When Graph Tokens Sink: A Mechanistic Analysis of Graph Language Models

Abstract: Graph LLMs (GLMs) have become a promising direction for adapting LLMs to graph learning tasks. By transforming graph topology and node information into graph tokens, GLMs allow LLMs to jointly process structured graph inputs and textual instructions. Yet, it remains unclear how LLMs internally interpret these graph tokens and whether graph tokens act as meaningful carriers of graph structure. In this work, we analyze how LLMs process graph information through graph-token behavior in representative GLM architectures. Findings. We find that the internal saliency of graph tokens in GLMs is not equivalent to graph information utilization. Graph sink tokens consistently emerge as activation-level outliers: they can be identified by massive activation values along a small set of hidden-state dimensions and are biased toward early graph-token positions. However, this activation-level saliency does not imply that these tokens are the main carriers of graph information. Unlike classical attention sinks in language and vision-LLMs, graph sink tokens do not necessarily attract the largest attention weights from query tokens. Through pruning, repositioning, and swapping interventions, we show that graph sink tokens are not the most important semantic or structural tokens for downstream prediction. Implications. Together, these results suggest that after current GLMs map graph structure into the LLM token space, the resulting graph-token representations do not naturally form a fully usable topology-aware internal representation; instead, they exhibit a decoupling between activation-level saliency and graph-semantic utility. This decoupling points to limitations in existing graph-token construction, placement, and alignment mechanisms.

• The paper demonstrates that graph sink tokens, defined by high activation outliers, do not carry critical semantic or structural information in GLMs.
• The paper shows that intervention experiments like pruning and token swapping result in minimal performance changes, questioning the tokens' functional roles.
• The paper employs logit lens analysis to reveal that sink tokens decode into generic domain tokens, underlining the disconnect between activation saliency and meaningful information.

Mechanistic Analysis of Graph Sink Tokens in Graph LLMs

Introduction

The proliferation of Graph LLMs (GLMs), which hybridize LLM architectures and graph-structured data representations, has presented both empirical advances and interpretability challenges in the machine learning community. By transforming graph topology and node attributes into discrete or learned graph tokens, GLMs enable transformers to process mixed sequences of structured graph and textual input. Despite improvements in benchmark performance, the mechanistic behavior of these models with respect to their internal graph-token processing remains insufficiently understood, particularly the so-called "sink" behavior—i.e., the emergence of tokens that become salient primarily in the activation space rather than by information content. This paper delivers an in-depth mechanistic analysis of graph sink tokens in two representative GLM designs: LLaGA and TEA-GLM, using a battery of interpretability and intervention methodologies.

Emergence and Localization of Graph Sink Tokens

Sink tokens in the context of GLMs are operationally defined as those displaying massive activations on a sparse collection of hidden-state dimensions. The analysis across datasets (Cora, Arxiv, PubMed) and models demonstrates a consistent pattern:

• Activation outliers are dominated by a small set of sink dimensions (e.g., dimensions 1512, 2533 in LLaMA-based GLMs).
• Graph sink tokens are strongly biased toward initial graph-token positions and are frequently populated by non-informative tokens, such as [PAD] in LLaGA.
• These strong activation spikes are stable across tasks (node classification, link prediction) and persist across multiple GLM architectures, as confirmed by findings in the main study as well as additional results on InstructGLM in the appendix.

This suggests that the mapping of graph structure into the LLM token space triggers model-internal architectural phenomena reminiscent of, but distinct from, classical attention sinks within language or multimodal transformers.

Decoupling of Activation Saliency and Attention Routing

Despite their internal prominence, graph sink tokens do not reliably receive dominant attention mass. Empirical analysis of average and layer-wise attention weights reveals:

• In TEA-GLM, query-to-graph attention is focused on later token positions, not the early sink-token slots where the largest activation spikes are observed.
• In LLaGA, some attention is routed to the sink region due to stable vertical attention bands, but these positions largely correspond to [PAD] tokens and do not reflect meaningful graph structure.
• There is no strict coupling between activation-level saliency and query-to-graph attention, corroborating findings from recent work on activation and attention sinks in other modalities (Sun et al., 5 Mar 2026).

Thus, GLMs exhibit a divergence between massive activation and attention-based information flow, implying that activation prominence alone is an unreliable diagnostic for semantic or topological informativeness of graph-token representations.

Functional Role of Graph Sink Tokens: Token-Level Interventions

A comprehensive suite of intervention experiments (pruning, swapping, repositioning) was conducted to ascertain whether graph sink tokens actually carry critical semantic or structural information for downstream tasks:

• Pruning the highest-magnitude sink tokens yields negligible impact on task performance (node classification and link prediction) in both LLaGA and TEA-GLM. In many cases, removing randomly selected non-sink tokens yields greater performance degradation.
• Repositioning or swapping sink tokens with non-sink tokens also results in marginal changes, indicating weak contextual importance of activation-salient tokens.
• Sink behavior in LLaGA is tied to prompt sparsity: higher attention to sink tokens occurs as the proportion of non-padded (informative) tokens decreases.

This demonstrates that activation-level outliers, i.e., graph sink tokens, are not the main semantic or structural carriers for GLM performance; they are largely architectural artifacts of the mapping and prompt design pipeline.

Mechanistic Interpretability: Logit Lens Analysis

To further elucidate whether graph sink tokens encode task-relevant or topology-aware information, a logit lens analysis was performed, decoding hidden states of graph tokens to the base LLM vocabulary:

• Sink-token states consistently decode into generic domain-level tokens (e.g., "paper" in citation graphs), with low decoding confidence, rather than reflective of graph topology, node class, or task labels.
• These patterns are robust across datasets (Cora, Arxiv, PubMed) and persist through multiple transformer layers.

This provides mechanistic evidence that GLMs, even with orchestrated graph-token construction and alignment, do not fully translate graph structure into interpretable, task-relevant semantic content in hidden states—activation-level outliers do not exhibit deep semantic alignment with the underlying graph.

Implications for GLM Design and Future Directions

These findings expose critical limitations in the prevailing methodologies for adapting LLMs to graph-structured data:

• The decoupling of activation saliency, attention weights, and semantic/topological information challenges the use of activation-based heuristics or attention analysis as proxies for interpretability in GLMs.
• Architectural design choices such as token alignment, placement, padding strategies, and graph-to-token mapping emerge as key factors influencing the internal representations but do not guarantee usable or interpretable topology-aware encodings.

Future research must not only refine graph-token construction and alignment but also develop rigorous mechanisms for ensuring that structural graph information is preserved and utilized within the LLM bottleneck, possibly through specialized architectural constraints, alignment objectives, or modified pretraining pipelines.

Conclusion

In summary, this paper presents systematic, mechanistic evidence that graph sink tokens—activation-level outliers induced by standard graph-token mapping procedures in GLMs—are not reliable carriers of graph-structural or semantic information. They do not dictate attention routing nor are they critical for model prediction, and their decoded representations are uninformative and domain-generic. These results underscore the need for rethinking tokenization and representation alignment in graph-based prompting for transformer architectures, opening concrete avenues for the principled design and interpretability of future GLMs.