{"id":"296d6f60-e279-4b61-b3cd-07da28f8dff6","arxiv_id":"2509.01660","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Jointly modeling news semantics and news intent as heterogeneous graphs, connected by dynamic alignment pseudo-nodes, improves fake news detection on four benchmarks versus state-of-the-art baselines.","lead":"Researchers propose InSide, a fake news detector that reads both the words of a news article and the article's underlying intent, modeled as two connected graphs. On four news datasets it reports consistently higher accuracy than existing state-of-the-art methods, including intent-aware baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LLM-extracted 'intent' is unvalidated and may be mere paraphrase; the framework-swap test in §4.3.3 only varies taxonomy, so the stable-intent mechanism is not established.","rationale":"The reader's weakest assumption is exactly the load-bearing premise I identify: that LLM-extracted coarse intents are faithful, distinct representations of news intent rather than paraphrases of surface semantics. The paper's own evidence does not test this; the framework-swap test only changes the prompt taxonomy, and no ablation replaces intents with summaries or random text. The empirical headline (SOTA on four datasets) could survive even if the mechanism is wrong, so this concern does not force a rejection, but it does mean the conceptual 'stable intent' claim is unsupported. Since the reader already issued a CONDITIONAL verdict on the basis of unverified empirical support and this concern is part of that, the verdict stands. My proposed summary-replacement experiment would settle the concern by testing whether intent-specific content, as opposed to any semantic paraphrase, is what drives the gains.","tokens_in":19474,"tokens_out":4712,"duration_ms":57131,"concrete_test":"On GossipCop (or Weibo), replace the coarse-grained intent generator prompts q with a single 'summarize the article' prompt, producing C_summary, and keep everything else identical. Retrain and retest InSide with these summary-based coarse nodes. If macF1 is statistically indistinguishable from the reported 0.8653 (or 0.8055 on Weibo), the 'intent' signal contributes no more than a surface paraphrase, and the stable-intent mechanism collapses. As a secondary check, compute BERTScore/entailment between generated intent texts and the source articles; if most are near-paraphrase, the claim is further weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central conceptual claim is that intent is stable while surface semantics shift, and that InSide exploits this stability. Yet the only evidence for intent fidelity is that the LLM is prompted with the authors' belief/plan/desire/outcome taxonomy. The framework-swap test (InSide(C') in §4.3.3) merely shows the method is insensitive to which taxonomy is used, not that the generated 'intents' denote anything beyond a paraphrase of the article. If the coarse intent embeddings are essentially re-encoded surface text, the performance gain attributed to 'intent' could arise entirely from extra model capacity (fine-grained nodes, pseudo-node alignment) or from LLM priors leaking veracity cues (e.g., the generator may implicitly comment on credibility). This would not necessarily falsify Table 2, but it would invalidate the paper's stated mechanism and the claim that gains come from stable intent. The paper provides no human validation of intent outputs, no comparison of intent vs. paraphrase embeddings, and no ablation replacing coarse intents with surface-level rewrites. The limitations section is silent on this missing support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InSide, a graph-based joint modeling framework for fake news detection that combines semantic signals (sentence and entity graphs) with LLM-extracted intent signals (coarse and fine-grained intent nodes). The method builds two heterogeneous graphs, applies dual-level local/global message passing, and introduces a dynamic pathway-based alignment module with pseudo nodes to bridge the semantic-intent representation gap. Experiments on PolitiFact, GossipCop, Weibo, and LLMFake report consistent improvement over intent-agnostic and intent-aware baselines on accuracy, macro-F1, AUC, and class-wise F1. The central conceptual claim is that news intent is a stable deception signal even when surface writing patterns shift, and that InSide exploits this stability through joint intent-semantic learning.","tokens_in":19685,"tokens_out":5237,"duration_ms":55869,"significance":"If the empirical results hold, InSide is a substantive advance for content-based fake news detection: it demonstrates that LLM-generated intent signals can be integrated with semantic graph structures and yields consistent gains across four datasets, with ablations showing each component contributes. The paper's strengths include evaluation under chronological splits (except LLMFake), comparison against strong intent-aware baselines, an extensibility study across semantic encoders and intent generators, and a case visualization of learned alignment pathways. The main weakness is that the 'intent' construct is not independently validated; the reported gains may stem from additional model capacity or LLM priors rather than from the hypothesized stable-intent mechanism. This gap is central to the paper's motivation and limits the interpretability of the results.","major_comments":[{"comment":"The load-bearing conceptual claim is that LLM-extracted coarse intents capture stable intent rather than surface paraphrase. However, no human validation of the generated intents is provided, and the framework-swap test InSide(C') only changes the taxonomy used in the prompt; it does not test whether the coarse nodes encode content distinct from a simple re-encoding of the article. To support the central mechanism, please add (i) human evaluation of a sample of generated intents, or (ii) a control ablation that replaces coarse intent nodes with surface-level LLM summaries or paraphrases of the news text. Without such a control, the performance gains could be attributed to extra capacity or to implicit veracity priors in the frozen LLM.","section":"Section 3.3 and Section 4.3.3"},{"comment":"The fine-grained intent node update (hf_ij <- hf_ij + softmax((hc_i + hf_ij) · Hsem_sen^T) · Hsem_sen) directly mixes coarse intent embeddings with sentence embeddings. Thus the 'intent' graph is not independent of semantics, and the w/o Intf ablation removes both the fine nodes and a substantial amount of semantic-conditioned parameters. The improvement might reflect increased capacity rather than intent-specific information. Please include an ablation that replaces fine-grained intent nodes with an equivalent-capacity semantic-only component, or otherwise controls for parameter count, so that the intent contribution is isolated.","section":"Section 3.3, fine-grained intent initialization"},{"comment":"The asterisk is defined as a paired t-test at 0.05 significance comparing InSide with the best baseline, but only three runs are reported. Please clarify whether the paired test is computed over test instances (which ignores run-level variability) or over the three runs (which would be a very small sample). Additionally, because five metrics across four datasets are tested, multiple-comparison correction or at least a conservative interpretation should be reported. This is important because the paper's central claim—'consistently outperforms all competitive methods'—relies entirely on this significance evidence.","section":"Section 4.2, Table 2"}],"minor_comments":[{"comment":"Typo: 'denoted as Lcoarse and Lcoarse' should be 'Lcoarse and Lfine'. Also, the sentence 'Lfine encodes logical dependencies among high-level intent aspects' seems to describe coarse-coarse edges; please clarify the edge semantics.","section":"Section 3.3"},{"comment":"Typo: 'the feature extractor of MEInt' should be 'DMInt'.","section":"Section 4.2"},{"comment":"Minor typos: 'PoltiFact' in Table 3 header and 'Weibo21' in Figure 3 axis labels; use consistent dataset names.","section":"Table 3 and Figure 3"},{"comment":"No code or data availability statement is provided. Given the reliance on LLM-generated intents and specific hyperparameters, releasing code and the exact prompts would materially help reproducibility.","section":"Reproducibility"},{"comment":"The prompt text says the underlined part is only for LLM-Int, but the underline is not visible in the manuscript. Please mark it explicitly (e.g., with brackets) so readers can distinguish the two variants.","section":"Prompt 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competent applied CIKM paper with a useful graph-alignment mechanism and strong benchmark results. However, the conceptual novelty rests on the validity of LLM-extracted 'intents' as a stable, semantics-independent signal, and the current evidence does not establish that. The requested controls (human validation, paraphrase replacement, and capacity-matched ablations) are feasible within the manuscript's scope and should be required before acceptance. The significance-testing ambiguity also needs clarification. No concerns about citation norms or scope fit; the paper fits CIKM well."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: InSide is a real architectural contribution and the empirical package is mostly clean. The headline claim—jointly modeling semantics and LLM-extracted intent as heterogeneous graphs with pseudo-node alignment beats strong baselines—is internally consistent: gains on all four datasets, paired t-tests, and ablations where every removed component hurts. The framework-swap robustness test is a nice touch. Credit where due: this is not a toy experiment.\n\nThe biggest soft spot is the intent construct. The paper never validates that the LLM's coarse intent outputs are anything more than rephrased surface text. The swap test in 4.3.3 only changes the taxonomy, not the underlying fidelity problem. An ablation that replaces coarse intents with a paraphrase of the article (or with random text) would tell you whether the gains come from intent, from extra model capacity, or from LLM priors. Without that, the 'stable intent' mechanism is a hypothesis, not a demonstrated result. The benchmark numbers can still be right; the explanation is unproven.\n\nReproducibility is the second issue. No code or data, several hyperparameters left implicit, and the tuning protocol isn't described. The LLM baselines are nearly degenerate on PolitiFact (macF1 around 0.41), so the claim of beating LLM baselines is weak there. LLMFake's random split also muddies the temporal-generalization narrative, though the paper does disclose it.\n\nNone of this is fatal. The paper is honest about its limitations and the empirical claims are specific and falsifiable. If the code and prompts ship, the table of results will be checkable. For a reader in fake news detection, this is worth a serious look; the architecture is novel and the benchmark results are sufficiently strong that ignoring them would be a mistake.\n\nMy recommendation: send it to review, but require the missing intent-fidelity analysis and the code/data release before acceptance. The central mechanism claim should be reworded as a hypothesis supported by ablations, not a validated finding.","headline":"Solid graph-based detector with consistent benchmark gains; the 'stable intent' story needs validation before it carries the weight the paper puts on it.","tokens_in":20308,"tokens_out":2357,"would_cite":true,"duration_ms":26553,"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":"InSide proposes that modeling the intent behind news, together with its semantic graph, detects fake news more reliably than semantic-only or fixed-intent systems, and reports consistent gains on four benchmarks.","keywords":["fake news detection","news intent modeling","intent-semantic joint learning","heterogeneous graph neural networks","graph alignment","large language model intent extraction","social media misinformation","temporal robustness"],"falsifier":"Run InSide against an ablated control in which each LLM-generated intent string is replaced by an equally long string formed from random sentences of the same news article, keeping every other component identical. If macro-F1 stays near 0.86 on GossipCop, intent content is not the operative signal. A complementary check: give human annotators the extracted intent strings without the news text and ask them to label fake versus real; chance-level performance would mean intent alone does not separate the classes.","tokens_in":19221,"feed_emoji":"📰","tokens_out":5884,"duration_ms":62792,"temperature":0.7,"pith_summary":"The paper tries to show that a news item's intent—the purpose or motivation behind writing it—is a durable clue for fake news detection, even when the surface wording shifts. It claims that intent can be extracted from content with a frozen large language model and combined with sentence- and entity-level semantics through graph-based joint learning, and that this combination is more robust than semantic-only or fixed-category intent models. On PolitiFact, GossipCop, Weibo, and LLMFake, the proposed InSide model reports consistently higher macro-F1 and fake-class F1 than the compared baselines, with the largest fake-class improvement about 9.12% on GossipCop. The reason this would matter: a content-only detector that tracks stable intent could catch newly published fake news before comments, propagation structure, or fact checks become available.","feed_headline":"Graph model reading news intent tops four fake-news benchmarks","feed_subtitle":"InSide joins LLM-extracted intent with sentence-entity graphs, lifting fake-class F1 by up to 9.12%.","key_machinery":"The carrier of the argument is a pair of heterogeneous graphs—one for semantics (sentences plus entities with sliding-window and entity edges) and one for intent (LLM-derived coarse nodes plus learnable fine-grained nodes)—joined by dynamic pathway-based graph alignment through pseudo nodes. The pseudo nodes create a common space with edges to both graphs, use attention over typed edge features to decide which pathways matter, and are mean-pooled into the final classifier, so they act as the information bottleneck that fuses 'thoughts' and 'words'.","core_discovery":"At its core, the paper claims that writing style is flexible but intent is comparatively stable: two deceptive articles can sound very different yet share the same underlying intent, so intent offers a reliable target for detection. To use that idea, InSide converts one article into two heterogeneous graphs. The semantic graph has sentence nodes connected by a sliding window plus entity nodes that create long-range links; the intent graph has coarse nodes representing belief, plan, desire, and outcome, produced by prompting a frozen large language model, plus learnable fine-grained nodes that tie each coarse intent to the narrative. Both graphs are updated by local message passing and a supe","pith_inferences":["A direct experiment the paper does not run: adversarially rewrite fake news to change style while preserving intent; if InSide degrades less than semantic-only baselines, the 'intent is stable' mechanism is confirmed rather than just the architecture.","The framework could transfer to detecting manipulative framing of true stories, which the paper notes as a possibility, suggesting a broader early-warning use case for coordinated influence campaigns that reuse the same intent across languages.","Because coarse intent nodes come from a frozen LLM, quantizing or distilling the generator could lower deployment cost; whether a fine-tuned intent extractor would outperform the frozen one remains an open question."],"forward_implications":["Content-only detection can be made more practical for brand-new articles, because InSide needs no user comments, propagation trees, or external evidence at inference time.","Intent-aware modeling can be decoupled from any specific taxonomy: InSide with a nine-perspective intent framework stays competitive, suggesting the coarse-to-fine machinery, not the particular prompt set, carries much of the gain.","The joint framework scales with the backbone: replacing BERT with BERT-large or the intent generator with a 32B-parameter model improves results, while lighter modules still help.","Fake-class F1 is the practically important metric for moderation, and relative gains up to 9.12% on GossipCop mean fewer fake items slip through at the same operating point."],"supporting_citations":[{"why":"Supplies the survey definition of news intent that motivates the stable-intent premise.","marker":"Zhou and Zafarani (2020)"},{"why":"Provides the theory-driven intent framework (belief, plan, desire, outcome) used for coarse-grained intent nodes.","marker":"Wang et al. (2025)"},{"why":"The classification-based intent baseline DM-Inter that InSide is designed to surpass.","marker":"Wang et al. (2024a)"},{"why":"Provides the PolitiFact dataset used in the main comparison.","marker":"Popat et al. (2018)"},{"why":"Provides the FakeNewsNet repository that supplies the GossipCop dataset.","marker":"Shu et al. (2020a)"},{"why":"Provides the Weibo benchmark dataset.","marker":"Sheng et al. (2022)"},{"why":"Provides the LLMFake dataset of LLM-generated misinformation.","marker":"Chen and Shu (2023)"},{"why":"Supplies the chronological train/validation/test split and the entity-debiasing baseline ENDEF.","marker":"Zhu et al. (2022)"},{"why":"GenFEND, the strongest semantic baseline across several metrics, which InSide must beat.","marker":"Nan et al. (2024)"}],"fun_headline_variants":["Intent beats style: Graph model detects fake news via deep thoughts","Fake news unmasked by intent, not just words","InSide: Graph AI sees the intent behind fake news","Stable intent signals help AI catch evolving fake news","Graph model uses news intent to top four fake-news tests"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the texts a frozen large language model produces when prompted for belief, plan, desire, and outcome really are a faithful, non-redundant record of the news's intent; if those texts are just paraphrases of the article's surface meaning, the reported gains could come from extra model capacity or LLM priors rather than from intent itself.","fun_headline_variants_meta":{"raw":{"variants":["Intent beats style: Graph model detects fake news via deep thoughts","Fake news unmasked by intent, not just words","InSide: Graph AI sees the intent behind fake news","Stable intent signals help AI catch evolving fake news","Graph model uses news intent to top four fake-news tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1264,"prompt_tokens":766,"completion_tokens":498,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":510,"completion_tokens_details":{"reasoning_tokens":416}},"tokens_in":510,"tokens_out":498,"duration_ms":5859,"temperature":1.0,"reasoning_tokens":416,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:19:17.099534+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run InSide against an ablated control in which each LLM-generated intent string is replaced by an equally long string formed from random sentences of the same news article, keeping every other component identical. If macro-F1 stays near 0.86 on GossipCop, intent content is not the operative signal. A complementary check: give human annotators the extracted intent strings without the news text and ask them to label fake versus real; chance-level performance would mean intent alone does not separate the classes.","supporting_citations":[],"review_version":1}