{"id":"aea94d3a-c269-40ac-a706-a1fa5471ac4e","arxiv_id":"2505.08265","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using interchange interventions on a purpose-built synthetic graph dataset, the paper finds that frozen-LLM features carry node-level information into shallow GNN layers and that an attention-based token and prompt selector adds about 0.7 to 3.1 accuracy points.","lead":"This paper studies pipelines where a large language model produces node features that a graph neural network classifies, using causal interventions to locate where information enters. It introduces a controllable synthetic graph dataset, reports where LLM-derived information lands in the GNN, and adds an attention module that gives small consistent accuracy gains.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.5's proof of GNN-variable sufficiency rests on Appendix B.4 Eq. (31), an unproven bottleneck assertion; Empirical Finding 1's localization may be an artifact of searching only GNN layers.","rationale":"The reader's weakest assumption is that an exact bijective correspondence between GNN variables and the hand-built causal model variables exists. My concern sharpens that: the paper's own attempt to establish the needed correspondence, Theorem 3.5, is circular at Eq. (31) in Appendix B.4. That equation assumes the GNN first layer is a sufficient bottleneck for every LLM hidden variable, which is exactly the sufficiency claim the theorem must prove. This is not merely a missing proof: in a transformer with residuals, attention, and nonlinearities, an intermediate LLM state is not generally recoverable from, or interventionally equivalent to, a downstream GNN layer. The empirical findings and the AT module are still valuable. The AT results in Table 2 show consistent positive differences with standard deviations and multiple seeds, and the CCSG dataset is a real artifact. But Empirical Finding 1 is the paper's main analytical output, so the unverified Eq. (31) is load-bearing. A conditional verdict is appropriate: require the authors to test Eq. (31) directly and to search LLM-layer candidates before claiming causal localization. If the test fails, the findings should be reframed as properties of the search procedure rather than as mechanism identification.","tokens_in":38260,"tokens_out":8747,"duration_ms":93514,"concrete_test":"Directly test Eq. (31) on the published CCSG node-level task: instrument Llama3+GCN and, for a held-out set of (Gorig, Gdiff) pairs, compare INTINV(f,Gorig,Gdiff,eZ^i_j) with INTINV(f,Gorig,Gdiff,Zhat^1_j) for an intermediate LLM hidden variable eZ^i_j using the same metric D as Eq. (1). If the two quantities differ for any pair, Eq. (31) is false and the Theorem 3.5 sufficiency argument collapses. In the same run, include LLM hidden layers as candidates in the L_II minimization; if any LLM layer attains L_II below every GNN layer, Empirical Finding 1's localization to early GNN layers is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing step is Appendix B.4's Eq. (31). It asserts that because the GNN first-layer variable Zhat^1_j blocks all causal routes from X_j to Y_f, INTINV(f,Gorig,Gdiff,S)=INTINV(h,Gorig,Gdiff,Zhat^1_j) for every S in {X_j} union {eZ^i_j}, i.e., every LLM hidden variable is interventionally exchangeable with the GNN's first-layer representation. Causal blocking gives conditional independence, not interchange equivalence: Zhat^1_j is a non-injective, nonlinear function of the LLM layers, and patching an intermediate eZ^i_j (e.g., an attention or MLP state) can produce downstream activations that no valid input induces. The proof then uses Eq. (31) to replace LLM variables by GNN variables, which is precisely the sufficiency claim Theorem 3.5 is supposed to establish. Lemma B.2's IC-algorithm argument only recovers the architecture's dependency graph; it supplies no functional equations and no counterfactual equivalence. Since the L_II search in Figures 3-6 is restricted to GNN layers, the reported minima may reflect which layer is most sensitive to the patch, not where the LLM's information is consumed. This is the central mechanism claim (Empirical Finding 1), so the concern is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the LLM-enhancer-plus-GNN paradigm by constructing a synthetic graph dataset (CCSG) with controllable causal relationships, then applying interchange interventions to identify which internal variables of the combined model correspond to variables of a hand-specified high-level causal model. It presents a theoretical validation of this alignment procedure (Theorem 3.2, Corollary 3.3, Proposition 3.4, Theorem 3.5), draws three empirical findings about where LLM-provided information enters the GNN, and proposes an Attention-based Transmission (AT) module that selects LLM token features and prompts. Table 2 reports consistent accuracy gains from the AT module across Cora, Pubmed, and Instagram with multiple LLM and GNN backbones.","tokens_in":38606,"tokens_out":10032,"duration_ms":96695,"significance":"If the causal-analysis claims were valid, the paper would provide a useful mechanistic account of LLM-enhanced GNNs and a practical plug-in module. The empirical infrastructure is a genuine strength: the authors build a controlled synthetic dataset, use standard backbones, report means and standard deviations, and release code. The AT module gains are consistent across three datasets and three LLM/GNN combinations, with improvements between roughly +0.7 and +3.1 points. However, the central theoretical validation is not established: the proofs of Theorems 3.2 and 3.5 contain circular or unsupported steps, and Empirical Finding 1 is drawn from a search restricted to GNN layers. Consequently, the analytical findings should be regarded as hypotheses about layer sensitivity rather than validated causal mechanisms.","major_comments":[{"comment":"The proof of Theorem 3.2 is circular and relies on an unstated value-preservation assumption. Equation (16) replaces f_{Z_f}(G_orig) with eta^{-1}(h_{Z_h}(G_orig)), although the theorem only assumes a bijection between variable sets and does not state that eta maps the value taken by Z_f on an input to the value taken by Z_h on the same input. More importantly, Eq. (17) concludes INTINV(f,Gorig,Gdiff,Z_f) = INTINV(h,Gorig,Gdiff,Z_h) by treating f_latter on a patched input as equal to h_latter on the corresponding patched input; that is exactly the interchange-invariance property the theorem is supposed to derive, not a consequence of the stated premises. Thus the claimed theoretical validation of the L_II minimization does not follow.","section":"Section 3.2.3 / Appendix B.1, Eqs. (16)-(17)"},{"comment":"The claim that because Zhat^1_j blocks all causal routes from X_j to Y_f, one has INTINV(f,Gorig,Gdiff,S) = INTINV(f,Gorig,Gdiff,Zhat^1_j) for every S in {X_j} union {eZ^i_j} is false in general. Causal blocking gives conditional independence in the observational distribution, not equality of counterfactual interchange outcomes. Patching an intermediate LLM state eZ^i_j changes the subsequent LLM computation and therefore the value of Zhat^1_j in a way that need not coincide with the effect of patching Zhat^1_j directly, because Zhat^1_j is a non-injective, nonlinear function of the LLM layers. Since Eq. (31) is the only step that eliminates LLM variables from the intervention, Theorem 3.5 is unproven. The IC-algorithm argument in Lemma B.2 recovers only the architecture's dependency graph; it supplies no functional equations and no counterfactual equivalence.","section":"Appendix B.4, Eq. (31)"},{"comment":"Empirical Finding 1 is not supported by the reported experiments because the L_II search is restricted to GNN layers, as stated in Section 3.3. The observation that the smallest L_II among GNN layers occurs in layers 0-1 shows only that early GNN layers are more sensitive to patch-induced changes than later layers; it does not localize where the LLM's information is consumed, nor does it establish that the LLM output features 'serve the function of representing information at the node level and the raw data level.' The text also says this finding 'partially validates Theorem 3.5,' but since Theorem 3.5 is not established, that statement is unsupported. A control search over LLM intermediate variables, or an intervention that patches the entire LLM output, would be needed before this claim can be made.","section":"Section 3.3.2 / Figure 3"},{"comment":"The justification of Proposition 3.4 contains an incoherent inequality: the left-hand side TE_{zf,zf'}(Y_f) - TE_{zf,zf'}(Y_f) is identically zero, and the surrounding sentence ('we have total effect of Z_{f,b} upon the output Y is strictly similar to Z_h than Z_{f,a}') is not a well-formed claim. More substantively, the proposition asserts that reaching the minimal possible value L*_II implies exact interchange equality and hence equal total effects; this assumes the existence of a variable that attains the lower bound, which is precisely what the earlier theorems were meant to establish. As a result, Proposition 3.4 does not provide an independent justification of the method.","section":"Appendix B.3, Eq. (29) and Proposition 3.4"}],"minor_comments":[{"comment":"Step 4 says 'utilize Equation 2 to calculate the loss L_II,' but the loss is defined in Equation 1; Equation 2 is the high-level model formula for the running example.","section":"Section 3.2.2, Step 4"},{"comment":"The table lists 'Bipartite Graph' twice; the second row's description ('Outer nodes connected in a ring with a central node') appears to describe a different topology and should be re-labeled or corrected.","section":"Table 5"},{"comment":"The experimental details state that q was set to 10 for the AT module, while Table 6 reports results only for q = 1, 2, 3, 4; please clarify whether q = 10 was used in Table 2 and why the ablation stops at q = 4.","section":"Appendix D.1 / Table 6"}],"recommendation":"major_revision","confidential_remarks":"The theory section is the main obstacle. Theorem 3.2's proof is circular as written, and Theorem 3.5's Eq. (31) is not a valid inference from causal blocking. These are load-bearing because Empirical Findings 1-3 are presented as validated by the theory. The AT module results and the CCSG dataset are independent empirical contributions and could be publishable, but the authors would need to either provide a valid proof (unlikely given the general falsehood of Eq. (31)), or substantially reframe the paper around the empirical sensitivity analysis and the module, explicitly dropping the causal-identification claims. If the authors choose the latter, the title and abstract should be revised accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The empirical core is real and reproducible-looking: a new controlled synthetic graph dataset (CCSG), a causal-intervention study suggesting that frozen LLM features align best with early GNN layers and carry node-level/raw-level information, and an attention-based transmission module with small but consistent accuracy gains across three GNNs and three LLMs. The theoretical wrapper is not sound as written; the proofs are circular or gappy, and the paper leans on them too hard in the abstract.\n\nThe dataset is the strongest contribution. It is carefully built, with controllable semantics, varied topology, and a sensible high-level causal model. The L_II analysis is a reasonable application of interchange interventions, and the finding that LLM token features act as node-level information entering shallow GNN layers is interesting, even if it is roughly what the architecture suggests. The AT module is simple, plug-and-play, and honestly reported with standard deviations; the gains are modest but consistent, and the module is cheap enough to be useful.\n\nThe soft spots are real. Theorem 3.2's proof reaches Eq. (17) only by assuming the interchange-invariance it is supposed to establish, and the bijection premise already gives away the conclusion. More load-bearing, Theorem 3.5's claim that GNN variables are sufficient rests on Appendix B.4 Eq. (31), which is asserted: Zhat^1_j blocking causal routes gives conditional independence, not interchange equivalence with every LLM hidden variable. Causal blocking does not make a non-injective nonlinear layer interventionally substitutable for an attention or MLP state. And because the L_II search in Figures 3-6 is restricted to GNN layers, the localization results may partly reflect which layer is most sensitive to the patch, not where the LLM's information is consumed. That is the paper's central mechanism claim, so the concern is load-bearing. The AT protocol is also underspecified: the main text says q=10 in Appendix D, Table 6 reports q=1..4 and omits q=10, and there are no external baselines beyond the w/o AT ablation. These are fixable, but they matter.\n\nThe causal-abstraction machinery is imported from Geiger and colleagues, and Appendix A.1 itself notes Wu et al. 2023b already applied causal effect estimation to GNNs combined with LLMs, so the interpretive framework is incremental. That is fine if the empirical contribution carries it.\n\nBottom line: the dataset and the empirical findings deserve a serious referee. The theory should either be rigorously repaired or explicitly reframed as informal motivation. I would send it to review rather than desk reject, with a clear instruction to the authors to fix the proofs or cut them. I would also ask them to clarify the q setting and add a non-AT baseline that uses the same prompt diversity without the attention module.","headline":"A genuinely useful synthetic dataset and an interesting empirical localization result wrapped in a theoretical section whose proofs do not currently hold up; worth refereeing with a strong request to repair or reframe the theory.","tokens_in":39136,"tokens_out":1952,"would_cite":true,"duration_ms":21098,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"By swapping internal activations between graph pairs, the paper locates where a frozen LLM's features enter a GNN — at the node and raw-data level, consumed in the shallow layers — and turns this insight into an attention module that…","keywords":["LLM enhancers","graph neural networks","interchange intervention","causal mechanism identification","synthetic graph dataset","attention-based transmission","node classification","feature selection"],"falsifier":"Train the same GNN pipeline on CCSG with a frozen non-LLM feature extractor of similar size; if the same early-layer alignment pattern and $L_{II}$-accuracy correlation appear, the findings describe generic shallow-feature processing rather than LLM semantic knowledge. Alternatively, retrain the identical architecture with different seeds and rerun the interchange search: if the layer minimizing $L_{II}$ moves across seeds or sample pairs, the claim about where information lives in the model fails.","tokens_in":38049,"feed_emoji":"🧠","tokens_out":8903,"duration_ms":76187,"temperature":0.7,"pith_summary":"The paper tries to establish where a frozen large language model's contribution actually enters the common 'LLM-enhancer-plus-GNN' pipeline, in which an LLM writes node features and a graph neural network classifies the graph. Using a new synthetic dataset with built-in causal relationships (CCSG) and the interchange-intervention method from causal inference, it finds that the LLM's features carry node-level and raw-data-level semantic information, and are consumed in the shallow layers of the GNN. It further claims this early-layer placement is invariant to GNN depth and width, and that the alignment loss $L_{II}$ tracks model accuracy. Based on those findings it designs an attention-based transmission (AT) module that selects which LLM token features to pass on, and reports consistent accuracy gains between roughly +0.7 and +3.1 points on Cora, Pubmed, and Instagram across several LLM and GNN backbones.","feed_headline":"Swapping internal activations reveals where LLM features enter a GNN","feed_subtitle":"LLM features feed node-level semantics into early GNN layers; attention gating turns this into accuracy gains.","key_machinery":"The central tool is the interchange intervention loss $L_{II}$: a value of an internal variable $Z_f$ of the trained network, computed on a second graph sample, is swapped into the network processing a first sample, and the resulting output is compared, via cross-entropy, to the same swap performed in a hand-built high-level causal model $h(\\cdot)$. Searching over candidate hidden layers for the one that minimizes $L_{II}$ locates where the causal variables' information is encoded, and Theorems 3.2 to 3.5 convert that search into a statement about equal total effects. The CCSG dataset supplies the ground-truth causal variables ($\\Psi$, $\\Phi$, $\\Gamma$, $\\Omega$) that appear in those models, and the AT module — a small transformer that scores LLM output tokens across multiple prompts and reweights them into the node feature vector — operationalizes the finding.","core_discovery":"The central claim is that interchange interventions on the CCSG dataset reveal a precise division of labor in the LLM-enhancer-plus-GNN stack: with a fixed-parameter LLM, the features the LLM produces represent node-level and raw-data-level information, and the GNN internal variable that best matches the high-level causal model sits in layers 0 and 1, the early message-passing layers. The paper further claims that the GNN's internal structure after ingesting LLM input is largely invariant to model scale, that scaling up depth or width does not improve causal relation modeling, and that the minimum attainable interchange loss $L_{II}$ correlates with model accuracy. The design consequence is that the interface between the LLM output and the GNN input is a higher-leverage place to optimize than the backbone choice or model size, which motivates attention-weighted selection of token-level features as the AT module.","pith_inferences":["If the early-layer locus generalizes beyond CCSG, the design rule for hybrid LLM/GNN systems would be to architect the feature interface rather than the GNN capacity; a testable extension is to fine-tune the LLM enhancer and check whether the semantic information moves deeper into the stack or stays at the LLM output.","The even-layer-versus-odd-layer alignment asymmetry across depths could be probed by retraining the identical architecture with different seeds: if the asymmetry survives, it reflects a learned alternating internal structure rather than a property of the search procedure.","The $L_{II}$-as-capability link, if confirmed on other architectures, offers a cheap training-free probe for benchmarking model families without full evaluation runs."],"forward_implications":["Fixed-parameter LLM enhancers should feed their knowledge into the earliest GNN layers; routing LLM features through many rounds of message passing adds little under the paper's analysis.","Which token positions of the LLM output are selected as node features is a high-leverage design choice, reported as more consequential than the LLM family or GNN scale.","The interchange loss $L_{II}$ can serve as an indicator of model capability, since lower minimum alignment loss generally coincides with higher accuracy in the reported experiments.","Scaling up GNN depth or width mostly enlarges the structure without improving its ability to model the underlying causal relations, as the paper states as a hypothesis from Empirical Finding 2.","Attention-based selection over multiple prompts yields consistent gains across three datasets and three GNN backbones while leaving the LLM frozen, making the module plug-and-play."],"supporting_citations":[{"why":"Supplies the interchange intervention method that the paper applies to align low-level network variables with a high-level causal model.","marker":"Geiger et al., 2022a"},{"why":"Provides the definitions of total effect and structural causal models on which Theorem 3.2 and Theorem 3.5 rest.","marker":"Pearl, 2009"},{"why":"Establishes the node-as-smallest-variable-unit convention and prior synthetic causal graph datasets that CCSG extends.","marker":"Wu et al., 2022"},{"why":"Introduces the CRCG causal graph dataset that CCSG extends with controllable multi-order semantic relationships.","marker":"Gao et al., 2024"},{"why":"Defines the LLM-enhancer-plus-GNN framework that is the object of the causal analysis.","marker":"Chen et al., 2023"},{"why":"Provides the fixed-parameter LLM enhancer setting that the paper adopts and analyzes.","marker":"Liu et al., 2024"},{"why":"GCN is the primary GNN backbone analyzed, and Cora is one of the evaluation datasets.","marker":"Kipf & Welling, 2016"},{"why":"Llama 3 is the LLM enhancer whose output features are localized by the intervention analysis.","marker":"Dubey et al., 2024"}],"fun_headline_variants":["Early GNN layers are where LLM features do their work","Causal swaps reveal LLM slots into GNN layers 0-1","LLM tokens need attention gating to help GNNs shine","Scale doesn't matter: LLM features target early layers","Intervention test pinpoints LLM entry point in GNNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analysis assumes that a subset of the trained GNN's internal variables can be put into exact one-to-one correspondence with the variables of the hand-built causal models $h_{node}$ and $h_{graph}$; if no such correspondence really exists in a trained network, the layer that minimizes $L_{II}$ is only a proxy and the empirical findings partly describe the search procedure rather than the model's mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Early GNN layers are where LLM features do their work","Causal swaps reveal LLM slots into GNN layers 0-1","LLM tokens need attention gating to help GNNs shine","Scale doesn't matter: LLM features target early layers","Intervention test pinpoints LLM entry point in GNNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000312,"raw_usage":{"total_tokens":1744,"prompt_tokens":883,"completion_tokens":861,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":771}},"tokens_in":499,"tokens_out":861,"duration_ms":8618,"temperature":1.0,"reasoning_tokens":771,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:01:41.849506+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same GNN pipeline on CCSG with a frozen non-LLM feature extractor of similar size; if the same early-layer alignment pattern and $L_{II}$-accuracy correlation appear, the findings describe generic shallow-feature processing rather than LLM semantic knowledge. Alternatively, retrain the identical architecture with different seeds and rerun the interchange search: if the layer minimizing $L_{II}$ moves across seeds or sample pairs, the claim about where information lives in the model fails.","supporting_citations":[{"cited_title":"Causality: Models, Reasoning, and Inference","cited_arxiv_id":null,"evidence_quote":"Provides the definitions of total effect and structural causal models on which Theorem 3.2 and Theorem 3.5 rest."},{"cited_title":"Rethinking causal relationships learning in graph neural networks","cited_arxiv_id":null,"evidence_quote":"Introduces the CRCG causal graph dataset that CCSG extends with controllable multi-order semantic relationships."},{"cited_title":"One for all: Towards training one graph model for all classification tasks","cited_arxiv_id":null,"evidence_quote":"Provides the fixed-parameter LLM enhancer setting that the paper adopts and analyzes."}],"review_version":1}