cs.LGOct 22, 2026

On the Token Collapse Phenomenon in Textual Graph Soft Prompting

Authors: Andrea Giuseppe Di Francesco, Rishabh Jain, Franco Maria Nardini, Nicola Tonellotto, Pietro Liò, Fabrizio Silvestri

Organizations: Sapienza University of Rome, Rome, Italy · ISTI-CNR, Pisa, Italy · University of Cambridge, Cambridge, UK · University of Pisa, Pisa, Italy

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Recent methods for injecting graph structure into LLM prompts include graph soft prompting, in which a graph neural network (GNN) learns graph-specific tokens that capture relational information among entities in textual prompts. G-Retriever combines graph soft prompts with a verbalized description of the same graph to strengthen the textual prompt signal. We conduct an in-depth analysis of G-Retriever and uncover a pathological behavior we term token collapse. Across three benchmarks, three graph encoders, two LLM families, and both frozen and LoRA-adapted models, G-Retriever's graph tokens become almost indistinguishable across inputs. We provide evidence that the same failure is inherited by subsequent methods, even with multi-token pooling. Under the standard setup, swapping a token with the dataset mean, a token from another example, or one derived from a corrupted graph produces no resolved answer-level effect in accuracy. Moreover, a learned constant prompt that never accesses the graph matches or surpasses G-Retriever on all three benchmarks. These findings indicate that, rather than conveying graph content as intended, the collapsed token functions primarily as a generic soft prompt. We also examine conditions linked to this collapse and propose a two-stage procedure that yields non-collapsed tokens while maintaining competitive performance in most cases. However, token variability alone does not prove that the model uses graph information: via interventions and decoding probes, we show that depending on the configuration, a token can reflect the question, an answer prior, or graph-derived content that the LLM may or may not use. Overall, our results offer a practical approach to evaluating graph-token alignment rather than inferring it from task accuracy alone. Code to reproduce our experiments is available at https://github.com/difra100/Token_Collapse_KGQA.

Figures & tables

Figure 1: Graph tokens of all 1,628 WebQSP test questions , projected onto two principal components after removing the mean token μ\bm{\mu} , with pairwise cosines shown below.
Figure 2: Two-stage pipeline. Stage 1 aligns the question-conditioned graph encoder. Stage 2 freezes it and trains separate graph-token and prompt-token pathways. Orange modules are trained; blue modules are frozen.
DatasetModelScoreEncoder cosineToken cosine
ExplaGraphsG-Retriever (GT)85.560.6780.997
G-Retriever (GAT)85.200.4570.998
G-Retriever (GCN)78.880.4230.985
SceneGraphsG-Retriever (GT)81.350.6300.993
G-Retriever (GAT)80.200.6930.998
G-Retriever (GCN)80.550.4280.998
Table 1: Collapse across datasets, graph encoders, LLMs, and tuning regimes. Scores are accuracy for ExplaGraphs and SceneGraphs and Hit@1 for WebQSP .
DatasetrealΔ\Delta meanΔ\Delta swapΔ\Delta gswapnotxtΔ\Delta notxt-meanconst
ExplaGraphs85.56−0.18-0.18−1.26-1.26−1.08-1.0866.43−0.90-0.9085.56
SceneGraphs81.35−0.35-0.35−1.20-1.20 †−1.10-1.10 †30.20+0.45+0.4581.40
WebQSP70.33−1.84-1.84 *−0.92-0.92−0.55-0.5541.34+1.04+1.0471.25
Table 2: Role of G-Retriever’s collapsed token. real , notxt , and const are absolute scores; other columns report differences from their corresponding reference. Full results in Appendices I , J , and K .
EncoderProjectorEncoder cos.Step-0 token cos.Token cos.Score
trained end to endSigmoid0.6951.0000.99785.56
trained end to endGELU0.7870.9920.91473.29 ***
frozen after Stage 1Sigmoid0.0240.9870.99284.48
frozen after Stage 1GELU0.0240.0910.32357.76 ***
Table 3: Factors associated with collapse on ExplaGraphs . Step 0 precedes projector training, and significance markers compare task score with the first row. Full projector diagnostics and intervention results are in Appendices H and I .
DatasetModelScoreToken cos.Δ\Delta meanΔ\Delta swapΔ\Delta gswap
Llama-2-7B
WebQSPG-Retriever70.330.996−1.84-1.84 *−0.92-0.92−0.55-0.55
Ours69.900.142−3.13-3.13 ***−2.15-2.15 *−1.97-1.97 †
no entity features72.73\mathbf{72.73}0.125−6.14-6.14 ***−5.65-5.65 ***−1.97-1.97 *
no question69.840.169−0.06-0.06−3.69-3.69 ***−2.33\mathbf{-2.33} *
SceneGraphsG-Retriever81.35\mathbf{81.35}0.993−0.35-0.35−1.20-1.20 †−1.10-1.10 †
Table 4: Token geometry, performance, and intervention effects. Ours denotes the complete two-token pipeline; Intervention columns report condition minus real . Bold scores mark the highest score within each dataset–LLM block; bold gswap effects satisfy ∣Δ∣>2|\Delta|>2 and p<.05p<.05 . Complete interventions and pp -values are in Appendix I , and pipeline ablations are in Appendix O .
DatasetModelQ rankA rankG rankΔ\Delta Score gswapΔ\Delta NLL S/GSDiagnosis
WebQSPG-Retriever293317369−0.55-0.55+0.0183/+0.0182+0.0183/+0.0182soft prompt
Ours6560315−1.97-1.97 †+0.0895/+0.0332\mathbf{+0.0895}/\mathbf{+0.0332}masked
no entity features3451303−1.97-1.97 *+0.1111/+0.0262\mathbf{+0.1111}/+0.0262prior
no question120134329−2.33\mathbf{-2.33} *+0.0848/+0.0845\mathbf{+0.0848}/\mathbf{+0.0845}used
SceneGraphsG-Retriever271436248−1.10-1.10 †+0.0169/+0.0167+0.0169/+0.0167soft prompt
Ours47187427−5.30\mathbf{-5.30} ***+0.0229/+0.1665+0.0229/\mathbf{+0.1665}used
Table 5: Token content and use with Llama-2. Q, A, and G are median retrieval ranks (lower is better); constant controls are 407 for WebQSP , 500 for SceneGraphs (A: 476–500), and 139 for ExplaGraphs (A omitted). Δ\Delta Score is the with-text gswap effect; NLL reports swap / gswap . Bold denotes effects meeting the criteria in Section 4.1 .
Appendix figures & tables72 assets

Supplementary material from the paper’s appendix.

Appendix

 ExplaGraphsSceneGraphsWebQSP
Graphs2,766100,0004,737
Average nodes5.1719.131,370.89
Average edges4.2568.444,252.37
Node attributescommonsense conceptsobjects and attributesFreebase entities
Edge attributescommonsense relationsactions, spatial relationsFreebase relations
Taskstance predictionscene-graph QAknowledge-graph QA
Appendix
Table 6: GraphQA, textual graphs benchmark. Numbers are from He et al. (2024) .
ControlPerturbationChurnAccuracyExact McNemar pp
Null replicate, same weights, same conditionnone at all63.2%−1.04-1.04.247
MLPQ, swap3×10−43\times 10^{-4} , cosine ≥0.999999\geq 0.99999938%−1.72-1.72.015
No entity features, nodezero, features already zeroed43%−0.61-0.61.415
CONST, swapzero, all token rows identical0%0.000.001
Appendix
Table 7: The measured floor , WebQSP , n=1,628n=1{,}628 . Churn is the fraction of generated strings that change; no perturbation here carries graph information.
  information from the original sample retained
ConditionWhat it replacesquestionentitiesstructurerelations
Interventions at fixed weights
meantoken →\to training-set mean μ\bm{\mu}×\times×\times×\times×\times
swaptoken →\to another question’s token×\times×\times×\times×\times
gswapgraph →\to another question’s graph✓\checkmark×\times×\times×\times
nodenode features →\to resampled features✓\checkmark×\times✓\checkmark✓\checkmark
Appendix
Table 8: Information from the original sample retained under each condition. ✓\checkmark indicates that the information remains available when constructing the token; ×\times indicates that it is removed or replaced. The final three rows are separately trained ablations rather than fixed-weight interventions.
MethodConfigurationProjectorKK regimeMean KK
GRAG-PCST ( Hu et al., 2025 )released GAT + LoRALinear–Sigmoid–Linear111.000
GRETPool ( Grover et al., 2026 )TopK ( Gao & Ji, 2019 ) , ratio 0.440.44Linear–Sigmoid–Linear⌈0.44N⌉\lceil 0.44N\rceil7.848
–SAG ( Lee et al., 2019 ) , ratio 0.440.44Linear–Sigmoid–Linear⌈0.44N⌉\lceil 0.44N\rceil7.848
–DiffPool ( Ying et al., 2018 )Linear–Sigmoid–Linear888.000
–DiffPool ( Ying et al., 2018 )Linear–Sigmoid–Linear100100100.000
GraSP-PCST ( Wang et al., 2026 )SubgraphGNN ( Alsentzer et al., 2020 ) + LlamaLinear–Tanh–LinearNN16.719
Appendix
Table 9: Cross-method configurations. KK is the number of graph-derived tokens and NN the retained PCST node count. The projector column reports the final graph-to-LLM mapping.
MethodMean KKHit@1Inter cos.RRNorm ratioWithin cos.Eff. rank
GRAG-PCST1.00074.5700.9991710.02822342.935–10.300
TopK7.84871.4371.0000000.00015528.2581.000000–
SAG7.84872.7271.0000000.00053125.8020.999999–
DiffPool ( K=8K=8 )8.00052.7030.9997120.00928425.4770.999856–
DiffPool ( K=100K=100 )100.00061.3640.9999940.00129127.1701.000000–
GraSP-PCST16.71981.4500.9977830.0394424.2710.999465–
Appendix
Table 10: Cross-method token-collapse audit on WebQSP . Panel (a) reports geometry after mean-pooling each example’s KK tokens; within-example cosine is computed before pooling. Panels (b) and (c) report Hit@1 differences with exact McNemar pp -values in parentheses. Panel (b) uses the text-present real condition as reference; Panel (c) uses each model’s notxt-real condition. Significance symbols follow Section 4.1 . Results are single-seed and pp -values are unadjusted; NLL and multi-seed uncertainty are unavailable.
DatasetModelResidual ratioNorm ratio
ExplaGraphsG-Retriever (GT)0.05021.9 ×\times
G-Retriever (GAT)0.04024.4 ×\times
G-Retriever (GCN)0.12124.3 ×\times
SceneGraphsG-Retriever (GT)0.08884.8 ×\times
G-Retriever (GAT)0.04687.6 ×\times
G-Retriever (GCN)0.04897.0 ×\times
Appendix
Table 11: Residual and norm ratios of the trained tokens; norm ratio is relative to the median word embedding.
DatasetModelREALMEANSWAPNODEEDGENOTXT-REAL
ExplaGraphsG-Retriever (GAT)85.20−0.36-0.360.000.00−0.18-0.18+0.18+0.1876.90
G-Retriever (GCN)78.88+0.90+0.90+0.36+0.36+1.08+1.08+0.72+0.7269.86
SceneGraphsG-Retriever (GAT)80.20+0.60+0.60+0.20+0.20−0.20-0.20+0.10+0.1028.65
G-Retriever (GCN)80.55−0.10-0.10−1.05-1.05−1.25-1.25 †−1.25-1.25 †29.80
Appendix
Table 12: Interventions on the GAT and GCN encoders, as differences from REAL in accuracy points.
Figure 3: WebQSP , all 1,628 test questions, six models. G-Retriever 0.9963 cosine and 0.031 mean radius; with LoRA 0.9970 and 0.032; with Qwen2.5-3B 0.9975 and 0.022; our projector-only ablation 0.5425 and 0.479; our two-stage pipeline 0.1424 and 0.714; ours with LoRA 0.1702 and 0.686. Fine-tuning the LLM and changing the LLM family leave the token as collapsed as it was, and the projector-only ablation sits between the two regimes, which is the visual counterpart of the trade-off in Table 5 .
Figure 4: ExplaGraphs , all 554 test questions, five models: the three graph encoders of G-Retriever, with cosine 0.9974, 0.9983 and 0.9852 and mean radius 0.017, 0.012 and 0.026; our projector-only ablation, 0.3229 and 0.503; and our two-stage pipeline under the cross-fitted recipe, 0.0986 and 0.895. Collapse is the same for all three encoders.
Figure 5: SceneGraphs , the first 2,000 test questions, four models: G-Retriever with a graph transformer (cosine 0.9934, mean radius 0.077), with GAT (0.9980, 0.040), with GCN (0.9979, 0.032), and our projector-only ablation (0.5194, 0.736).
Figure 6: ExplaGraphs : the graph-free controls and the two causes of collapse. CONST (cosine 1.0000, radius 0.000) is a single point by construction and is the reference for a fully collapsed token; MLPQ (1.0000, 0.000) shows that a token computed from the question collapses just as completely; G-Retriever with GELU (0.9139, 0.116) shows that removing the saturating non-linearity alone is not enough while the encoder trains; Stage 1 with a Sigmoid projector (0.9918, 0.027) shows that an encoder whose outputs are nearly orthogonal still produces collapsed tokens. The last two panels are the visual form of Table 3 .
Input normToken normToken cosineResidual ratio
118.40.99990.009
1018.40.99190.090
10021.40.73460.600
Appendix
Table 13: Untrained projector (Equation 1 , default initialisation) applied to 500 random inputs of fixed norm. Collapse precedes training and is a property of the parameterisation.
ConditionAccuracyDifferencepp
REAL85.56––
MEAN85.38−- 0.181
SWAP84.30−- 1.260.118
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})85.20−- 0.360.727
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})84.48−- 1.080.146
NODE (real rows)84.66−- 0.900.227
Appendix
Table 14: G-Retriever (graph transformer) , 554 test questions.
ConditionAccuracyDifferencepp
REAL86.28––
MEAN84.84−- 1.440.134
SWAP85.02−- 1.260.265
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})84.84−- 1.440.134
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})83.57−- 2.71 ***0.000275
NODE (real rows)86.10−- 0.181
Appendix
Table 15: G-Retriever, initialisation 2 , 554 test questions.
ConditionAccuracyDifferencepp
REAL87.91––
MEAN87.18−- 0.720.219
SWAP87.73−- 0.181
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})87.73−- 0.181
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})87.91++ 0.001
NODE (real rows)88.09++ 0.181
Appendix
Table 16: G-Retriever, initialisation 3 , 554 test questions.
ConditionAccuracyDifferencepp
REAL89.71––
MEAN89.71++ 0.001
SWAP90.07++ 0.360.625
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})89.71++ 0.001
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})90.07++ 0.360.5
NODE (real rows)89.71++ 0.001
Appendix
Table 17: G-Retriever, initialisation 4 , 554 test questions.
ConditionAccuracyDifferencepp
REAL85.20––
MEAN84.84−- 0.360.754
SWAP85.20++ 0.001
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})85.02−- 0.181
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})85.20++ 0.001
NODE (real rows)85.02−- 0.181
Appendix
Table 18: G-Retriever (GAT) , 554 test questions.
ConditionAccuracyDifferencepp
REAL78.88––
MEAN79.78++ 0.900.405
SWAP79.24++ 0.360.86
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})80.14++ 1.260.23
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})77.44−- 1.440.215
NODE (real rows)79.96++ 1.080.345
Appendix
Table 19: G-Retriever (GCN) , 554 test questions.
ConditionAccuracyDifferencepp
REAL85.56––
MEAN85.02−- 0.540.375
SWAP85.56++ 0.001
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})85.02−- 0.540.375
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})84.84−- 0.720.125
NODE (real rows)85.56++ 0.001
Appendix
Table 20: CONST , 554 test questions.
ConditionAccuracyDifferencepp
REAL85.56––
MEAN85.02−- 0.540.25
SWAP85.56++ 0.001
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})85.02−- 0.540.25
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})85.56++ 0.001
NODE (real rows)85.56++ 0.001
Appendix
Table 21: CONST at word norm , 554 test questions.
ConditionAccuracyDifferencepp
REAL84.84––
MEAN85.02++ 0.181
SWAP85.20++ 0.360.688
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})85.02++ 0.181
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})85.20++ 0.360.5
NODE (real rows)84.84++ 0.001
Appendix
Table 22: MLPQ , 554 test questions.
ConditionAccuracyDifferencepp
REAL73.29––
SWAP72.20−- 1.080.631
Appendix
Table 23: G-Retriever with GELU (no Stage 1) , 554 test questions.
ConditionAccuracyDifferencepp
REAL84.48––
MEAN84.48++ 0.001
SWAP84.66++ 0.181
Appendix
Table 24: Stage 1 + Sigmoid projector , 554 test questions.
ConditionAccuracyDifferencepp
REAL57.76––
MEAN58.84++ 1.080.735
SWAP51.08−- 6.68 *0.0167
NODE (real rows)56.50−- 1.260.685
NODE (Gaussian)48.56−- 9.21 ***0.000616
EDGE (rewired)51.62−- 6.14 *0.02
Appendix
Table 25: Stage 1 + GELU, projector only , 554 test questions.
ConditionAccuracyDifferencepp
REAL85.74––
MEAN86.10++ 0.360.832
SWAP85.20−- 0.540.749
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})86.10++ 0.360.832
NODE (real rows)85.20−- 0.540.648
NODE (Gaussian)86.10++ 0.360.824
Appendix
Table 26: Ours, joint (the method) , 554 test questions.
ConditionAccuracyDifferencepp
REAL76.35––
MEAN72.74−- 3.61 *0.0265
SWAP77.44++ 1.080.545
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})72.92−- 3.43 *0.037
NODE (real rows)77.08++ 0.720.636
NODE (Gaussian)75.45−- 0.900.511
Appendix
Table 27: Ours, joint, cross-fitted , 554 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL84.48––
MEAN87.55++ 3.07 **0.00232
SWAP84.48++ 0.001
NODE (real rows)84.66++ 0.181
EDGE (rewired)84.66++ 0.181
GSWAP (other graph)84.12−- 0.360.754
Appendix
Table 28: Ours, no entity features , 554 test questions.
ConditionAccuracyDifferencepp
REAL81.35––
MEAN81.00−- 0.350.614
SWAP80.15−- 1.20 †0.0566
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})80.30−- 1.050.119
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})82.05++ 0.700.338
NODE (real rows)81.55++ 0.200.699
Appendix
Table 29: G-Retriever (graph transformer) , 2000 test questions.
ConditionAccuracyDifferencepp
REAL80.20––
MEAN80.80++ 0.600.213
SWAP80.40++ 0.200.781
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})80.60++ 0.400.434
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})80.15−- 0.051
NODE (real rows)80.00−- 0.200.757
Appendix
Table 30: G-Retriever (GAT) , 2000 test questions.
ConditionAccuracyDifferencepp
REAL80.55––
MEAN80.45−- 0.100.921
SWAP79.50−- 1.050.137
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})80.35−- 0.200.752
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})79.00−- 1.55 *0.0178
NODE (real rows)79.30−- 1.25 †0.0725
Appendix
Table 31: G-Retriever (GCN) , 2000 test questions.
ConditionAccuracyDifferencepp
REAL81.40––
MEAN81.25−- 0.150.25
SWAP81.40++ 0.001
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})81.25−- 0.150.25
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})80.95−- 0.45 **0.00391
NODE (real rows)81.40++ 0.001
Appendix
Table 32: CONST , 2000 test questions.
ConditionAccuracyDifferencepp
REAL69.15––
MEAN48.95−- 20.20 ***6.8×10−486.8\times 10^{-48}
SWAP34.10−- 35.05 ***1.11×10−1271.11\times 10^{-127}
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})49.05−- 20.10 ***2.05×10−472.05\times 10^{-47}
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})38.85−- 30.30 ***4.36×10−1164.36\times 10^{-116}
NODE (real rows)60.25−- 8.90 ***7.43×10−157.43\times 10^{-15}
Appendix
Table 33: Stage 1 + GELU, projector only , 2000 test questions.
ConditionAccuracyDifferencepp
real72.50––
mean71.10−1.40-1.40.182
swap69.10−3.40-3.40 ***6.31×10−46.31\times 10^{-4}
swap (permutation 2)67.85−4.65-4.65 ***2.32×10−62.32\times 10^{-6}
swap (permutation 3)67.85−4.65-4.65 ***1.21×10−51.21\times 10^{-5}
μ+0(ti−μ)\bm{\mu}+0(\mathbf{t}_{i}-\bm{\mu})70.50−2.00-2.00 *.0481
Appendix
Table 34: Ours, joint, standard recipe , 2,000 SceneGraphs test questions. The reference is real above the midrule and notxt-real below it.
ConditionAccuracyDifferencepp
REAL77.40––
MEAN77.30−- 0.100.918
SWAP77.20−- 0.200.801
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})77.35−- 0.051
NODE (real rows)77.50++ 0.100.856
NODE (Gaussian)77.35−- 0.051
Appendix
Table 35: G-Retriever, half split , 2000 test questions.
ConditionAccuracyDifferencepp
REAL76.90––
MEAN75.40−- 1.50 †0.0829
SWAP72.45−- 4.45 ***2.02×10−72.02\times 10^{-7}
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})74.95−- 1.95 *0.0242
NODE (real rows)75.85−- 1.050.158
NODE (Gaussian)76.10−- 0.800.323
Appendix
Table 36: Ours, joint, cross-fitted, initialisation 1 , 2000 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL77.55––
MEAN76.45−- 1.100.191
SWAP75.10−- 2.45 **0.00146
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})76.50−- 1.050.214
NODE (real rows)78.55++ 1.000.116
NODE (Gaussian)78.15++ 0.600.366
Appendix
Table 37: Ours, joint, cross-fitted, initialisation 2 , 2000 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL79.65––
MEAN73.90−- 5.75 ***2.1×10−112.1\times 10^{-11}
SWAP70.05−- 9.60 ***1.2×10−231.2\times 10^{-23}
NODE (real rows)79.75++ 0.100.73
EDGE (rewired)80.05++ 0.400.56
GSWAP (other graph)80.00++ 0.350.54
Appendix
Table 38: Ours, no entity features , 2000 test questions. Best epoch 3 of 5, with validation loss 0.326, 0.287, 0.283, 0.2826, 0.302 over epochs 0 to 4.
ConditionAccuracyDifferencepp
REAL70.33––
MEAN68.49−- 1.84 *0.0285
SWAP69.41−- 0.920.311
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})70.70++ 0.370.687
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})70.95++ 0.610.495
NODE (real rows)70.39++ 0.061
Appendix
Table 39: G-Retriever (graph transformer) , 1628 test questions.
ConditionAccuracyDifferencepp
REAL74.32––
MEAN73.89−- 0.430.634
SWAP73.59−- 0.740.382
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})73.89−- 0.430.636
NODE (real rows)73.34−- 0.980.224
NODE (Gaussian)73.59−- 0.740.359
Appendix
Table 40: G-Retriever + LoRA , 1628 test questions.
ConditionAccuracyDifferencepp
REAL55.47––
MEAN54.98−- 0.490.48
SWAP54.79−- 0.680.428
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})54.98−- 0.490.48
NODE (real rows)54.85−- 0.610.378
NODE (Gaussian)54.67−- 0.800.232
Appendix
Table 41: G-Retriever with Qwen2.5-3B , 1628 test questions.
ConditionAccuracyDifferencepp
REAL64.62––
MEAN49.88−- 14.74 ***1.26×10−301.26\times 10^{-30}
SWAP54.73−- 9.89 ***4.62×10−154.62\times 10^{-15}
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})49.88−- 14.74 ***1.26×10−301.26\times 10^{-30}
μ+2(ti−μ)\bm{\mu}+2(\mathbf{t}_{i}-\bm{\mu})37.65−- 26.97 ***8.63×10−828.63\times 10^{-82}
NODE (real rows)62.22−- 2.40 *0.0289
Appendix
Table 42: Stage 1 + GELU, projector only , 1628 test questions.
ConditionAccuracyDifferencepp
REAL69.90––
MEAN66.77−- 3.13 ***0.000865
SWAP67.75−- 2.15 *0.0387
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})67.38−- 2.52 **0.00807
NODE (real rows)68.67−- 1.230.216
NODE (Gaussian)68.55−- 1.350.157
Appendix
Table 43: Ours, joint (the method) , 1628 test questions.
ConditionAccuracyDifferencepp
REAL74.57––
MEAN73.46−- 1.110.195
SWAP72.24−- 2.33 **0.00839
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})72.60−- 1.97 *0.0192
NODE (real rows)73.77−- 0.800.332
NODE (Gaussian)73.40−- 1.170.161
Appendix
Table 44: Ours, joint + LoRA , 1628 test questions.
ConditionAccuracyDifferencepp
REAL71.25––
MEAN71.44++ 0.180.69
SWAP71.25++ 0.001
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})71.44++ 0.180.69
NODE (real rows)71.25++ 0.001
NODE (Gaussian)71.25++ 0.001
Appendix
Table 45: CONST , 1628 test questions.
ConditionAccuracyDifferencepp
REAL70.70––
MEAN70.21−- 0.490.115
SWAP68.98−- 1.72 *0.015
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})70.21−- 0.490.115
NODE (real rows)70.70++ 0.001
NODE (Gaussian)70.70++ 0.001
Appendix
Table 46: MLPQ , 1628 test questions.
ConditionAccuracyDifferencepp
REAL65.79––
MEAN63.70−- 2.09 **0.00187
SWAP63.88−- 1.90 **0.00573
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})63.70−- 2.09 **0.00187
NODE (real rows)64.99−- 0.800.25
NODE (Gaussian)65.17−- 0.610.373
Appendix
Table 47: Ours, joint, Qwen2.5-3B , 1628 test questions.
ConditionAccuracyDifferencepp
REAL57.13––
MEAN56.57−- 0.550.478
SWAP55.71−- 1.410.104
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})56.70−- 0.430.595
NODE (real rows)56.82−- 0.310.704
NODE (Gaussian)56.57−- 0.550.452
Appendix
Table 48: G-Retriever, Qwen2.5-3B, matched budget , 1628 test questions.
ConditionAccuracyDifferencepp
REAL70.58––
MEAN70.02−- 0.550.538
SWAP70.15−- 0.430.654
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})69.90−- 0.680.46
NODE (real rows)70.33−- 0.250.784
NODE (Gaussian)70.58++ 0.001
Appendix
Table 49: G-Retriever, half split, initialisation 1 , 1628 test questions.
ConditionAccuracyDifferencepp
REAL67.75––
MEAN67.69−- 0.061
SWAP67.26−- 0.490.614
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})67.57−- 0.180.89
NODE (real rows)67.26−- 0.490.598
NODE (Gaussian)67.63−- 0.120.942
Appendix
Table 50: G-Retriever, half split, initialisation 2 , 1628 test questions.
ConditionAccuracyDifferencepp
REAL70.02––
MEAN70.02++ 0.001
SWAP69.78−- 0.250.836
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})68.86−- 1.170.22
NODE (real rows)69.16−- 0.860.356
NODE (Gaussian)70.52++ 0.490.617
Appendix
Table 51: G-Retriever, half split, initialisation 3 , 1628 test questions.
ConditionAccuracyDifferencepp
REAL63.14––
MEAN59.71−- 3.44 **0.00341
SWAP57.80−- 5.34 ***6.9×10−66.9\times 10^{-6}
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})60.01−- 3.13 **0.0045
NODE (real rows)61.49−- 1.660.127
NODE (Gaussian)62.16−- 0.980.385
Appendix
Table 52: Projector only, cross-fitted , 1628 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL63.64––
MEAN60.69−- 2.95 **0.0075
SWAP58.42−- 5.22 ***1.41×10−61.41\times 10^{-6}
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})59.28−- 4.36 ***7.12×10−57.12\times 10^{-5}
NODE (real rows)60.87−- 2.76 **0.00953
NODE (Gaussian)62.59−- 1.040.324
Appendix
Table 53: Late-prompt variant, cross-fitted , 1628 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL64.56––
MEAN60.32−- 4.24 ***5.2×10−55.2\times 10^{-5}
SWAP60.87−- 3.69 ***0.000608
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})61.86−- 2.70 **0.00801
NODE (real rows)63.27−- 1.290.216
NODE (Gaussian)63.21−- 1.350.196
Appendix
Table 54: Projector dropout 0.5, cross-fitted , 1628 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL67.20––
MEAN68.55++ 1.350.143
SWAP66.89−- 0.310.794
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})68.92++ 1.72 †0.0511
NODE (real rows)69.23++ 2.03 *0.0281
NODE (Gaussian)68.49++ 1.290.172
Appendix
Table 55: Text-dropout variant, cross-fitted , 1628 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL69.53––
MEAN68.00−- 1.54 †0.0967
SWAP66.34−- 3.19 **0.00151
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})67.87−- 1.66 †0.0732
NODE (real rows)66.28−- 3.26 ***0.000443
NODE (Gaussian)68.24−- 1.290.174
Appendix
Table 56: Ours, joint, cross-fitted, initialisation 1 , 1628 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL66.89––
MEAN65.11−- 1.78 †0.0888
SWAP65.85−- 1.040.322
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})63.70−- 3.19 **0.00139
NODE (real rows)64.68−- 2.21 *0.0298
NODE (Gaussian)65.29−- 1.600.101
Appendix
Table 57: Ours, joint, cross-fitted, initialisation 2 , 1628 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL68.43––
MEAN68.18−- 0.250.848
SWAP68.30−- 0.120.948
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})68.73++ 0.310.797
NODE (real rows)68.37−- 0.061
NODE (Gaussian)68.80++ 0.370.737
Appendix
Table 58: Ours, joint, cross-fitted, initialisation 3 , 1628 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL70.95––
MEAN69.78−- 1.170.166
SWAP70.02−- 0.920.279
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})69.53−- 1.41 †0.0845
NODE (real rows)69.47−- 1.47 †0.0669
NODE (Gaussian)69.59−- 1.35 †0.0945
Appendix
Table 59: Ours, joint + LoRA, cross-fitted , 1628 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL72.17––
MEAN71.93−- 0.250.805
SWAP71.99−- 0.180.859
μ+0⋅(ti−μ)\bm{\mu}+0\cdot(\mathbf{t}_{i}-\bm{\mu})70.82−- 1.35 †0.0882
NODE (real rows)71.87−- 0.310.738
NODE (Gaussian)71.56−- 0.610.456
Appendix
Table 60: G-Retriever + LoRA, cross-fitted , 1628 test questions, cross-fitted recipe.
ConditionAccuracyDifferencepp
REAL72.73––
MEAN66.58−- 6.14 ***1.28×10−91.28\times 10^{-9}
SWAP67.08−- 5.65 ***5.88×10−95.88\times 10^{-9}
NODE (real rows)72.11−- 0.610.415
EDGE (rewired)72.30−- 0.430.664
GSWAP (other graph)70.76−- 1.97 *0.0347
Appendix
Table 61: Ours, no entity features , 1628 test questions.
ConditionAccuracyDifferencepp
REAL69.84––
MEAN69.78−- 0.061
SWAP66.15−- 3.69 ***0.000365
NODE (real rows)67.75−- 2.09 *0.0367
EDGE (rewired)69.35−- 0.490.636
GSWAP (other graph)67.51−- 2.33 *0.0247
Appendix
Table 62: Ours, no question conditioning , 1628 test questions.
DatasetModelREALzero tokennorm →\rightarrow word×0.5\times 0.5×2\times 2
ExplaGraphsG-Retriever (GT)85.56−81.95-81.95 ***−81.95-81.95 ***−8.12-8.12 ***−3.61-3.61 ***
CONST85.56−83.39-83.39 ***−80.14-80.14 ***−23.83-23.83 ***−2.17-2.17 †
CONST at word norm85.56−83.39-83.39 ***0.000.00−80.69-80.69 ***−78.52-78.52 ***
SceneGraphsG-Retriever (GT)81.35−70.05-70.05 ***−70.00-70.00 ***−4.50-4.50 ***−0.80-0.80
CONST81.40−69.20-69.20 ***−66.60-66.60 ***−6.90-6.90 ***−1.95-1.95 **
WebQSPG-Retriever (GT)70.33−59.64-59.64 ***−65.48-65.48 ***−7.31-7.31 ***−4.24-4.24 ***
Appendix
Table 63: Conditions that keep the token’s direction and change its length, as differences from REAL. † p<.10p<.10 , * p<.05p<.05 , ** p<.01p<.01 , *** p<.001p<.001 , exact McNemar, unmarked p≥.10p\geq.10 . Blank cells were not run. ExplaGraphs 554 questions, SceneGraphs the first 2,000, WebQSP 1,628.
ModelTrainable parametersExplaGraphsSceneGraphsWebQSP
G-Retriever31,492,09985.5681.3570.33
const10,492,92885.5681.4071.25
mlpq14,690,30484.84–70.70
pt40,96060.6070.2252.21
ours, Stage 220,990,97985.7472.5069.90
Appendix
Table 64: Graph-free controls. Score is accuracy on ExplaGraphs and SceneGraphs and Hit@1 on WebQSP ; parameter counts are from the WebQSP checkpoints with the LLM frozen and its embedding table excluded. Dashes are configurations we did not train.
DatasetModelREALMEANSWAPGSWAP
WebQSPG-Retriever0.7558+0.0131+0.0131 [+.0097,+.0166][+.0097,+.0166]+0.0183+0.0183 [+.0137,+.0229][+.0137,+.0229]+0.0182+0.0182 [+.0136,+.0228][+.0136,+.0228]
ours0.8540−0.0159-0.0159 [−.0288,−.0041][-.0288,-.0041]+0.0895+0.0895 [+.0727,+.1074][+.0727,+.1074]+0.0332+0.0332 [+.0185,+.0461][+.0185,+.0461]
ours, no entity features0.8213+0.0378+0.0378 [+.0257,+.0506][+.0257,+.0506]+0.1111+0.1111 [+.0954,+.1261][+.0954,+.1261]+0.0262+0.0262 [+.0175,+.0348][+.0175,+.0348]
ours, no question0.8517+0.0070+0.0070 [−.0071,+.0203][-.0071,+.0203]+0.0848+0.0848 [+.0671,+.1020][+.0671,+.1020]+0.0845+0.0845 [+.0668,+.1015][+.0668,+.1015]
G-Retriever, Qwen2.5-3B1.0367+0.0109+0.0109 [+.0038,+.0185][+.0038,+.0185]+0.0734+0.0734 [+.0623,+.0845][+.0623,+.0845]+0.0737+0.0737 [+.0626,+.0850][+.0626,+.0850]
ours, Qwen2.5-3B1.0553+0.0123+0.0123 [+.0066,+.0177][+.0066,+.0177]+0.0168+0.0168 [+.0106,+.0226][+.0106,+.0226]+0.0144+0.0144 [+.0076,+.0211][+.0076,+.0211]
Appendix
Table 65: Paired gold-answer NLL differences, condition minus real , with 95% bootstrap intervals; positive means the intervention made the gold answer less likely. Every cell whose pp -value we measured is significant under a Wilcoxon signed-rank test (largest p=10−4p=10^{-4} ), so the criterion we apply is the effect size: the same weights and condition rerun give +0.0003+0.0003 [−0.0003,+0.0008][-0.0003,+0.0008] , and we require 0.030.03 , a hundred times that floor, before reading a likelihood effect. The Qwen2.5-3B rows are the matched-budget pair of Table 4 .
DatasetModelQ rankA rankG rankΔ\Delta Score gswapΔ\Delta NLL S/GSDiagnosis
Llama-2-7B
WebQSPG-Retriever293317369−0.55-0.55+0.0183/+0.0182+0.0183/+0.0182soft prompt
mlpq407407407–––
Ours6560315−1.97-1.97 †+0.0895/+0.0332\mathbf{+0.0895}/\mathbf{+0.0332}masked
no entity features3451303−1.97-1.97 *+0.1111/+0.0262\mathbf{+0.1111}/+0.0262prior
no question120134329−2.33\mathbf{-2.33} *+0.0848/+0.0845\mathbf{+0.0848}/\mathbf{+0.0845}used
Appendix
Table 66: Complete token-content and use results across LLMs. Q, A, and G are median retrieval ranks; lower is better. Δ\Delta Score is the with-text gswap effect, and NLL reports swap / gswap . Bold score effects satisfy ∣Δ∣>2|\Delta|>2 and p<.05p<.05 ; bold NLL values satisfy the likelihood criterion. Diagnoses also use Tables 4 , 70 , and 71 . The binary ExplaGraphs answer is omitted.
RunTargetValidation top-1Graph gap
SceneGraphs , with questionfull answer sentence0.4330.035→0.2040.035\rightarrow 0.204
SceneGraphs , without questionfull answer sentence0.084Same embeddings, N/A
SceneGraphs , scene-disjoint halvesfull answer sentence0.3760.035→0.2640.035\rightarrow 0.264
WebQSP , half splitgold answers0.469N/A
ExplaGraphs , half splitverbalised graph0.501N/A
Appendix
Table 67: Stage 1 measured without the LLM. Top-1 is computed against all validation targets of the benchmark, 19,997 on SceneGraphs .
VariantHit@1Token cosineMEANSWAPNODE
Joint Stage 2, as used in the paper69.530.1778−1.54-1.54 †−3.19-3.19 **−3.26-3.26 ***
Projector only63.140.5700−3.44-3.44 ***−5.34-5.34 ***−1.66-1.66
Prompt token added after the projector63.640.5700−2.95-2.95 ***−5.22-5.22 ***−2.76-2.76 ***
Dropout 0.5 on the projector64.560.7563−4.24-4.24 ***−3.69-3.69 ***−1.29-1.29
40% of verbalised edges hidden in training67.200.1847+1.35+1.35−0.31-0.31+2.03+2.03 **
Appendix
Table 68: Alternatives to the joint Stage 2. WebQSP , cross-fitted recipe, 1,628 test questions. MEAN, SWAP and NODE are differences from each model’s own REAL, with the exact McNemar test: † p<.10p<.10 , * p<.05p<.05 , ** p<.01p<.01 , *** p<.001p<.001 , unmarked p≥0.1p\geq 0.1 .
DatasetModelrealToken cos.Δ\Delta meanΔ\Delta swapΔ\Delta nodeΔ\Delta edgeΔ\Delta gswap
Llama-2-7B
WebQSPprojector only64.620.543−14.74-14.74 ***−9.89-9.89 ***−2.40-2.40 *+0.06+0.06–
Ours69.900.142−3.13-3.13 ***−2.15-2.15 *−1.23-1.23−0.80-0.80−1.97-1.97 †
no entity features72.730.125−6.14-6.14 ***−5.65-5.65 ***−0.61-0.61−0.43-0.43−1.97-1.97 *
no question69.840.169−0.06-0.06−3.69-3.69 ***−2.09-2.09 *−0.49-0.49−2.33-2.33 *
cross-fitted69.530.178−1.54-1.54 †−3.19-3.19 **−3.26-3.26 ***−0.55-0.55−2.83-2.83 **
Appendix
Table 69: Pipeline variants under fixed-weight interventions. Ours denotes the complete two-token pipeline; indented rows are retrained diagnostic variants. Projector only is the single-token model with the aligned frozen encoder but no separate prompt token. Intervention columns report condition minus real , so negative values indicate worse performance. Lower token cosine indicates greater input dependence. Dashes denote conditions not evaluated.
DatasetModelnotxtΔ\Delta meanΔ\Delta swapΔ\Delta gswapWith text
WebQSPG-Retriever41.34+1.04+1.04−0.18-0.18−0.86-0.8670.33
Ours28.75−2.09-2.09 *−2.40-2.40 *−2.27\mathbf{-2.27} *69.90
no entity features28.38−7.49-7.49 ***––72.73
no question30.90+0.43+0.43––69.84
SceneGraphsG-Retriever30.20+0.45+0.45−1.40-1.40–81.35
Ours32.60−6.60-6.60 ***−10.85-10.85 ***−3.10\mathbf{-3.10} **72.50
Appendix
Table 70: Llama-2 evaluation without the verbalised graph. Intervention columns report differences from notxt ; the final column gives the corresponding score with the verbalised graph present. Bold gswap effects satisfy ∣Δ∣>2|\Delta|>2 and p<.05p<.05 . Dashes denote conditions not evaluated.
DatasetModelnotxtΔ\Delta meanΔ\Delta swapΔ\Delta gswapWith text
WebQSPG-Retriever29.30+0.86+0.86−2.03-2.03 †−1.84-1.84 †57.13
Ours34.34+0.31+0.31+1.11+1.11+1.11+1.1165.79
no entity features30.10+1.97+1.97 *––58.54
no question32.74+0.61+0.61––60.14
SceneGraphsG-Retriever39.50+1.95+1.95 *––76.05
Ours43.90−7.10-7.10 ***––72.65
Appendix
Table 71: Qwen2.5-3B evaluation without the verbalised graph. Intervention columns report differences from notxt ; the final column gives the corresponding score with the verbalised graph present. Bold gswap effects satisfy ∣Δ∣>2|\Delta|>2 and p<.05p<.05 . Dashes denote conditions not evaluated.
DatasetStage 2AccuracyBaselineMEANSWAPNODEEDGE
WebQSPprojector only63.1470.52−3.44-3.44 ***−5.34-5.34 ***−1.66-1.66+1.04+1.04
WebQSPprojector + prompt69.5370.58−1.54-1.54 †−3.19-3.19 **−3.26-3.26 ***−0.55-0.55
SceneGraphsprojector + prompt, run 176.9078.57−1.50-1.50 †−4.45-4.45 ***−1.05-1.05−1.95-1.95 ***
SceneGraphsprojector + prompt, run 277.5578.57−1.10-1.10−2.45-2.45 **+1.00+1.00+0.30+0.30
ExplaGraphsprojector + prompt76.3580.87−3.61-3.61 *+1.08+1.08+0.72+0.72+0.18+0.18
Appendix
Table 72: Cross-fitted runs, in which Stage 1 and Stage 2 see disjoint halves of the training questions. Accuracy is compared with the G-Retriever baseline trained on the same half; the intervention columns are differences from the model’s own REAL, with the exact McNemar test. WebQSP 1,628 test questions, SceneGraphs the first 2,000, ExplaGraphs 554.
ModelRun 1Run 2Run 3Mean ±\pm sdMEANSWAPNODE
G-Retriever70.5867.7570.0269.45±1.5069.45\pm 1.50≤0.53\leq 0.53≤0.53\leq 0.53≤0.53\leq 0.53
Ours, joint69.5366.8968.4368.28±1.3368.28\pm 1.33−1.19-1.19 **−1.45-1.45 **−1.84-1.84 ***
Appendix
Table 73: Three initialisations on WebQSP , cross-fitted recipe, 1,628 test questions each. The intervention columns are differences from REAL pooled over the three runs, with the exact McNemar test.
More in this topic

Graph Prompt Learning

See all 10 papers

Explore similar work

CardsList

  1. Is One Token All It Takes? Graph Pooling Tokens for LLM-based GraphQA

    Apr 1, 2026Ankit Grover, Lodovico Giaretta, Rémi Bourgerie +1LLM Reasoning with GraphsGraph Pooling

    KTH Royal Institute of Technology, Stockholm, Sweden · RISE Research Institutes of Sweden, Stockholm, Sweden

  2. Revisiting Graph-Tokenizing Large Language Models: A Systematic Evaluation of Graph Token Understanding

    May 5, 2026Zhongjian Zhang, Yue Yu, Mengmei Zhang +3LLM EvaluationLLM Reasoning with Graphs

    Beijing University of Posts and Telecommunications · China Telecom Bestpay · Beihang University

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

    Jun 2, 2026Ding Zhang, Runtao Zhou, Wenqing Zheng +3LLM InterpretabilityMechanistic Interpretability

    University of Virginia · Capital One