{"id":"d0c7f14b-0726-4f25-a8c1-a7696dbc514d","arxiv_id":"2506.10014","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"NOCL lets an LLM handle node, edge, and graph tasks on text and non-text graphs by compressing each node's description into one semantic embedding and turning the graph into a text prompt.","lead":"NOCL is a new way to make large language models solve graph tasks without graph neural networks. It compresses each node's description into a single embedding so one LLM can handle node, link, and graph predictions in a text-based format.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Possible target-edge leakage in link prediction: the graph descriptor in §3.4 lists every edge of the induced subgraph, so the queried edge tuple may appear verbatim in the prompt; if unmasked, the Cora/PubMed link results in Table 3 do not measure link prediction.","rationale":"The central claim is that NOCL uniformly solves node, edge, and graph tasks without message passing and generalizes in zero-shot settings. Link prediction is one of the three task levels and produces the largest apparent margin over MPNN baselines (Cora 0.8965 vs. roughly 0.68). If the target edge tuple is present in the graph descriptor, the link prediction experiment does not require structural reasoning: the LLM can answer by locating the tuple in the edge list. That would invalidate a major part of the empirical support for the paper's unified-task claim, even though node classification and graph classification results could still stand. I did not choose the reader's 1-hop truncation as the primary concern because, while real, it is explicitly acknowledged in Section 6 and would not by itself explain a link prediction result that is suspiciously high. I also recognize the paper's genuine strengths: the node-concept compression is well motivated, the efficiency numbers in Table 6 are useful, the limitations section is honest, and the code is promised in the supplementary material. The proposed test is deliberately cheap: inspect the descriptor construction and, if needed, mask the target edge. If the leakage is confirmed, the paper needs a major revision rather than outright acceptance; if masking is already present and the results survive, the concern is resolved. Since the issue is empirical and directly checkable, CONDITIONAL is the appropriate disposition, with the explicit condition that link prediction be re-evaluated with the target edge masked.","tokens_in":20516,"tokens_out":10618,"duration_ms":116930,"concrete_test":"Using the released code referenced in Appendix E, trace the data-generation routine for link-level tasks and print the graph descriptor for a held-out positive test pair and a matched negative pair. Check whether the tuple (vi,vj) appears in the edge list. Then rerun Cora link prediction after removing the target edge tuple from every descriptor, or otherwise masking it, while keeping all hyperparameters unchanged, and compare ROC_AUC to Table 3; if AUC drops substantially toward the MPNN baselines or chance, the reported 0.8965 is an artifact of edge leakage.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing unresolved issue is whether link prediction is actually being tested. Section 3.4 defines graph representation descriptors as a complete listing of nodes followed by all edges, with tuples <|EDGE|> i j for every edge in the induced subgraph. For edge-level tasks, the same section says the target node vi is placed first, then all nodes in its induced graph Gvi are generated, followed by vj and Gvj, with edges from Gvi preceding edges from Gvj. For a positive test pair (vi,vj), the edge (vi,vj) belongs to Gvi by definition, so the descriptor contains the tuple <|EDGE|> vi vj; for a negative pair it does not. The paper never states that this target edge is masked. If it is not masked, the LLM can answer \"connect / not connect\" by checking whether the tuple is present, reducing link prediction to string membership. Section 4.1 only describes the 1-hop induced subgraph cap at 11 nodes and the linear projection head for ROC_AUC; no masking is mentioned. Under this reading, the Cora supervised link result (0.8965) and the PubMed zero-shot result (0.6351) in Table 3 are not evidence of structural link prediction, and the unified-task claim is weakened for edge-level tasks. The reader's 1-hop truncation concern is real but secondary; even a correct 1-hop protocol would not fix leakage if the target edge is listed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NOCL, an MPNN-free framework that converts node features into natural-language descriptions, encodes them into compact 'node concept' embeddings via a pretrained language model and a linear connector, and feeds graph representation descriptors (node and edge listings) into a LoRA-tuned LLM. The authors claim that this unified, text-based formulation handles node, link, and graph classification on both textual-attributed and non-textual graphs, with large token-length reductions and strong zero-shot generalization, and they report experiments on ogbn-arxiv, Cora, PubMed, MUTAG, and ogbg-molhiv.","tokens_in":20827,"tokens_out":7110,"duration_ms":66857,"significance":"If the reported results hold, NOCL would be a useful step toward graph foundation models that do not rely on message passing, with concrete efficiency gains (up to 93.9% token reduction, substantial memory and time savings) and a unified task format. The paper also provides a reproducibility link and discusses limitations honestly in Section 6. However, several load-bearing aspects of the evaluation—particularly the link prediction protocol and the zero-shot setting—require scrutiny before the claims can be accepted.","major_comments":[{"comment":"The link prediction evaluation does not mask the target edge in the graph representation descriptor. Section 3.4 states that the descriptor lists all edges, and for edge-level tasks edges from the first target node's induced graph Gvi precede edges from the second target node's induced graph Gvj. For a positive test pair (vi, vj), the edge (vi, vj) is by definition in Gvi, so the descriptor contains the tuple <|EDGE|> vi vj; for a negative pair it does not. Because the prompt asks 'Should node vi connect node vj?' with the full descriptor visible, the LLM can answer by checking string membership rather than by structural reasoning. The paper never states that this edge is masked (Section 4.1 only describes the 1-hop cap and the linear projection head). Consequently, the supervised Cora link results (0.8842/0.8965) and zero-shot PubMed link results (0.6057/0.6351) do not currently support the claimed link prediction capability. The authors should specify and implement a masking protocol (e.g., removing the queried edge from the descriptor) and re-report all link-level numbers.","section":"3.4, Table 3"},{"comment":"The 'zero-shot' results are partially self-referential because the node concept connector is tuned with unlabeled node descriptions from the same target datasets used for zero-shot evaluation. Section 4.1 states 'We utilize all nodes from these datasets to generate the corresponding node descriptions and node concepts, which are then employed in training the connector,' and Appendix B shows that title/abstract recovery data are generated for all nodes, with only the class-prediction task restricted to training nodes. Thus, the PubMed and Cora zero-shot numbers are obtained with a model whose connector has already seen the raw text of those graphs. This weakens the claim of generalization to unseen domains made in Section 1. The authors should either exclude target-dataset text from connector tuning entirely and re-run the zero-shot experiments, or explicitly relabel the setting as 'transductive zero-label' and adjust the claims accordingly.","section":"3.3, 4.1, Appendix B"},{"comment":"The ROC_AUC numbers for link and graph tasks are not produced by text generation but by a learned linear projection applied to the LLM's last hidden state. Section 4.1 states: 'To adapt LLM outputs for the ROC_AUC evaluation metric, we flatten the last hidden state corresponding to the first output token of the LLM and apply a linear projection to produce the final numeric prediction.' This contradicts the paper's claim in Section 3.4 that NOCL does not rely on specialized heads or task-specific formats. It is also unclear how this projection head is trained relative to the LoRA adapter, and how it transfers in the zero-shot PubMed link prediction. Please report the exact training protocol for the projection head, or use a text-derived scoring rule (e.g., answer-token probability) for all tasks.","section":"4.1, Tables 3-4"},{"comment":"No uncertainty quantification is reported, and MUTAG has only 188 graphs with an 80/20 split, giving roughly 38 test graphs. The claimed improvement over the best MPNN on MUTAG (0.7262 vs 0.6310 for the 1B model) could be within sampling noise. The authors acknowledge the absence of error bars in the checklist but still make strong comparative statements (e.g., 'exceeds the best-performing MPNN by a margin of 0.09'). At minimum, report standard deviations over multiple random splits or bootstrap confidence intervals for MUTAG and for the zero-shot results, or temper the claims accordingly.","section":"4.1, Table 3, Checklist item 7"}],"minor_comments":[{"comment":"The formula 'The overall token length of graph representation descriptors for G is 4 + 2n + 3n' should read '4 + 2n + 3m', since the second term counts edge tuples and there are m edges, not n.","section":"3.4"},{"comment":"There are typos such as 'The LLaV A seires' and 'pretraind'; these should be corrected.","section":"5"},{"comment":"In Table 8, the 'Ave. Node' and 'Ave Edge' columns are empty for the citation networks and molecular graphs; clarify whether these are per-graph averages and fill in the values for consistency.","section":"Table 8"},{"comment":"The authors answer 'No' to the question about licenses but then provide the names of the Llama 3.2 and Apache 2.0 licenses; please clarify whether full license texts and terms are included with the released code and data.","section":"Checklist item 12"},{"comment":"The declaration of LLM usage is inconsistent: the paper answers 'No' with the justification that the LLM is used only for writing, whereas the core methodology is built around LoRA fine-tuning of Llama-3.2. This should be corrected to accurately reflect the use of LLMs in the research method.","section":"Checklist item 16"},{"comment":"The base model zero-shot results for PubMed show 0.0000 and 0.0002 accuracy, which appear degenerate; a brief explanation of these outputs would help the reader interpret the comparison.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the unmasked target edge in the link prediction descriptor, which calls the Table 3 link results into question. The zero-shot claim also needs re-scoping because the connector is tuned on unlabeled text from the target datasets. The paper's efficiency findings and released code are strengths, and I believe the approach is promising if these evaluation protocols are fixed and the experiments re-run."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is good: compress each node description into one PLM embedding token, then feed a standardized graph descriptor into a LoRA-tuned LLM. That combination is new as far as I can tell, and the efficiency gains are real—token reductions of 80–94% and training on a single RTX 4090. The supervised ogbn-arxiv and ogbg-molhiv numbers are respectable, and the paper ships code, which is more than many papers in this area do.\n\nThe problem is link prediction. In §3.4, the graph descriptor lists every edge of the induced subgraph. For a positive test pair (vi,vj), the edge (vi,vj) is in vi's 1-hop subgraph by construction, so the tuple appears verbatim in the prompt. The paper never says this target edge is masked. If it isn't, the model can answer by string membership—\"does the edge tuple appear?\"—which is not link prediction. That would invalidate the Cora supervised and PubMed zero-shot link results in Table 3. This is a load-bearing flaw, not a minor gap.\n\nThere are secondary issues. The \"zero-shot\" claims are diluted: the connector is trained on unlabeled text from the same PubMed and Cora corpora used for zero-shot evaluation, so those corpora are not fully unseen. No error bars are reported, which matters a lot for MUTAG (188 graphs, 0.09 margin). And the ROC_AUC metric comes from a linear projection head on the LLM's hidden state, which undercuts the \"no specialized heads\" framing. The 1-hop truncation at 11 nodes is a real but secondary limitation, and Section 6 acknowledges it.\n\nThe framework deserves serious attention, and the node-concept compression is likely reusable beyond this paper. But as submitted, the link prediction claims don't hold up, and the zero-shot framing needs to be more careful. This is a fixable paper rather than a dead one.\n\nRecommendation: send to review, but require the authors to address masking, report error bars, and re-run the zero-shot evaluation without connector training on target corpora. A good referee can get this into shape.","headline":"NOCL's node-concept compression is genuinely useful, but its link prediction results are likely invalid because the prompt may literally contain the edge being predicted.","tokens_in":21374,"tokens_out":2454,"would_cite":false,"duration_ms":25896,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a large language model can handle node, edge, and graph classification on text-attributed and molecular graphs without any message-passing module, by translating node features into compact semantic embeddings and…","keywords":["graph foundation models","large language models","node concept embedding","zero-shot learning","message passing neural networks","graph representation descriptors","LoRA instruction tuning","text-attributed graphs"],"falsifier":"Construct a synthetic benchmark where the class label is determined entirely by a node's second- or third-hop structure and the 1-hop features are class-balanced noise; if NOCL's accuracy drops to chance while an MPNN with message passing stays high, the sufficiency of the 11-node induced subgraph is refuted.","tokens_in":20291,"feed_emoji":"🧠","tokens_out":5813,"duration_ms":53311,"temperature":0.7,"pith_summary":"The paper introduces NOCL, a framework that lets an instruction-tuned LLM solve graph tasks directly from text prompts instead of propagating messages through a GNN. The central idea is to turn each node's raw features into a natural-language description, compress that description into a short learned embedding with a pretrained language model, and then feed the LLM a graph descriptor that lists nodes and edges. This recasts node, link, and graph classification as ordinary next-token prediction, so one model can serve all three task levels without specialized output heads. If the approach holds, graph learning no longer needs message passing, and a single LLM can transfer zero-shot to unseen graph datasets.","feed_headline":"No message passing needed: one LLM handles node, link, and graph tasks","feed_subtitle":"Node-concept embeddings shrink input tokens by up to 94%, and zero-shot node classification beats LLM-MPNN hybrids on Cora.","key_machinery":"The load-bearing objects are the node concept embedding and the graph representation descriptor. A node concept is a compact semantic embedding produced by feeding a node's natural-language description through a pretrained language model; a graph descriptor is a token sequence that lists nodes' concept embeddings and then their edges, delimited by special tokens like <|BON|>, <|NC|>, and <|BOE|>. Together they let the LLM see both attributes and adjacency in one prompt, and they convert every downstream task into a question answered by next-token generation.","core_discovery":"NOCL claims that graph structure can be traded for language: a frozen LLM, lightly adapted with LoRA, can classify nodes, predict links, and classify whole graphs by reading graph representation descriptors built from node concept embeddings. The node concept is a fixed-size vector made by encoding a node's description with a sentence transformer, which cuts input token counts by up to 93.9% while keeping the semantics. With node-level and link-level tasks re-expressed as graph-level questions over induced subgraphs, all tasks become text comprehension, and experiments on Cora, PubMed, ogbn-arxiv, MUTAG, and ogbg-molhiv report supervised accuracy competitive with MPNNs and better zero-shot transfer than existing LLM baselines.","pith_inferences":["Editorial inference: if the 1-hop-truncation assumption is robust, NOCL-style models could scale to much larger graphs by subsampling neighborhoods, a setting the paper explicitly leaves untested.","Editorial inference: the paper's framing suggests that structural regularities used by MPNNs, such as oversmoothing and locality bias, can be bypassed entirely; a testable corollary is that NOCL should maintain accuracy on graphs with long-range dependencies that hurt shallow MPNNs.","Editorial inference: replacing hand-written non-TAG templates with an LLM-generated or learned description step would likely improve generalization to new molecule types, since the current templates require expert curation.","Editorial inference: one could probe how much structure the LLM actually uses by ablating edge lines from the graph descriptor; if accuracy stays high, the model may be relying mostly on node semantics rather than topology."],"forward_implications":["A single LoRA-tuned LLM can solve node, link, and graph classification together, with no message-passing module and no task-specific heads.","Token compression from node concepts (up to 93.9%) brings training memory low enough for commercial GPUs, enabling datasets that previously caused out-of-memory failures.","Zero-shot transfer improves over LLM baselines on Cora and matches supervised MPNN link prediction on PubMed, suggesting labels are not strictly required for competitive graph reasoning.","Because outputs come from next-token prediction, the same framework extends beyond fixed labels to free-form graph explanations and questions.","The approach extends LLM use from text-attributed graphs to molecular graphs, where node features are converted into natural-language descriptions."],"supporting_citations":[{"why":"Supplies the LoRA low-rank adaptation used to fine-tune the LLM during downstream instruction tuning.","marker":"[17]"},{"why":"Provides the sentence transformer (all-mpnet-base-v2) that encodes node descriptions into node concept embeddings.","marker":"[34]"},{"why":"Establishes the LLaVA-style modality connector structure the paper follows to bridge the PLM embedding space to the LLM.","marker":"[26]"},{"why":"Contributes the prompt-based reformulation that turns node- and edge-level tasks into graph-level questions.","marker":"[38]"},{"why":"Supplies the ogbn-arxiv and ogbg-molhiv datasets used for supervised node and graph classification.","marker":"[18]"},{"why":"Provides the link prediction evaluation protocol (train/validation/test split with negative sampling) used on Cora and PubMed.","marker":"[22]"},{"why":"Introduces the BERT-style pretrained language model family that underpins the PLM encoder for node descriptions.","marker":"[9]"},{"why":"Serves as the hybrid LLM-MPNN baseline whose results are compared in the node classification tables.","marker":"[41]"}],"fun_headline_variants":["LLM reads graphs as text without message passing","Node concepts shrink tokens 94%, enabling zero-shot graph tasks","No MPNN, just LLM: NOCL unifies all graph tasks","Graph tasks become text comprehension via node concepts","One frozen LLM for node, link, and graph tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For node and link tasks, NOCL assumes that a node's 1-hop neighborhood capped at ten neighbors contains enough structural signal for correct classification; if multi-hop context is essential, the reported performance would not hold.","fun_headline_variants_meta":{"raw":{"variants":["LLM reads graphs as text without message passing","Node concepts shrink tokens 94%, enabling zero-shot graph tasks","No MPNN, just LLM: NOCL unifies all graph tasks","Graph tasks become text comprehension via node concepts","One frozen LLM for node, link, and graph tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1499,"prompt_tokens":974,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":442}},"tokens_in":590,"tokens_out":525,"duration_ms":4960,"temperature":1.0,"reasoning_tokens":442,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:18:53.373870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a synthetic benchmark where the class label is determined entirely by a node's second- or third-hop structure and the 1-hop features are class-balanced noise; if NOCL's accuracy drops to chance while an MPNN with message passing stays high, the sufficiency of the 11-node induced subgraph is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the LLaVA-style modality connector structure the paper follows to bridge the PLM embedding space to the LLM."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the hybrid LLM-MPNN baseline whose results are compared in the node classification tables."}],"review_version":1}