LLM processing in favor of node-centric windowed reasoning, and use LLM reasoning to complement rather than replace temporal GNNs. The method is presented as especially suitable for dynamic link prediction and destination retrieval in systems such as e-commerce, email, or Q&A graphs, where interaction text is informative (Wang et al., 23 Sep 2025).