Paper deep dive
IRIS: Reusable Identity Representations from Frozen LLMs for Entity Alignment
Xinran Liu, Shengtao Li, Shouqian Shi, Ge Wang, Xin-Wei Yao
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Entity alignment (EA) identifies entities across knowledge graphs (KGs) that refer to the same real-world object. Conventional EA methods mainly exploit explicit graph structures and textual fields, which often provide insufficient semantic understanding to recognize the same entity under heterogeneous descriptions and distinguish it from semantically similar entities. Although large language models (LLMs) offer deeper entity understanding, existing LLM-based EA methods largely use this capability for auxiliary generation or candidate-conditioned decisions. Consequently, such understanding is not distilled into a stable and directly comparable identity space, leaving alignment tied to specific KG pairs or candidate sets and requiring repeated processing as the matching context changes. To address these limitations, we propose IRIS (Identity Representations from Internal States), a training-free framework that constructs for each entity an iris-like signature encoding its distinctive and stable identity characteristics. IRIS derives these signatures by eliciting identity-oriented contextual representations from a frozen LLM, thereby forming a shared space in which each entity is encoded once and can be aligned across different KGs through direct similarity comparison, without pair-dependent representation construction or candidate-wise LLM inference. Across four established EA benchmarks and two frozen LLM backbones, the best IRIS variants achieve Hits@1 scores of 100.00, 99.38, 98.31, and 97.99 on D-Y-15K V2, DBP-WIKI, ICEWS-WIKI, and ICEWS-YAGO, respectively.
Tags
Links
- Source: https://arxiv.org/abs/2607.25579v1
- Canonical: https://arxiv.org/abs/2607.25579v1
Trouble viewing inline? Open PDF directly →
Full Text
46,557 characters extracted from source content.
Expand or collapse full text
IRIS: Reusable Identity Representations from Frozen LLMs for Entity Alignment Xinran Liu*, Shengtao Li*, Shouqian Shi†, Ge Wang, Xin-Wei Yao *Equal contribution. †Corresponding author: sqlite@nju.edu.cn Abstract Entity alignment (EA) identifies entities across knowledge graphs (KGs) that refer to the same real-world object. Conventional EA methods mainly exploit explicit graph structures and textual fields, which often provide insufficient semantic understanding to recognize the same entity under heterogeneous descriptions and distinguish it from semantically similar entities. Although large language models (LLMs) offer deeper entity understanding, existing LLM-based EA methods largely use this capability for auxiliary generation or candidate-conditioned decisions. Consequently, such understanding is not distilled into a stable and directly comparable identity space, leaving alignment tied to specific KG pairs or candidate sets and requiring repeated processing as the matching context changes. To address these limitations, we propose IRIS (Identity Representations from Internal States), a training-free framework that constructs for each entity an iris-like signature encoding its distinctive and stable identity characteristics. IRIS derives these signatures by eliciting identity-oriented contextual representations from a frozen LLM, thereby forming a shared space in which each entity is encoded once and can be aligned across different KGs through direct similarity comparison, without pair-dependent representation construction or candidate-wise LLM inference. Across four established EA benchmarks and two frozen LLM backbones, the best IRIS variants achieve Hits@1 scores of 100.00, 99.38, 98.31, and 97.99 on D-Y-15K V2, DBP-WIKI, ICEWS-WIKI, and ICEWS-YAGO, respectively. Introduction Entity alignment (EA) identifies entities across knowledge graphs (KGs) that refer to the same real-world object and is fundamental to knowledge integration. Existing methods learn comparable representations from graph structures, attributes, and textual descriptions through knowledge graph embeddings Chen et al. (2017); Sun et al. (2017, 2018), graph neural networks Wang et al. (2018); Cao et al. (2019); Wu et al. (2019); Sun et al. (2020a); Mao et al. (2020, 2021), or multi-view and interaction-based methods Zhang et al. (2019); Tang et al. (2020). Reliable EA, however, requires more than matching explicit graph and textual patterns. The same entity may appear under different languages, abbreviations, naming conventions, or incomplete descriptions, while different entities may share similar names, types, and relational contexts. Conventional EA methods often lack the semantic depth needed to recognize the same entity across heterogeneous descriptions and distinguish it from semantically similar entities. Large language models (LLMs) provide stronger semantic understanding and have recently been introduced into EA for entity description generation, information augmentation, and candidate reasoning. ChatEA performs candidate retrieval followed by LLM-based reasoning Jiang et al. (2024a); LLM4EA uses LLM-generated annotations with active learning Chen et al. (2024); HLMEA combines candidate filtering with LLM selection Jin et al. (2025); and EasyEA integrates LLM-based summarization into a multi-stage alignment framework Cheng et al. (2025). These methods demonstrate the value of LLMs for interpreting heterogeneous entity information. Nevertheless, existing LLM-based methods mainly use this understanding for auxiliary generation or candidate-conditioned decisions. The resulting semantic knowledge is not retained as a stable and directly comparable identity representation. Consequently, representation construction or matching remains tied to specific KG pairs, cross-graph interactions, or candidate sets. Changes in the counterpart KG or candidate context may therefore require repeated representation learning, interaction, or LLM inference, limiting both representation reuse and alignment efficiency. To address these limitations, we propose IRIS, a training-free framework that constructs for each entity an iris-like signature encoding its distinctive and stable identity characteristics. IRIS directs the contextual representations of a frozen LLM toward entity identity and converts its semantic understanding into a reusable vector representation. Like an iris used for identification, each signature is designed to distinguish one entity from others while remaining stable across heterogeneous surface realizations. Each IRIS signature is independently derived from evidence in the entity’s resident KG, without access to a counterpart graph or candidate set. Entity representations can therefore be computed once, placed in a shared comparison space, and reused across different KG combinations. Alignment is subsequently performed through direct vector similarity, avoiding pair-dependent representation construction and candidate-wise LLM inference. The main contributions are summarized as follows: • We propose IRIS, an identity-oriented representation elicitation framework that converts the contextual understanding of a frozen large language model into stable and discriminative entity signatures, without entity-alignment supervision or parameter updates. • IRIS reformulates entity alignment as independent entity encoding followed by direct vector comparison. Each entity is encoded once from its resident knowledge graph and reused across different graph combinations, avoiding pair-dependent representation construction and repeated candidate-wise LLM inference. • Extensive experiments on D-Y-15K V2, DBP-WIKI, ICEWS-WIKI, and ICEWS-YAGO demonstrate the effectiveness and generality of IRIS. The best IRIS variants achieve Hits@1 scores of 100.00, 99.38, 98.31, and 97.99, respectively, outperforming the strongest compared LLM-based baseline by 10.31 and 4.49 percentage points on the two ICEWS benchmarks. Related Work Entity Representation Learning for Entity Alignment Entity representation learning is central to entity alignment. Early embedding-based methods learn cross-KG comparable representations from relational triples and seed alignments Chen et al. (2017); Sun et al. (2017, 2018). Graph-based approaches further aggregate neighborhood and relation information through graph neural networks Wang et al. (2018); Cao et al. (2019); Wu et al. (2019); Sun et al. (2020a); Mao et al. (2020, 2021). Subsequent methods combine structural information with entity names and attributes Zhang et al. (2019), or directly model interactions between candidate entity pairs Tang et al. (2020). These methods progressively enrich entity representations, but their cross-graph comparability is generally established within a predefined alignment task through seed supervision, joint optimization, or cross-graph interaction. IRIS instead constructs an identity representation for each entity independently from its own KG and performs cross-graph alignment only after representation construction. LLM-Enhanced Entity Alignment Large language models have recently been introduced into EA to provide semantic knowledge beyond explicit KG evidence. LLM4EA obtains LLM-generated alignment annotations through active learning and mitigates annotation noise Chen et al. (2024). EasyEA uses LLM-based summarization together with representation fusion and candidate selection Cheng et al. (2025). ChatEA retrieves candidate entities and applies conversational reasoning to refine alignment decisions Jiang et al. (2024a), while HLMEA combines candidate filtering with LLM-based selection Jin et al. (2025). These methods generally integrate LLM outputs into task-specific alignment pipelines rather than explicitly constructing reusable identity representations. In contrast, IRIS directly extracts identity-oriented contextual representations from a frozen LLM and preserves them as reusable entity signatures independent of counterpart entities and candidate sets. Decoder-Only LLMs for Representation Learning Recent studies have demonstrated that decoder-only LLMs can be repurposed for representation learning. SGPT develops pooling and fine-tuning strategies for semantic search with GPT-style models Muennighoff (2022). Echo Embeddings improves causal representations through repeated input construction Springer et al. (2025). LLM2Vec adapts decoder-only LLMs through attention modification and representation-oriented training BehnamGhader et al. (2024), while NV-Embed develops a specialized general-purpose embedding model Lee et al. (2025). These methods primarily target general textual similarity. Entity alignment additionally requires representations that remain stable across heterogeneous descriptions of the same entity while separating different entities with similar semantics. IRIS addresses this identity-specific requirement through structured KG contextualization and semantic readout, without modifying or optimizing the frozen LLM. The representation quality of an LLM also varies across Transformer layers. Previous work has studied intrinsic dimension, anisotropy, and expansion–compression behavior in Transformer representations Valeriani et al. (2023); Razzhigaev et al. (2024). IRIS uses intrinsic dimension as a label-free signal to select a fixed readout layer window for each backbone. Method Problem Setup Consider a collection of K knowledge graphs, denoted by =G(k)k=1KG=\G^(k)\_k=1^K. Each graph G(k)=(ℰ(k),ℛ(k),(k))G^(k)=(E^(k),R^(k),T^(k)) contains a set of entities, relations, and relational triples. Given an entity e∈ℰ(k)e ^(k), IRIS constructs its identity representation ez_e using only the information associated with e in its resident KG. This construction does not access another KG, cross-graph candidates, or the source–target role of e in a particular alignment task. Because entity representations constructed from different KGs share a common comparison space, each representation can be retained and reused when the counterpart KG changes. Overview IRIS constructs a reusable identity signature for each entity in three stages, as illustrated in Figure 1. First, it organizes the entity and its local KG evidence into a structured textual context and applies contextual identity completion to recover a complementary full English name and entity type. Second, it constructs multiple name views and independently extracts identity-oriented representations from each retained view using a frozen LLM. Third, it aggregates the selected token and layer representations and fuses the resulting view representations into a unified entity signature. Since each entity is encoded independently using only evidence from its resident KG, the resulting signatures can be cached and directly compared across different KGs. Figure 1: Overview of IRIS. (a) For each entity, IRIS independently organizes identity evidence from its resident KG and applies contextual identity completion to obtain a complementary full English name and entity type. (b) The original name, available aliases, and contextually completed name form multiple identity views. Each retained view is processed by an identity-oriented prompt, followed by intrinsic-dimension-based layer selection and token- and layer-aware aggregation. The resulting view representations are fused into a reusable entity signature. (c) Independently constructed signatures are stored in a shared representation space and used for cross-KG entity alignment through direct similarity comparison. The vector index illustrates a scalable retrieval implementation; all reported experiments use exact cosine-similarity ranking. Structured Local KG Context Construction IRIS organizes the information associated with each entity into a structured textual context, including its name, type, properties, and local relations. Each entity is represented as (name, type, properties), where properties denotes the key–value pairs included in the entity record. Together, these fields form a structured description of the entity. Each outgoing relation is represented as this node -[relation]-> target, where target follows the same entity format. This serialization preserves relation direction and incorporates the neighboring entity’s name, type, and properties into the local context. All information is collected solely from the entity’s resident KG. Contextual Identity Completion The same entity may be described using different languages, incomplete names, or abbreviations. IRIS therefore uses the structured local context to recover a complementary full English name and type while retaining the original KG information. Given the serialized entity and its local relations, IRIS adopts the following generation template: Focus on entity deduplication and disambiguation. Entity: entity Relationships: relations Determine the entity’s full English name and type. Answer: "[full name]gj" with type: [type]gj. Here, entity denotes the structured description of the target entity, and relations contains its serialized local relations. The underlined fields indicate content generated by the frozen LLM. IRIS first generates the full English name, then appends the fixed phrase " with type: and continues generating the type from the same KV cache. The completed name and type supplement the original KG fields with additional identity cues while preserving the information provided by the KG. Discriminative Multi-View Representation Extraction Complementary Name-View Construction. IRIS retains the original KG name as the primary view because it is directly grounded in the source data. Available aliases and the contextually completed full English name are treated as auxiliary views. These name variants may express the same entity through translations, expanded abbreviations, more complete names, or alternative word orders, thereby providing complementary descriptions of its identity. To avoid introducing redundant views, IRIS compares each auxiliary name with the original name using character-level edit distance. Variants whose distance does not exceed one third of the shorter name length are removed, and only sufficiently distinct auxiliary names are retained for subsequent representation extraction. Identity-Oriented Semantic Readout. For each retained name view, IRIS constructs the following representation prompt: Focus on entity deduplication and disambiguation in Knowledge Graph. The entity: entity. Context: relations What is the core semantics of the entity entity? The core semantics of entity "label" (type) The label field is instantiated with the current name view, while type contains the corresponding entity type. The final label–type expression follows the complete entity context. Under causal attention, its hidden states incorporate the preceding entity description and local relations while remaining explicitly anchored to the identity being represented. IRIS extracts the post-MLP hidden states corresponding to the label and type tokens at the outputs of the selected Transformer blocks. Quotation marks and other fixed delimiters are excluded from the readout. This design concentrates the representation on an explicit entity-specific semantic target. Token- and Layer-Aware Aggregation. The label and type tokens contribute differently to entity representation. IRIS therefore applies a progressive position-weighting scheme to the label tokens. Regular label tokens receive linearly increasing weights according to their relative positions, allowing the complete name to be preserved while reducing the dominance of generic components appearing earlier in the sequence. Label tokens containing special symbols are assigned a smaller fixed weight to suppress formatting-related noise, while type tokens receive another fixed weight because they provide complementary categorical information. Let e,vnP_e,v^n, e,vsP_e,v^s, and e,vtP_e,v^t denote the regular label tokens, special-symbol label tokens, and type tokens, respectively. Suppose that the regular label contains m tokens. For each selected token position j, let rj∈0,…,m−1r_j∈\0,…,m-1\ denote its zero-based position among the regular label tokens when j∈e,vnj _e,v^n. Its weight is defined as wj=w0+αrjm,j∈e,vn,wsym,j∈e,vs,wtype,j∈e,vt.w_j= casesw_0+α r_jm,&j _e,v^n,\\ w_sym,&j _e,v^s,\\ w_type,&j _e,v^t. cases (1) Here, w0w_0 is the base weight assigned to the first regular label token, α controls the progressive increase across the label, and wsymw_sym and wtypew_type are the fixed weights assigned to special-symbol and type tokens. The quality of the extracted representations also varies across Transformer layers. IRIS selects the readout depth using a label-free intrinsic-dimension criterion. For each frozen backbone, it estimates a layer-wise intrinsic-dimension trajectory once on an unlabeled calibration corpus using a trimmed TwoNN estimator Facco et al. (2017). Following the expansion–compression pattern observed in Transformer representations Valeriani et al. (2023); Razzhigaev et al. (2024), IRIS selects the first local minimum after the initial expansion and aggregates a small window of neighboring layers. The selected layer window is fixed for each backbone and reused across all EA datasets. Let e,v=e,vn∪e,vs∪e,vtP_e,v=P_e,v^n _e,v^s _e,v^t denote all selected label and type token positions, and let ℒL denote the selected layer window. The representation of entity e under name view v is computed as e,v=Norm(1|ℒ|∑ℓ∈ℒ∑j∈e,vwjjℓ∑j∈e,vwj),z_e,v=Norm ( 1|L| _ _j _e,vw_jh_j _j _e,vw_j ), (2) where jℓh_j denotes the post-MLP hidden state of token j at the output of Transformer layer ℓ , and Norm(⋅)Norm(·) denotes ℓ2 _2 normalization. The concrete token weights and selected layer windows are provided in the implementation details. Multi-View Representation Fusion. The original KG name provides a representation directly grounded in the entity record, while the retained auxiliary names contribute complementary name information. IRIS first averages the representations of all retained auxiliary views and then combines the result with the original-name representation: e=Norm((1−λ)e,o+λ|e+|∑v∈e+e,v),z_e=Norm ((1-λ)z_e,o+ λ|V_e^+| _v _e^+z_e,v ), (3) where e,oz_e,o denotes the representation extracted from the original KG name, e+V_e^+ is the set of retained auxiliary name views, and λ controls their contribution. When no auxiliary view is retained, IRIS directly uses e,oz_e,o. Unified Similarity-Based Alignment Each IRIS representation is constructed independently using only information from the entity’s resident KG. Representations from different KGs can therefore be compared directly in a shared representation space. Since all entity representations are ℓ2 _2-normalized, the alignment score between entities eie_i and eje_j is computed as s(ei,ej)=ei⊤ej.s(e_i,e_j)=z_e_i z_e_j. (4) For standard EA evaluation, the similarity scores between the two entity sets are used to rank candidate counterparts. Representation construction requires no cross-graph retrieval, pair-dependent prompt construction, or candidate-wise LLM inference. Once an entity has been encoded, its representation can be cached and reused when aligning the resident KG with other KGs. Experiments Method D-Y-15K V2 DBP-WIKI ICEWS-WIKI ICEWS-YAGO Hits@1 Hits@10 MRR Hits@1 Hits@10 MRR Hits@1 Hits@10 MRR Hits@1 Hits@10 MRR AlignE 86.4 97.0 90.2 56.6 82.7 65.5 5.7 26.1 12.2 1.9 11.8 5.5 BootEA 95.0 98.6 96.3 74.8 89.8 80.1 7.2 27.5 13.9 2.0 12.0 5.6 GCN-Align 82.6 94.9 87.2 49.4 75.6 59.0 4.6 18.4 9.3 1.7 8.5 3.8 RDGCN 85.4 93.2 88.3 97.4 99.4 98.0 6.4 20.2 9.6 2.9 9.7 4.2 Dual-AMN 97.5 99.3 98.1 98.3 99.6 99.1 8.3 28.1 14.5 3.1 14.4 6.8 LLM4EA 97.9 99.6 98.5 – – – – – – – – – BERT-INT – – – 99.6 99.7 99.6 56.1 70.0 60.7 75.6 85.9 79.3 Simple-HHEA – – – 97.5 99.1 98.8 72.0 87.2 75.4 84.7 91.5 87.0 ChatEA – – – 99.5 100.0 99.8 88.0 94.5 91.2 93.5 95.5 94.4 IRIS-Llama 3.1-8B 100.00 100.00 100.00 99.38 99.89 99.60 98.31 99.26 98.70 97.92 99.05 98.34 IRIS-Qwen 2.5-7B 100.00 100.00 100.00 99.34 99.93 99.59 98.31 99.36 98.67 97.99 99.05 98.37 Table 1: Main results on four entity alignment benchmarks. Scores are percentages. Baseline results on D-Y-15K V2 are taken from the unified GPT-4 annotation setting reported by LLM4EA Chen et al. (2024), while baseline results on DBP-WIKI, ICEWS-WIKI, and ICEWS-YAGO are taken from the unified evaluation reported by ChatEA Jiang et al. (2024a). Boldface and underlining denote the best and second-best results, respectively. Results tied at the same rank receive the same formatting. Baseline scores retain the numerical precision reported in their corresponding sources, while IRIS results are reported with two decimal places. A dash indicates that the corresponding result was not reported under the same benchmark setting. Experimental Setup Datasets. We evaluate IRIS on four established entity alignment benchmarks: D-Y-15K V2, DBP-WIKI, ICEWS-WIKI, and ICEWS-YAGO. D-Y-15K V2 is the dense OpenEA benchmark constructed from DBpedia and YAGO. We use the same dataset version and evaluation setting as LLM4EA Chen et al. (2024). DBP-WIKI aligns entities between DBpedia and Wikidata, while ICEWS-WIKI and ICEWS-YAGO align event-oriented entities from ICEWS with Wikidata and YAGO, respectively. For these three benchmarks, we follow the dataset versions and evaluation settings used by ChatEA Jiang et al. (2024a). The two ICEWS benchmarks exhibit substantially greater structural and descriptive heterogeneity than DBP-WIKI. Evaluation Metrics. We report Hits@1, Hits@10, and mean reciprocal rank (MRR), expressed as percentages. Hits@k measures the proportion of source entities whose correct counterparts appear among the top-k ranked target entities. MRR averages the reciprocal rank of the correct counterpart and evaluates the overall ranking quality. Since Hits@1 directly measures whether the correct counterpart is ranked first, we use it as the primary metric in the ablation studies. Baselines. We compare IRIS with representative embedding-based, graph-based, interaction-based, and LLM-enhanced EA methods. The compared methods include AlignE Sun et al. (2020b), BootEA Sun et al. (2018), GCN-Align Wang et al. (2018), RDGCN Wu et al. (2019), Dual-AMN Mao et al. (2021), BERT-INT Tang et al. (2020), Simple-HHEA Jiang et al. (2024b), LLM4EA Chen et al. (2024), and ChatEA Jiang et al. (2024a). For D-Y-15K V2, baseline results are taken from the unified GPT-4 annotation setting reported by LLM4EA Chen et al. (2024). For DBP-WIKI, ICEWS-WIKI, and ICEWS-YAGO, baseline results are taken from the unified evaluation reported by ChatEA Jiang et al. (2024a). Results unavailable under the corresponding benchmark setting are marked with a dash. Implementation Details. For the main experiments, we instantiate IRIS with Llama 3.1-8B-Instruct Grattafiori et al. (2024) and Qwen 2.5-7B-Instruct Qwen Team (2025) as frozen decoder-only backbones. For brevity, they are denoted as Llama 3.1-8B and Qwen 2.5-7B, respectively, in all tables. No entity alignment labels are used to update the model parameters. To examine whether the effects of individual components generalize across model families and scales, the ablation studies additionally include Qwen 3.5-4B Qwen Team (2026). For each entity, IRIS organizes its name, type, properties, and outgoing relations using the formats described in the method section. The contextual identity completion stage generates a complementary full English name and entity type from the local KG context. The original KG name is always retained as the primary name view, while available aliases and the contextually completed full English name are treated as auxiliary views. To avoid repeatedly encoding nearly identical name forms, IRIS computes the character-level Levenshtein distance d(a,b)d(a,b) between the original name a and each auxiliary name b. The auxiliary name is retained only when d(a,b)>⌊min(|a|,|b|)/3⌋d(a,b)> (|a|,|b|)/3 ; otherwise, it is removed as a near-duplicate. The same criterion is used across all datasets and backbones. For each retained name view, IRIS extracts post-MLP hidden states from the label and type tokens at the outputs of the selected Transformer blocks. Quotation marks and other fixed delimiters are excluded from the readout. Let the regular label tokens be indexed by i=0,…,m−1i=0,…,m-1. Their weights are defined as wilabel=1.0+0.5im.w_i^label=1.0+0.5 im. (5) A label token containing special symbols is assigned a weight of 0.250.25, while each type token receives a weight of 0.750.75. The selected token representations are then combined through weighted averaging. For each backbone, IRIS computes its layer-wise intrinsic-dimension trajectory once using unlabeled WikiText-103 sequences Merity et al. (2017). At each Transformer layer, key-projection states are mean-pooled over valid token positions to obtain sequence representations, and their intrinsic dimension is estimated using a trimmed TwoNN estimator Facco et al. (2017). The key-projection states are used only for label-free layer selection, whereas the final entity representations are extracted from the post-MLP hidden states at the selected layers. We compare four layer windows: the final layers (L1L_1), the global minimum-ID region (L2L_2), the global maximum-ID region (L3L_3), and the first local minimum after the initial ID expansion (L4L_4). Using one-based Transformer-layer indices, the respective windows are [30,31,32][30,31,32], [24,25,26][24,25,26], [12,13,14][12,13,14], and [1,2,3,4][1,2,3,4] for Llama 3.1-8B; [26,27,28][26,27,28], [24,25,26][24,25,26], [18,19,20][18,19,20], and [2,3,4,5][2,3,4,5] for Qwen 2.5-7B; and [30,31,32][30,31,32], [1,2,3][1,2,3], [18,19,20][18,19,20], and [3,4,5][3,4,5] for Qwen 3.5-4B. IRIS uses L4L_4, whose window is fixed for each backbone and reused across all EA datasets. When multiple auxiliary name views are retained, their normalized representations are first averaged. IRIS then combines the resulting auxiliary representation with the original-name representation using equal weights: e=Norm(12e,o+12e,a),z_e=Norm ( 12z_e,o+ 12z_e,a ), (6) where e,oz_e,o denotes the original-name representation and e,az_e,a denotes the average representation of the retained auxiliary views. When no auxiliary view is retained, e,oz_e,o is used directly. All final entity representations are ℓ2 _2-normalized. For each benchmark, we compute the complete cross-KG similarity matrix using cosine similarity and rank all target entities for each source entity. IRIS does not use a separate candidate-generation stage, reranking, pair-dependent prompts, or candidate-conditioned LLM inference. Main Results Table 1 reports the main results on the four benchmarks. IRIS achieves the best Hits@1 performance on D-Y-15K V2, ICEWS-WIKI, and ICEWS-YAGO, while remaining competitive on the nearly saturated DBP-WIKI benchmark. On D-Y-15K V2, both IRIS-Llama 3.1-8B and IRIS-Qwen 2.5-7B achieve 100.00 on Hits@1, Hits@10, and MRR. Compared with LLM4EA, the strongest reported baseline under the same benchmark setting, IRIS improves Hits@1 by 2.10 percentage points. This result demonstrates that independently constructed identity representations can support highly accurate alignment without entity-alignment supervision or candidate-conditioned LLM inference. The advantages of IRIS are particularly pronounced on the more heterogeneous ICEWS benchmarks. On ICEWS-WIKI, both IRIS variants achieve 98.31 Hits@1, outperforming ChatEA by 10.31 percentage points. IRIS-Qwen 2.5-7B reaches 99.36 Hits@10, while IRIS-Llama 3.1-8B obtains the highest MRR of 98.70. On ICEWS-YAGO, IRIS-Qwen 2.5-7B achieves 97.99 Hits@1 and 98.37 MRR, exceeding ChatEA by 4.49 and 3.97 percentage points, respectively. On DBP-WIKI, IRIS-Llama 3.1-8B obtains 99.38 Hits@1 and 99.60 MRR, while IRIS-Qwen 2.5-7B obtains 99.34 Hits@1 and 99.93 Hits@10. Although this benchmark is already nearly saturated, IRIS remains within 0.22 percentage points of the strongest Hits@1 result while avoiding alignment-specific parameter training and candidate-wise LLM reasoning. The two frozen backbones exhibit highly consistent performance. Their Hits@1 scores differ by at most 0.07 percentage points across the four benchmarks. This consistency indicates that the effectiveness of IRIS is not restricted to a particular frozen LLM backbone. Ablation Studies We conduct controlled ablations to examine three components of IRIS: token-aware weighting, contextual identity completion, and intrinsic-dimension-based readout-layer selection. The experiments use Llama 3.1-8B, Qwen 2.5-7B, and Qwen 3.5-4B. All ablation studies report Hits@1 to provide a compact and consistent comparison. Within each ablation, all settings other than the evaluated component are held fixed. For each backbone, the “With” columns and L4L_4 correspond to the same complete IRIS configuration. Consequently, the results for Llama 3.1-8B and Qwen 2.5-7B reproduce the Hits@1 scores reported in Table 1. Effects of Token-Aware Weighting and Contextual Identity Completion. For token-aware weighting, we replace the weighted aggregation in IRIS with uniform averaging over the same label and type tokens. For contextual identity completion, we remove the generated full English name and entity type, retaining only the original KG information and available aliases. The two column pairs in Table 2 are independent controlled comparisons. Dataset Backbone Token-Aware Weighting Contextual Identity Completion Without With Without With D-Y-15K V2 Llama 3.1-8B 100.00 100.00 99.31 100.00 Qwen 2.5-7B 100.00 100.00 99.25 100.00 Qwen 3.5-4B 100.00 100.00 99.27 100.00 DBP-WIKI Llama 3.1-8B 98.55 99.38 98.01 99.38 Qwen 2.5-7B 98.23 99.34 97.93 99.34 Qwen 3.5-4B 97.89 99.03 97.57 99.03 ICEWS-WIKI Llama 3.1-8B 98.03 98.31 97.51 98.31 Qwen 2.5-7B 98.03 98.31 97.47 98.31 Qwen 3.5-4B 97.61 98.09 96.95 98.09 ICEWS-YAGO Llama 3.1-8B 97.34 97.92 97.30 97.92 Qwen 2.5-7B 97.85 97.99 97.24 97.99 Qwen 3.5-4B 96.50 97.76 97.05 97.76 Table 2: Component ablations of token-aware weighting and contextual identity completion. “Without” and “With” indicate whether the corresponding component is excluded or included. The two column pairs represent independent controlled comparisons. Scores are Hits@1 percentages. The better result within each comparison is shown in bold; ties are both boldfaced. As shown in Table 2, token-aware weighting preserves the saturated results on D-Y-15K V2 and improves all nine dataset–backbone combinations on the remaining benchmarks, with gains of 0.14–1.26 percentage points. The largest gains for the two main backbones occur on DBP-WIKI, reaching 0.83 and 1.11 points for Llama 3.1-8B and Qwen 2.5-7B, respectively. These improvements indicate that uniformly averaging all readout tokens can dilute the contribution of informative name and type components. By assigning different weights according to token position and function, IRIS better preserves entity-specific identity cues while reducing the influence of formatting-related or less discriminative tokens. The consistent gains across backbones further suggest that this effect is not tied to a particular model architecture. Contextual identity completion improves all twelve dataset–backbone combinations by 0.62–1.46 points, with the largest gains of 1.37–1.46 points on DBP-WIKI. By recovering a more explicit English name and type from the resident-KG context, it supplements the original KG fields and aliases with complementary identity evidence. The consistent gains across datasets and backbones indicate that this additional information is particularly useful when the original KG fields and available aliases do not fully expose the entity’s identity. Effect of Readout-Layer Selection. To evaluate the intrinsic-dimension-based readout-layer selection strategy, we compare four layer aggregation schemes. L1L_1 uses a fixed window over the final Transformer layers. L2L_2 and L3L_3 use windows around the global minimum-ID and global maximum-ID layers, respectively. L4L_4, which is used by IRIS, selects a window around the first local minimum after the initial intrinsic-dimension expansion. The backbone-specific layer indices are provided in the implementation details. Dataset Backbone L1L_1 L2L_2 L3L_3 L4L_4 D-Y-15K V2 Llama 3.1-8B 99.81 94.27 93.27 100.00 Qwen 2.5-7B 99.99 99.95 97.93 100.00 Qwen 3.5-4B 98.87 100.00 92.48 100.00 DBP-WIKI Llama 3.1-8B 98.58 91.25 90.15 99.38 Qwen 2.5-7B 98.57 98.79 97.70 99.34 Qwen 3.5-4B 97.78 98.99 86.70 99.03 ICEWS-WIKI Llama 3.1-8B 96.04 88.25 84.21 98.31 Qwen 2.5-7B 95.84 95.88 88.91 98.31 Qwen 3.5-4B 94.62 97.95 81.86 98.09 ICEWS-YAGO Llama 3.1-8B 96.51 85.62 74.59 97.92 Qwen 2.5-7B 96.25 96.21 81.98 97.99 Qwen 3.5-4B 93.40 97.70 66.87 97.76 Table 3: Effect of readout-layer selection. L1L_1 uses a fixed window over the final Transformer layers; L2L_2 and L3L_3 use windows around the global minimum-ID and global maximum-ID layers, respectively; and L4L_4 selects a window around the first local minimum after the initial intrinsic-dimension expansion. L4L_4 is used by IRIS. Scores are Hits@1 percentages. The best result for each dataset–backbone combination is shown in bold; ties are both boldfaced. As shown in Table 3, L4L_4 achieves the highest or tied-highest Hits@1 across every evaluated dataset–backbone combination. The alternative strategies exhibit greater sensitivity to the backbone and dataset. Although L2L_2 performs strongly with Qwen 3.5-4B, it substantially degrades Llama 3.1-8B on DBP-WIKI and both ICEWS benchmarks. The global maximum-ID strategy L3L_3 performs particularly poorly on the heterogeneous ICEWS benchmarks. The final-layer strategy L1L_1 remains competitive in some settings but is consistently outperformed by L4L_4. These results support the layer-selection rule introduced in the method section. The first local minimum after the initial intrinsic-dimension expansion provides a more stable readout region than either the final-layer window or the global intrinsic-dimension extrema. Conclusion This paper presented IRIS, a training-free framework for constructing reusable entity identity representations from frozen decoder-only LLMs. IRIS organizes the information available in each entity’s resident KG into a structured context, applies contextual identity completion to recover complementary name and type cues, and extracts identity-oriented representations from selected post-MLP hidden states. It further combines multiple name views through token-aware aggregation and intrinsic-dimension-based layer selection, yielding a unified representation without entity-alignment supervision, parameter updates, or candidate-conditioned inference. By encoding each entity independently, IRIS decouples representation construction from subsequent cross-KG matching. The resulting vectors can therefore be cached, directly compared, and reused when the counterpart KG changes. Experiments on four benchmarks with Llama 3.1-8B and Qwen 2.5-7B demonstrate consistently strong performance, with particularly large improvements on the heterogeneous ICEWS datasets. The ablation results further verify the individual contributions of contextual identity completion, token-aware aggregation, and intrinsic-dimension-based readout-layer selection. Overall, these findings suggest that the contextual knowledge encoded in frozen LLMs can be elicited into a shared identity-oriented representation space, offering a practical and promising approach to accurate, efficient, and reusable entity alignment across knowledge graphs. References BehnamGhader et al. (2024) Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961, 2024. Cao et al. (2019) Yixin Cao, Zhiyuan Liu, Chengjiang Li, Zhiyuan Liu, Juanzi Li, and Tat-Seng Chua. Multi-channel graph neural network for entity alignment. In Anna Korhonen, David Traum, and Lluís Màrquez, editors, Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1452–1461, Florence, Italy, July 2019. Association for Computational Linguistics. doi: 10.18653/v1/P19-1140. URL https://aclanthology.org/P19-1140/. Chen et al. (2017) Muhao Chen, Yingtao Tian, Mohan Yang, and Carlo Zaniolo. Multilingual knowledge graph embeddings for cross-lingual knowledge alignment. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17, pages 1511–1517, 2017. doi: 10.24963/ijcai.2017/209. URL https://doi.org/10.24963/ijcai.2017/209. Chen et al. (2024) Shengyuan Chen, Qinggang Zhang, Junnan Dong, Wen Hua, Qing Li, and Xiao Huang. Entity alignment with noisy annotations from large language models. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems, volume 37, pages 15097–15120. Curran Associates, Inc., 2024. doi: 10.52202/079017-0482. URL https://proceedings.neurips.c/paper_files/paper/2024/file/1b57aaddf85ab01a2445a79c9edc1f4b-Paper-Conference.pdf. Cheng et al. (2025) Jingwei Cheng, Chenglong Lu, Linyan Yang, Guoqing Chen, and Fu Zhang. EasyEA: Large language model is all you need in entity alignment between knowledge graphs. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors, Findings of the Association for Computational Linguistics: ACL 2025, pages 20981–20995, Vienna, Austria, 2025. Association for Computational Linguistics. ISBN 979-8-89176-256-5. doi: 10.18653/v1/2025.findings-acl.1080. URL https://aclanthology.org/2025.findings-acl.1080/. Facco et al. (2017) Elena Facco, Maria d’Errico, Alex Rodriguez, and Alessandro Laio. Estimating the intrinsic dimension of datasets by a minimal neighborhood information. Scientific reports, 7(1):12140, 2017. Grattafiori et al. (2024) Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. Jiang et al. (2024a) Xuhui Jiang, Yinghan Shen, Zhichao Shi, Chengjin Xu, Wei Li, Zixuan Li, Jian Guo, Huawei Shen, and Yuanzhuo Wang. Unlocking the power of large language models for entity alignment. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7566–7583, Bangkok, Thailand, August 2024a. Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.408. URL https://aclanthology.org/2024.acl-long.408/. Jiang et al. (2024b) Xuhui Jiang, Chengjin Xu, Yinghan Shen, Yuanzhuo Wang, Fenglong Su, Zhichao Shi, Fei Sun, Zixuan Li, Jian Guo, and Huawei Shen. Toward practical entity alignment method design: Insights from new highly heterogeneous knowledge graph datasets. In Proceedings of the ACM Web Conference 2024, pages 2325–2336, 2024b. Jin et al. (2025) Xiongnan Jin, Zhilin Wang, Jinpeng Chen, Liu Yang, Byungkook Oh, Seung-won Hwang, and Jianqiang Li. HLMEA: Unsupervised entity alignment based on hybrid language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 11888–11896, 2025. doi: 10.1609/aaai.v39i11.33294. Lee et al. (2025) Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. In Y. Yue, A. Garg, N. Peng, F. Sha, and R. Yu, editors, International Conference on Learning Representations, volume 2025, pages 79310–79333, 2025. URL https://proceedings.iclr.c/paper_files/paper/2025/file/c4bf73386022473a652a18941e9ea6f8-Paper-Conference.pdf. Mao et al. (2020) Xin Mao, Wenting Wang, Huimin Xu, Yuanbin Wu, and Man Lan. Relational reflection entity alignment. In Proceedings of the 29th ACM international conference on information & knowledge management, pages 1095–1104, 2020. Mao et al. (2021) Xin Mao, Wenting Wang, Yuanbin Wu, and Man Lan. Boosting the speed of entity alignment 10 ×: Dual attention matching network with normalized hard sample mining. In Proceedings of the Web Conference 2021, W ’21, pages 821–832, New York, NY, USA, 2021. Association for Computing Machinery. ISBN 9781450383127. doi: 10.1145/3442381.3449897. URL https://doi.org/10.1145/3442381.3449897. Merity et al. (2017) Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id=Byj72udxe. Muennighoff (2022) Niklas Muennighoff. Sgpt: Gpt sentence embeddings for semantic search. arXiv preprint arXiv:2202.08904, 2022. Qwen Team (2025) Qwen Team. Qwen2.5 technical report, 2025. URL https://arxiv.org/abs/2412.15115. Qwen Team (2026) Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. URL https://qwen.ai/blog?id=qwen3.5. Razzhigaev et al. (2024) Anton Razzhigaev, Matvey Mikhalchuk, Elizaveta Goncharova, Ivan Oseledets, Denis Dimitrov, and Andrey Kuznetsov. The shape of learning: Anisotropy and intrinsic dimensions in transformer-based models. In Yvette Graham and Matthew Purver, editors, Findings of the Association for Computational Linguistics: EACL 2024, pages 868–874, St. Julian’s, Malta, March 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.findings-eacl.58. URL https://aclanthology.org/2024.findings-eacl.58/. Springer et al. (2025) Jacob Mitchell Springer, Suhas Kotha, Daniel Fried, Graham Neubig, and Aditi Raghunathan. Repetition improves language model embeddings. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=Ahlrf2HGJR. Sun et al. (2017) Zequn Sun, Wei Hu, and Chengkai Li. Cross-lingual entity alignment via joint attribute-preserving embedding. In International semantic web conference, pages 628–644. Springer, 2017. Sun et al. (2018) Zequn Sun, Wei Hu, Qingheng Zhang, and Yuzhong Qu. Bootstrapping entity alignment with knowledge graph embedding. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 4396–4402. International Joint Conferences on Artificial Intelligence Organization, 2018. doi: 10.24963/ijcai.2018/611. URL https://doi.org/10.24963/ijcai.2018/611. Sun et al. (2020a) Zequn Sun, Chengming Wang, Wei Hu, Muhao Chen, Jian Dai, Wei Zhang, and Yuzhong Qu. Knowledge graph alignment network with gated multi-hop neighborhood aggregation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 222–229, 2020a. Sun et al. (2020b) Zequn Sun, Qingheng Zhang, Wei Hu, Chengming Wang, Muhao Chen, Farahnaz Akrami, and Chengkai Li. A benchmarking study of embedding-based entity alignment for knowledge graphs. arXiv preprint arXiv:2003.07743, 2020b. Tang et al. (2020) Xiaobin Tang, Jing Zhang, Bo Chen, Yang Yang, Hong Chen, and Cuiping Li. Bert-int: A bert-based interaction model for knowledge graph alignment. In Christian Bessiere, editor, Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, pages 3174–3180. International Joint Conferences on Artificial Intelligence Organization, 7 2020. doi: 10.24963/ijcai.2020/439. URL https://doi.org/10.24963/ijcai.2020/439. Main track. Valeriani et al. (2023) Lucrezia Valeriani, Diego Doimo, Francesca Cuturello, Alessandro Laio, Alessio Ansuini, and Alberto Cazzaniga. The geometry of hidden representations of large transformer models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 51234–51252. Curran Associates, Inc., 2023. doi: 10.52202/075280-2230. URL https://proceedings.neurips.c/paper_files/paper/2023/file/a0e66093d7168b40246af1cddc025daa-Paper-Conference.pdf. Wang et al. (2018) Zhichun Wang, Qingsong Lv, Xiaohan Lan, and Yu Zhang. Cross-lingual knowledge graph alignment via graph convolutional networks. In Ellen Riloff, David Chiang, Julia Hockenmaier, and Jun’ichi Tsujii, editors, Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 349–357, Brussels, Belgium, 2018. Association for Computational Linguistics. doi: 10.18653/v1/D18-1032. URL https://aclanthology.org/D18-1032/. Wu et al. (2019) Yuting Wu, Xiao Liu, Yansong Feng, Zheng Wang, Rui Yan, and Dongyan Zhao. Relation-aware entity alignment for heterogeneous knowledge graphs. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 5278–5284. International Joint Conferences on Artificial Intelligence Organization, 7 2019. doi: 10.24963/ijcai.2019/733. URL https://doi.org/10.24963/ijcai.2019/733. Zhang et al. (2019) Qingheng Zhang, Zequn Sun, Wei Hu, Muhao Chen, Lingbing Guo, and Yuzhong Qu. Multi-view knowledge graph embedding for entity alignment. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 5429–5435. International Joint Conferences on Artificial Intelligence Organization, 7 2019. doi: 10.24963/ijcai.2019/754. URL https://doi.org/10.24963/ijcai.2019/754.