When Links Mislead: Label-Free Structural Polarity for Zero-Shot Classification of Web Pages
Authors: Hanshuai Cui, Zhiqing Tang, Zhi Yao, Weijia Jia, Jun Luo
Organizations: School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai 519087, China · College of Computing and Data Science, Nanyang Technological University, Singapore
Web platforms such as digital libraries, online stores and university sites must classify new pages before anyone labels them. Large language models (LLMs) can label each page from its text, but these labels are noisy, and links could correct them only where linked pages share classes. Existing zero-shot methods ignore the links, or pull a page toward its neighbours as label propagation and graph neural networks (GNNs) do, which hurts where linked pages differ in class. Others weight propagation by an LLM-estimated homophily that costs extra calls and drifts across LLMs. We find that the sign with which links should enter a prediction, which we call structural polarity, is readable without true labels. It is the excess agreement of LLM labels and text-prototype predictions along links. We thus propose ZeroPolar, which gates one signed propagation step by this statistic. Training-free ZeroPolar mixes prototype scores with any number of LLM labels, and ZeroPolar-T denoises the labels with a text classifier once half of the pages carry one. With one GPT-4o label per page, ZeroPolar-T reaches the highest mean accuracy on eleven graphs (73.99%), 6.01 points above LLM-BP and significantly above all eighteen baselines we compare. With LLM labels on only 5% of the pages, training-free ZeroPolar leads a GNN trained on these labels by 10.80 points.
Figures & tables
Figure 1. ZeroPolar measures links instead of trusting them. A professor’s page has four course in-links, so a step toward them gives course, while ZeroPolar finds that labels barely agree along the graph’s links and pushes away. On the left, a professor's home page from the Wisconsin graph, whose true class is faculty, has four in-links from course pages, with a key for solid same-class and dashed different-class links. On the right, two stacked lanes. The upper lane, Trust links, assumes that linked pages share a class, pulls the page toward its neighbours and labels it course, marked wrong. The lower lane, \sys{}, reads that zero-shot labels along the graph's links agree barely above chance, sets the marker on the minus side of a minus and plus scale, pushes the page away from its neighbours and keeps faculty, marked correct.
Figure 2. Reading polarity on eleven graphs. (a, c) LLM estimates of the homophily H (§ 3.1 ) cost 500 calls and drift from the truth. (b, d) Excess agreement along edges (§ 4.2 ) splits the graphs at δ . Four panels in a two-by-two grid, two schematics in the top row and two scatter plots in the bottom row. First, a small graph whose sampled edges go to two LLMs that answer whether the endpoints share a class, giving one homophily estimate below a dashed true value and one above it by a drift. Second, the excess agreement, the share of edges whose endpoints carry the same label minus the chance agreement, which sets the sign to plus when it reaches delta. Third, homophily estimated by GPT-4o-mini and DeepSeek-V4-Flash against the true homophily on eleven graphs, with the two estimates on Children on either side of the diagonal and DeepSeek-V4-Flash well above it. Fourth, the excess agreement of LLM labels and of prototype predictions against the true excess agreement, with all homophilous graphs above the threshold line and all WebKB graphs below it.
Figure 3. Overview of ZeroPolar. The gate (b) reads σ (§ 4.2 ) from the signals in (a), and the budget routes it to ZeroPolar (c) or ZeroPolar-T (d). The bottom row shows two measured cases, on Cora and Texas. Framework diagram in four panels and a bottom row. Panel a shows a graph of web pages and an LLM that reads each page with the class names and labels it with a class, some of them wrong, and a text encoder whose embeddings meet class prototypes built from twenty LLM labels per class, giving the prototype prediction of each node. Panel b draws the prototype predictions and the LLM labels on two small graphs with solid same-label and dashed different-label edges, computes the excess agreement of each, takes their maximum and compares it with delta, which sets the sign sigma, plus to pull toward the neighbours and minus to push away. A bus carries sigma to panel c, \sys{}, if fewer than half of the test nodes carry an LLM label, and to panel d, \sysT{}, otherwise. \sys{} mixes softmax scores with the one-hot LLM label, while \sysT{} trains an MLP on the LLM labels. Both apply the same signed step, whose formula sits in panel d, and predict the class with the highest score. The bottom row shows a Cora paper whose MLP label Rule Learning becomes Probabilistic Methods after a plus step, since the MLP gives most of its in-neighbours Probabilistic Methods, and a Texas page whose MLP label faculty becomes student after a minus step that pushes it away from the faculty label of most of its in-neighbours.
Figure 4. Accuracy change from one step of Eq. ( 1 ) on prototype scores, toward or away from the neighbours. The better sign flips where the statistic s crosses δ (§ 4.2 ), and the gate picks it on every graph. A scatter over eleven graphs of the accuracy change from one propagation step toward the neighbours, blue dots, or away from them, brown triangles, against the label-free statistic s, with a vertical dashed line at delta. Left of the line lie the four WebKB graphs, where the step toward the neighbours loses about ten points and the step away loses little or gains. Right of it lie the seven homophilous graphs, where the step toward the neighbours gains a little and the step away loses much more. Orange rings mark the sign that the gate picks, which is the better sign on every graph.
Design graphs
Other graphs
Mean
Method
Calls
Cora
Cite
Pub
Wiki
Corn
Hist
Child
Sport
Tex
Wisc
Wash
Design
Other
All
Vanilla ( Li et al., 2025 )
0
68.610
67.099
76.363
63.171
85.288
58.335
28.940
59.372
78.342
75.572
66.783
72.106
61.224
66.171
LLM direct ( OpenAI, 2024b )
N
70.295
64.770
89.858
66.102
45.550
53.303
30.764
66.358
63.102
56.604
48.908
67.315
53.173
59.601
Training-free
LLM labels + LP ( Zhou et al., 2003 )
N
75.092
68.355
91.202
72.738
48.168
61.858
32.800
71.596
67.914
58.491
51.092
71.111
57.292
63.573
LLM-BP ( Wang et al., 2025b )
20C+500
72.589
69.508
75.547
67.746
83.264
59.860
24.807
61.920
81.658
77.748
73.144
73.731
63.190
67.981
LLM-BP (appr.) ( Wang et al., 2025b )
20C
71.415
68.658
76.813
67.964
84.921
59.490
29.402
61.512
79.394
75.648
70.044
73.954
62.582
67.751
Table 1. Zero-shot accuracy (%) with GPT-4o labels on eleven graphs.
Method
Computer
arXiv
Mean
Vanilla ( Li et al., 2025 )
36.41
67.43
51.92
LLM direct
62.10
75.46
68.78
Training-free
LLM labels + LP ( Zhou et al., 2003 )
63.92
75.46
69.69
LLM-BP ( Wang et al., 2025b )
51.99
66.37
59.18
LLM-BP (appr.) ( Wang et al., 2025b )
46.85
67.70
57.27
Gated switch
63.92
75.46
69.69
Table 2. Accuracy (%) with Qwen3.8-27B labels on two additional graphs, Computer and arXiv.
Figure 5. Mean accuracy on the homophilous against the WebKB graphs, where the dashed line joins the baselines that no other beats on both (a), and the lead of ZeroPolar-T by in-degree with 95% CIs (b), with GPT-4o labels. Two panels. Left, a scatter of the homophilous-graph mean against the WebKB mean, with baselines marked by family and a dashed line through LLM-BP, LLM-BP (appr.), ACM-GCN h64 and SAGE h256, which no other baseline beats on both axes. \sysT{}, a star near 71 and 79 percent, lies above and to the right of every baseline, \sys{} is close to it, and GOFA-T lies left of the frame. Right, bars for the lead of \sysT{} in four in-degree buckets over SAGE h64, MLP h256, LLM-BP and \sys{} with interval bars, about 5 to 8 points over LLM-BP in every bucket and near zero over SAGE h64 at six or more in-edges.
Variant
Design
Other
All
ZeroPolar-T
No step (MLP only)
76.69
69.44
72.74
Step always +
74.39
64.11
68.78
Step always −
42.47
55.28
49.45
ZeroPolar-T
78.68
70.08
73.99
ZeroPolar
No gate (always + )
74.47
60.18
66.68
No step
77.95
65.20
71.00
Table 3. Controls that remove or flip one part of each implementation, GPT-4o labels.
GPT-4o
Qwen3-8B
Method
All
Other
All
Other
Addl.
LLM direct
48.98
43.79
50.10
43.19
27.41
LLM-BP ( Wang et al., 2025b )
58.54
49.27
49.15
42.63
26.89
Gated switch
62.75
53.09
52.75
42.73
30.52
ZeroPolar
63.38
53.56
55.52
46.27
29.65
SAGE h64 ( Hamilton et al., 2017 )
62.71
53.74
54.75
45.10
30.35
Table 4. Macro-F1 averaged over seeds and then over graphs.
Figure 6. Eleven-graph mean accuracy against LLM calls, for ZeroPolar against SAGE h64 and Locle at few labels (a) and for the budget switch of § 4.1 against the best baseline (b), shaded where ours leads. Two line panels on logarithmic axes of LLM calls over the eleven graphs. Left, at few labels, \sys{} starts near 63 percent at about 800 calls with 5 prototype labels per class, reaches about 67 percent at about 1.3 thousand calls with 10, and stays near 67 to 69 percent up to about 49 thousand calls, while SAGE h64 rises from about 42 percent at about 800 calls to about 64 percent and Locle with the zero-shot cache rises from about 54 to about 62 percent over its four budgets between about 1.8 thousand and 14 thousand calls, both below \sys{} throughout. Right, from about 64 thousand to about 257 thousand calls around the full budget of one LLM label per test node, our line, which follows the budget switch between the two implementations, rises from about 70 percent to about 74 percent at the full budget and stays flat beyond it. The best baseline, the per-budget maximum of all \nNbaseWbaselines, stays near 69 percent up to about 90 thousand calls and rises to about 73 percent at the full budget. Our line is above it across the whole axis, and the gap is shaded.
Qwen3.8-27B
Qwen3-8B
Llama-3.1-8B
Mistral-7B
Method
Other
All
Addl.
Other
All
Addl.
Other
All
Addl.
Other
All
Vanilla
61.87
66.31
51.92
50.79
58.40
38.55
53.69
60.01
38.16
52.28
58.49
LLM direct
68.91
72.15
68.78
53.70
59.34
44.37
54.12
58.73
47.27
53.66
56.69
Training-free
LLM labels + LP
68.91
72.15
69.69
53.65
59.33
44.64
54.07
58.74
47.31
53.92
57.04
LLM-BP
63.66
66.36
59.18
51.85
55.60
43.22
57.87
59.31
41.20
49.10
45.23
LLM-BP (appr.)
62.75
67.54
57.27
52.15
60.08
43.98
56.19
62.33
42.88
54.05
60.47
Table 5. Mean accuracy (%) with open LLM annotators. A dash marks a setting not run.
Figure 7. Eleven-graph mean accuracy with GPT-4o labels when one setting of ZeroPolar or ZeroPolar-T changes. Six small line panels, one per setting, beta plus, beta minus, delta, tau, w and the number of steps, on a shared accuracy axis from 66 to 76 percent. ZeroPolar stays near 72 percent across beta plus, beta minus, delta and tau and drops with more steps. ZeroPolar-T, drawn for w, delta and steps, stays near 74 percent and drops with two steps.
Method
k=1
k=3
k=5
k=10
Vanilla
46.51
64.12
68.88
73.61
Vanilla + NA
41.78
57.66
63.30
68.64
LLM-BP
47.88
65.13
69.10
73.34
LLM-BP (appr.)
47.70
65.45
69.64
74.03
ZeroPolar, no LLM labels
47.32
64.76
69.02
73.82
ZeroPolar
67.63
73.66
74.91
76.16
Table 6. Eleven-graph mean accuracy in the few-shot setting of LLM-BP with k true labels per class and GPT-4o labels.
Qwen3-8B
Llama-3.1-8B
Method
Inst.
Reddit
Photo
Mean
Inst.
Reddit
Photo
Mean
LLM-BP ( Wang et al., 2025b )
62.75
50.50
44.33
52.53
61.67
52.02
24.80
46.16
ZeroPolar
61.05
54.52
44.82
53.46
60.53
55.42
48.30
54.75
MLP h256 ( Chen et al., 2024d )
42.28
46.57
42.79
43.88
44.65
47.23
48.52
46.80
SAGE h256 ( Hamilton et al., 2017 )
41.99
46.72
43.04
43.92
44.80
47.12
48.93
46.95
SAGE h64 ( Hamilton et al., 2017 )
41.59
46.11
42.96
43.56
45.09
46.76
49.67
47.17
Table 7. Accuracy on three further Web platforms with Qwen3-8B or Llama-3.1-8B as the annotator.
Figure 8. Eleven-graph mean accuracy as injected noise replaces GPT-4o labels by random classes, for ZeroPolar-T and baselines (a) and for two gates of ZeroPolar (b). Two line panels against the injected noise rate from zero to forty percent. Left, \sysT{} has the highest mean at every rate, above SAGE h256, SAGE h64, MLP h256, the gated switch and a flat LLM-BP line near 68 percent. Right, \sys{} with the larger of the two statistics declines slowly, while the gate that reads only the LLM labels falls steeply at high noise, with solid lines for clean prototype labels and dashed lines for noisy ones.
Figure 9. Two nodes repaired by the signed step of ZeroPolar-T , with the gate at +1 on Cora and −1 on Texas. Frames mark correct and wrong classes. Two cases stacked. Above, a Cora paper on a non-axiomatic reasoning system whose GPT-4o label is correct, whose MLP prediction is Rule Learning and whose seven in-neighbours all belong to Probabilistic Methods, and the plus step returns it to Probabilistic Methods. Below, a Texas home page that GPT-4o and the MLP label faculty, whose three in-neighbours are a course, a faculty and a project page, and the minus step moves it to student.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 10. Eleven-graph mean accuracy against LLM calls for both implementations and every baseline. Line plot on a logarithmic axis of LLM calls with a three-column legend below it. LLM-BP (appr.) and LLM-BP appear as short curves over their prototype budgets near 63 to 69 percent with dashed flat lines beyond their last point, and Vanilla, which makes no call, is a dashed horizontal line near 66 percent. SAGE h64 and \sysT{} climb from about 40 percent at the smallest budgets to about 72 and 74 percent at the full budget, while \sys{} stays near 68 percent and rises to about 72 percent. Locle with the zero-shot cache is a short curve over its four budgets from about 54 to 62 percent at a few thousand calls. The other baselines, including four heterophily GNNs, appear as single points at one label per test node between about 60 and 73 percent.
Design graphs
Other graphs
Mean
Method
Calls
Cora
Cite
Pub
Wiki
Corn
Hist
Child
Sport
Tex
Wisc
Wash
Design
Other
All
Trained
MLP h64 ( Chen et al., 2024d )
N
75.461
69.512
90.112
74.055
66.318
69.137
35.137
75.227
84.848
71.195
65.211
75.091
66.793
70.565
MLP h256 ( Chen et al., 2024d )
N
75.154
69.486
90.281
72.630
75.742
68.824
34.333
75.263
87.522
75.094
70.742
76.658
68.630
72.279
GCN h64 ( Kipf and Welling, 2017 )
N
77.245
71.720
86.934
74.426
42.932
68.840
34.723
78.153
65.241
58.994
37.991
70.651
57.324
63.382
GCN h256 ( Kipf and Welling, 2017 )
N
76.507
71.447
86.629
74.038
41.361
68.002
33.925
77.656
67.558
59.119
59.825
69.997
61.014
65.097
SAGE h256 ( Hamilton et al., 2017 )
N
77.122
71.551
89.579
74.420
72.775
68.371
35.493
77.896
82.175
75.220
72.926
77.089
68.680
72.502
Appendix
Table 8. Zero-shot accuracy (%) of every trained baseline with GPT-4o labels on eleven graphs.
Design graphs
Other graphs
Mean
Method
Cora
Cite
Pub
Wiki
Corn
Hist
Child
Sport
Tex
Wisc
Wash
Design
Other
All
GraphGPT ( Tang et al., 2024 )
17.48
13.93
42.94
33.59
10.18
12.31
9.94
4.53
18.48
12.35
20.64
23.62
13.04
17.85
LLaGA ( Chen et al., 2024a )
11.62
19.52
7.56
10.98
12.57
7.95
10.09
1.84
15.51
15.09
10.48
12.45
10.16
11.20
OFA ( Liu et al., 2024 )
20.36
41.31
28.18
30.77
29.84
8.25
3.05
15.18
11.77
4.80
6.04
30.09
8.18
18.14
ZeroG ( Li et al., 2024 )
60.40
50.35
74.68
46.74
10.47
36.55
12.72
14.27
53.48
12.66
8.30
48.53
23.00
34.60
UniGLM ( Fang et al., 2025 )
45.57
52.26
70.33
55.05
23.03
44.24
21.48
33.46
21.39
27.16
24.01
49.25
28.62
38.00
Appendix
Table 9. Zero-shot accuracy (%) of six published graph models reported by LLM-BP, next to ours with GPT-4o labels.
Method
Computer
arXiv
Mean
Vanilla ( Li et al., 2025 )
33.16
43.93
38.55
LLM direct
45.30
43.44
44.37
Training-free
LLM labels + LP ( Zhou et al., 2003 )
45.84
43.44
44.64
LLM-BP ( Wang et al., 2025b )
42.92
43.52
43.22
LLM-BP (appr.) ( Wang et al., 2025b )
43.79
44.17
43.98
Gated switch
45.84
43.93
44.89
Appendix
Table 10. Accuracy (%) with Qwen3-8B labels on the two additional graphs, Computer and arXiv.
Qwen3-8B
Llama-3.1-8B
Mistral-7B
Method
Other
All
Addl.
Other
All
Addl.
Other
All
SAGE h64 ( Hamilton et al., 2017 )
57.86
63.44
54.54
54.26
59.23
51.01
57.03
60.63
GPR-GNN h256 ( Chien et al., 2021 )
57.53
62.70
55.16
55.23
58.96
51.45
56.46
58.91
H2GCN h64 ( Zhu et al., 2020 )
59.11
64.40
53.09
55.15
59.86
50.41
57.47
60.44
FAGCN h256 ( Bo et al., 2021 )
58.40
64.19
54.31
53.98
58.79
49.62
57.93
58.83
ACM-GCN h64 ( Luan et al., 2022 )
58.00
63.70
52.46
54.95
59.00
49.44
57.17
60.17
Appendix
Table 11. Mean accuracy (%) of SAGE h64, the four heterophily GNNs and ZeroPolar-T with three open LLM annotators.
Baseline
Setting
Difference [95% CI]
Ahead
MLP h64 ( Chen et al., 2024d )
Patience 50
+3.57[+2.87,+4.33]
MLP h256 ( Chen et al., 2024d )
Patience 50
+1.17[+0.62,+1.76]
GCN h64 ( Kipf and Welling, 2017 )
Patience 50
+7.95[+6.89,+9.05]
GCN h256 ( Kipf and Welling, 2017 )
Patience 50
+8.72[+7.66,+9.81]
SAGE h256 ( Hamilton et al., 2017 )
Patience 50
+0.97[+0.29,+1.71]
SAGE h64 ( Hamilton et al., 2017 )
Patience 50
+1.66[+0.99,+2.35]
Appendix
Table 12. ZeroPolar-T minus each trained baseline with a patience of 50 on the eleven graphs with GPT-4o labels, and over ten training seeds per setting.
Zero-shot text classification assigns texts to categories without task-specific labeled training data, using language models, embeddings, or label descriptions.
Flagship large language models (LLMs) offer rich general knowledge and strong reasoning capabilities, but their high inference cost and latency limit their use for large-scale data labeling. We study zero-shot text classification with semantically meaningful labels and task descriptions, without task-specific labeled examples or classifier feedback. We propose SLKR (Semantic Label Knowledge Retrieval), a label-centric retrieval-augmented generation framework that separates offline knowledge construction from lightweight online inference. Offline, a flagship LLM independently generates label definitions, retrieval keywords, and boundary rules for potentially confusable label pairs, forming a reusable label-knowledge store. Because this construction operates over the label space rather than the input corpus, its one-time flagship cost can be amortized across large-scale inference. Online, hybrid semantic and lexical retrieval first identifies the top-K candidate labels and then selects up to M input-relevant boundary rules connecting candidate-label pairs. A lightweight LLM performs the final classification in a single call using the retrieved definitions and rules. Across three public benchmarks and a real-world multilingual book-classification dataset, SLKR achieves the best results on all four datasets with Qwen3.5-4B and on three of four with Ministral-3-3B, while using compact input contexts and no task-specific labeled examples. Boundary ablations further show that retrieving a small set of input-relevant rules preserves performance close to using all boundaries with substantially less context, supporting efficient reuse of flagship-generated label knowledge for lightweight classification.
Daoping Wu, Dong Wang, Siyu Liu +5
Zhejiang Lab, Hangzhou, China · School of Computer Science and Software Engineering, Harbin Institute of Technology, China · School of Mechanical Engineering, Shanghai Jiao Tong University, China
Link sign prediction (LSP) aims to infer the positive or negative polarity of unobserved links in signed networks. Signed Graph Neural Networks (SGNNs) usually rely on signed-graph structural priors, including structural balance and homophily-like similarity, to guide message passing and prediction. These priors describe population-level tendencies, not guarantees for individual target edges. Their failures are especially costly under severe sign imbalance, where errors on minority and locally conflicting relations are harder to detect and correct. We propose Polarity-Asymmetric Structural Calibration (PASC), a target-edge structural-prior calibration framework for signed link prediction. PASC constructs a structure-only prior representation, estimates a target-edge structural prior score, and compares this score with a local signed-context cue to derive a conflict residual. The residual calibrates signed attention aggregation, target-edge gated fusion, and regime-adaptive optimization. Experiments on five real-world signed network datasets show that PASC consistently achieves the best Macro-F1 among representative baselines, with competitive AUC, Binary-F1, and Micro-F1. Structural-shift experiments further suggest reduced dependence on dense-neighborhood and local-closure shortcuts. Source code is available at https://github.com/iqqGGGGGGG/PASC-for-LSP.
Qiqi Gao, Wenzhuo Song, Xueyan Liu
Northeast Normal University Changchun, Jilin, China · Jilin University Changchun, Jilin, China
The POLAR SemEval-2026 Shared Task aims to detect online polarization and focuses on the classification and identification of multilingual, multicultural, and multi-event polarization. Accurate computational detection of online polarization is challenging due to nuanced rhetoric, implicit framing, and the high cost of human-in-the-loop annotation. Building on recent findings that contextual prompting enables large language models to function as strong polarization detectors, we present a two-stage approach for detecting polarization in social media text that combines structured supervised fine tuning with Direct Preference Optimization (DPO) refinement. We fine tune Qwen 2.5-7B-Instruct with LoRA using an interpretable slot-filling template (target, claim type, manifestation checklist, and justification). We then apply DPO with automatically generated preference pairs to reduce costly false negatives. Our submitted system achieves 0.7664 Macro-F1 on the English test set. Post-submission experiments with Mistral-Nemo-Instruct-2407 and LLM-judge-filtered preference pairs further improve to 0.8162 Macro-F1 (not submitted to CodaBench), surpassing the organiser baseline of 0.7802.
Atharva Gupta, Dhruv Kumar, Yash Sinha
Birla Institute of Technology and Science, Pilani, India