{"id":"5e070628-2f11-4439-962b-5d81bd0428c4","arxiv_id":"2507.13622","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"IP2 adds entity-level interest probing at both headline and cross-article levels and reports state-of-the-art recommendation accuracy on MIND-small, MIND-large, and Adressa-1week.","lead":"This paper presents a news recommendation model that uses the people, places, and organizations named in headlines to model what catches a reader's attention and what they read next. The model learns these name-based signals directly from news text, without needing an external knowledge graph, and reports the best published accuracy scores on three public news datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The [ent] signature entity is trained only to match the full title (Eq. 12), so the claim that it isolates a leading entity and thereby probes intra-news scanning interest is unsupported; the 1.0 AUC ablation gain may come from a generic second representation.","rationale":"I read the paper in good faith and accept that its empirical results are strong: public code is provided, ablations show each component contributes, and the gains over baselines are consistent across datasets. The reader's verdict of CONDITIONAL is appropriate. My stress-test identifies the same weakest assumption: the paper's central behavioral and architectural claim — that the [ent] signature entity captures a leading intra-news entity and that this drives the performance — is not directly supported by any loss term or evaluation. The contrastive objective aligns e_i with the entire title embedding, so the signature entity could simply be a second, title-derived representation that helps through multi-view learning rather than through entity-level interest probing. This does not invalidate the empirical SOTA claim, but it weakens the paper's stated novelty and its alignment with the scanning-title-reading-clicking behavior model. The concrete test above would settle whether the entity-specific interpretation holds. Since the reader already conditioned acceptance on testing the behavioral mechanism and reporting additional experimental details, my analysis does not move the verdict; it reinforces the existing condition.","tokens_in":15873,"tokens_out":3685,"duration_ms":44068,"concrete_test":"Use the released code to load the pretrained entity memory M and SEE. Sample 500 MIND-small titles with at least three entity annotations each. For each title, (a) record the softmax attention weights from the [ent] token to all entity positions in the last SEE layer, and (b) compute cosine similarities between the signature entity e_i and each constituent entity embedding E_{e_j}, as well as between e_i and the title embedding h_i. If the average maximum attention weight is below 0.5, or if e_i is consistently more similar to h_i than to any single entity embedding, the 'leading entity probing' interpretation fails. As a follow-up, replace the SEE output with the mean of entity embeddings (retaining the same contrastive loss) and rerun the w/o Intra ablation; if the 1.0 AUC gap persists, the gain is not from entity-specific probing but from having a second title-aligned view.","verdict_should_be":"UNCHANGED","load_bearing_attack":"IP2's central contribution is entity-guided interest probing at two levels, but the intra-news mechanism is not actually enforced. In Section 4.1.2, the Signature Entity Encoder (SEE) prepends a learnable [ent] token to the entity list and uses its output as the 'signature entity' e_i. The only supervision for e_i comes from the contrastive loss in Eq. (12), where the positive pair is (e_i, h_i) and h_i is the full-title embedding from BERT. Nothing in the loss requires e_i to be close to any particular entity embedding, to have concentrated attention on a single leading entity, or to be distinct from a generic title summary. The 'probing' language in Section 4.2 and the attention-weight interpretation in the Section 5.6 case study are post-hoc; attention weights in the SEE are not supervised and could be diffuse. The ablation w/o Intra (Table 4) shows that removing the entity-title contrastive loss costs about 1.0 AUC, but this is equally consistent with e_i acting as a complementary second view of the whole title, initialized with useful semantics through the contrastive objective. The behavioral claim that IP2 'probes the leading entity during scanning' is therefore an untested interpretation rather than an established result, and it is load-bearing because the paper frames its novelty and its alignment with behavioral science around this mechanism. A secondary but concrete reproducibility issue: Section 4.2 states alpha, beta, and delta sum to 1, while Section 5.1.2 sets them to 0.3, 0.2, and 0.2, summing to 0.7.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IP2, a news recommendation model that aims to model entity-guided reading interest at two levels: an intra-news level, where a Signature Entity Encoder (SEE) aggregates entities in a title into a single 'signature entity' representation trained by contrastive pre-training against the title embedding, and an inter-news level, where a dual-tower user encoder uses cross attention between title and entity streams to combine semantic and entity-guided user interest. The model is evaluated on MIND-small, MIND-large, and Adressa-1week, with extensive baselines, ablations, and a case study, and the authors report state-of-the-art results. The paper also releases source code. The central SOTA claim is supported by the reported experiments, but the paper's behavioral interpretation of the intra-news mechanism is not directly evidenced, and there are reproducibility and statistical reporting issues.","tokens_in":16238,"tokens_out":3940,"duration_ms":45575,"significance":"If the empirical results are correct, IP2 is a practically useful contribution: it outperforms a broad set of neural and knowledge-aware baselines on three dataset variants, and it does so without requiring an external knowledge graph. The two-stage contrastive pre-training recipe and the cross-attention dual-tower user encoder are concrete, reproducible architectural ideas. The paper also provides a public code repository, which strengthens the reproducibility of the work. The main significance risk is that the paper frames its novelty around 'probing' the leading entity during scanning, but the loss function does not actually enforce that the signature entity corresponds to a leading entity or even to any single entity. The ablation results show that the entity-title contrastive loss is useful, but they do not establish the proposed cognitive interpretation. This needs to be addressed before the claimed mechanism can be accepted.","major_comments":[{"comment":"The signature entity e_i is supervised only by the contrastive objective in Eq. (12), where the positive pair is (e_i, h_i) and h_i is the pooled embedding of the entire title. Nothing in this loss forces the [ent] output to concentrate on one leading entity, to match a particular entity embedding, or to be distinct from a generic title summary. Consequently, the claims in Section 4.2 and Figure 3 that SEE 'probes' intra-news entity interest and that the case-study attention weights in Section 5.6 reveal the leading entity are not established by the presented evidence. The 1.0 AUC drop of w/o Intra in Table 4 is equally consistent with e_i acting as a second, complementary view of the whole title. Please provide a direct test of the probing mechanism, for example by comparing SEE attention or e_i against human- or entity-level salience annotations, or by ablating with the [ent] output replaced by a mean-pooled entity representation.","section":"Section 4.1.2 and Eq. (12)"},{"comment":"Section 4.2 states that the contrastive weights alpha, beta, and delta sum to 1, but Section 5.1.2 reports alpha=0.3, beta=0.2, delta=0.2, which sum to 0.7. The exact pre-training objective is therefore ambiguous and the reported setting cannot be reproduced as written. Please correct the inconsistency and state the exact loss weights used in the experiments.","section":"Section 4.2 and Section 5.1.2"},{"comment":"The headline improvements over the best baseline are modest, approximately 0.8 to 1.0 AUC on MIND-small and 1.0 to 1.4 AUC on MIND-large, yet only five-run averages are reported, without standard deviations, confidence intervals, or error bars. The significance statement that improvements are significant at the 0.05 level with a paired t-test does not say which baseline or baselines were tested. Please report per-seed variance and state the exact comparison protocol. In addition, Table 3 compares only four baselines on Adressa, so the claim that IP2 'outperforms all compared methods in all cases' should be qualified by noting that knowledge-graph-aware baselines could not be run on that dataset.","section":"Section 5.2, Tables 2 and 3"}],"minor_comments":[{"comment":"The caption and axis labels of Figure 4 are difficult to parse: 'Random' and 'TransE' appear on the x-axis, but the curves are labeled only by 'w/ CP' and 'w/o CP', and no numerical values or error bars are shown. Please add a clear legend and report numeric results with variance.","section":"Section 5.4 and Figure 4"},{"comment":"The case study is qualitative and the attention visualization is post-hoc; the text should explicitly state that the highlighted attention weights are not supervised and should not be read as direct evidence for the probing mechanism.","section":"Section 5.6 and Figure 7"},{"comment":"The text contains a typo, 'TranE', which should be 'TransE'.","section":"Section 5.4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-executed news recommender paper, not a breakthrough. The main new thing is replacing knowledge-graph entity embeddings with a contrastive pre-training step that aligns a learned entity embedding with the BERT title embedding, then feeding both title and entity signals into a dual-tower user encoder with cross attention. The method is coherent, the ablations are honest, and the code is public. The numbers are strong: roughly +1 AUC on MIND-small, +1.4 on MIND-large, and +4.5 on Adressa over the best baselines. On Adressa, the fact that it works without a KG is a real practical point, since Norwegian news KGs are thin.\n\nThe soft spot is the behavioral story. The abstract and intro claim the [ent] token in the Signature Entity Encoder 'probes the leading entity during scanning' and that cross attention models 'title reading.' But the [ent] output is trained only to match the whole title embedding via the contrastive loss in Eq. (12). Nothing forces it to isolate a dominant entity. So the entity-guided scanning claim is not actually enforced; the ~1.0 AUC ablation gain from removing the intra-news contrastive loss is just as consistent with [ent] being a useful second view of the whole title. The case study attention weights look plausible, but they are not supervised. I'd treat 'behavioral mechanism' as an interpretive gloss, not a finding.\n\nOn the empirical side, the main caveats: no standard deviations anywhere, only five-seed averages and a t-test statement. That makes it hard to judge whether the +0.8 to +1.4 AUC gaps are meaningful, though they are consistent across four metrics and three datasets. The Adressa baseline set is thin (four methods), and IP2 uses a Norwegian BERT, which likely helps compared to general BERT baselines. Also a concrete reproducibility error: Section 4.2 says alpha, beta, delta sum to 1, but Section 5.1.2 sets them to 0.3, 0.2, 0.2.\n\nThe paper deserves a serious referee. It is well-written, the MIND experiments are extensive, the baseline list is broad, and the KG-free angle is worth engaging with. I would not desk-reject. My own verdict would be conditional: request variance reporting, fix the alpha/beta/delta inconsistency, and add a direct test of whether [ent] actually captures entity salience, say, a probing task against gold entity labels. If those are addressed, the SOTA claim would be close to established.","headline":"Solid news recommendation paper with a genuinely useful KG-free entity pretraining trick, but the behavioral 'probing' interpretation is not actually enforced by the loss.","tokens_in":16798,"tokens_out":2410,"would_cite":true,"duration_ms":27810,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"IP2 claims that modeling entity-guided interest at two levels—which entity dominates headline scanning and how entity preference across reading history shapes title reading—makes a news recommender outperform all compared state-of-the-art…","keywords":["news recommendation","entity-guided interest","contrastive pre-training","dual-tower user encoder","cross attention","signature entity","knowledge-graph-free","click prediction"],"falsifier":"An ablation that keeps the contrastive pretraining and dual-tower user encoder but replaces the signature entity encoder's [ent] output with the average of the encoded entity vectors would settle the mechanism: if the score drop is far below the roughly 1.0 AUC drop of the full w/o Intra ablation, the leading-entity probing is not the source of the reported gains.","tokens_in":15644,"feed_emoji":"📰","tokens_out":8120,"duration_ms":83239,"temperature":0.7,"pith_summary":"News recommendation usually treats named entities as extra semantic clues. This paper argues that entities play a distinct behavioral role: during the brief scanning of a headline one entity catches the eye, and during the later title-reading and click stage a reader's accumulated entity preferences guide whether the title is worth reading. The proposed IP2 probes the scanning-stage signal by compressing all entities in a headline into one signature entity vector and aligning it with the title through contrastive pretraining; it probes the second signal with a dual-tower user encoder where entity interest and title-meaning interest attend to each other before the click score is computed. On MIND-small, MIND-large, and Adressa-1week, IP2 reports the best AUC, MRR, nDCG@5, and nDCG@10 among all methods compared, including gains of about 0.8–1.0 AUC on MIND-small, 1.0–1.4 on MIND-large, and about 4.5 on Adressa. If the mechanism is real, entity guidance is not a side channel but a primary driver of reading and clicking behavior.","feed_headline":"IP2 beats every compared news recommender by modeling entity interest","feed_subtitle":"A two-level entity model tracks scanning and title-reading signals, reporting top AUC, MRR, and nDCG on MIND and Adressa.","key_machinery":"The machinery that carries the argument is the signature entity encoder plus the dual-tower cross-attention user encoder, trained in two stages. The signature entity encoder is a small transformer over entity embeddings with a prepended [ent] token; during contrastive pretraining, the [ent] output is matched against the title embedding so the entity memory takes on meaning from the news text rather than from an external knowledge graph. The user encoder runs two parallel attention towers over the reading history, one querying titles against entities and the other querying entities against titles, and a learnable gate mixes the two user preference vectors; the same gating is applied to the candidate news embedding. The cross-attention link is the component that lets inter-news entity interest adjust title-reading interest, and the learnable aggregation is what the paper uses to adapt the strength of entity guidance per user and per news item.","core_discovery":"The paper's central claim is that a news recommender should model entity interest at exactly the two places the reading behavior chain puts it. At the intra-news level, a signature entity encoder prepends a learnable [ent] token to the entities mentioned in a title, stacks transformer layers over them, and uses the [ent] output as the title's signature entity; a contrastive pretraining loss pulls this signature entity toward the title embedding and pushes it away from other titles, which the paper interprets as teaching the entity memory which entity is the leading attractor during scanning. At the inter-news level, a dual-tower user encoder derives one preference vector from title meanings and one from entities, and cross-attention lets each tower query the other so that entity interest can calibrate title reading interest before the final click prediction. The paper reports that IP2 outperforms all compared neural and knowledge-aware baselines on every metric on MIND-small, MIND-large, and Adressa-1week, and that ablations removing the intra-news contrastive pretraining or the inter-news cross-attention each cost roughly 1.0 AUC on MIND-small, with the inter-news removal costing the most.","pith_inferences":["One testable extension is to quantify how much of the gain comes from bridging unlabeled phrases: the case study suggests the cross-attention link connects a phrase like 'Black Friday' to a labeled entity even when the phrase itself is not annotated, so performance on headlines with very few annotated entities could be measured separately.","The behavioral story predicts measurable differences at the scanning stage itself, so an eye-tracking or dwell-time study on headlines with one dominant entity would test whether the [ent] representation tracks actual first-pass attention rather than just click patterns.","An ablation that replaces the signature entity encoder with a second copy of the title encoder, keeping the same contrastive and cross-attention setup, would separate the value of a dedicated entity stream from the value of having any second aligned stream at all."],"forward_implications":["If the two-level entity signal is the cause of the gains, news recommenders can build entity representations from unlabeled title corpora and click logs, avoiding knowledge graphs that age quickly and miss new entities.","The same scan-title-click decomposition gives a design template for other short-text recommendation domains, where a salient token often directs first-pass attention.","The entity tower and its gate weight provide a readable account of why an item was ranked: the gate says whether entity interest or title meaning dominated for that user.","Because the contrastive pretraining stage requires no interaction logs, the method can be warmed up on large news collections before seeing any user data, which the paper says helps on smaller datasets such as MIND-small."],"supporting_citations":[{"why":"Behavioral-science evidence that screen-based reading is scanning-oriented; grounds the three-step scan-title-click behavior chain.","marker":"[19]"},{"why":"Entity-aware contrastive learning of sentence embeddings that the signature entity-title pretraining adapts.","marker":"[23]"},{"why":"The temperature-scaled contrastive loss used in the pretraining objective.","marker":"[4]"},{"why":"Multi-head self-attention news and user encoder that supplies the base architecture for the dual towers and is a primary baseline.","marker":"[34]"},{"why":"Entity-based persona and contrastive learning baseline whose reported results are used for comparison.","marker":"[18]"},{"why":"Global and local entity-graph-enhanced news and user representation model used as a main knowledge-aware baseline.","marker":"[39]"},{"why":"MIND dataset and the AUC, MRR, nDCG@5, nDCG@10 evaluation protocol.","marker":"[37]"},{"why":"Adressa-1week dataset used as the second benchmark.","marker":"[8]"}],"fun_headline_variants":["IP2 models entity interest at scanning and title-reading to beat news baselines","Entity interest at two reading stages lifts IP2 to state-of-the-art news rec","Two-level entity interest probing: IP2 tops news recommendation benchmarks","IP2: entity interest at two reading steps drives news recommendation gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the entity encoder's summary vector really captures the single most attention-grabbing entity in a headline, rather than simply reproducing whatever the title means as a whole.","fun_headline_variants_meta":{"raw":{"variants":["IP2 models entity interest at scanning and title-reading to beat news baselines","Entity interest at two reading stages lifts IP2 to state-of-the-art news rec","Two-level entity interest probing: IP2 tops news recommendation benchmarks","IP2: entity interest at two reading steps drives news recommendation gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000737,"raw_usage":{"total_tokens":3344,"prompt_tokens":1046,"completion_tokens":2298,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":662,"completion_tokens_details":{"reasoning_tokens":2219}},"tokens_in":662,"tokens_out":2298,"duration_ms":17985,"temperature":1.0,"reasoning_tokens":2219,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:20:34.868834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An ablation that keeps the contrastive pretraining and dual-tower user encoder but replaces the signature entity encoder's [ent] output with the average of the encoded entity vectors would settle the mechanism: if the score drop is far below the roughly 1.0 AUC drop of the full w/o Intra ablation, the leading-entity probing is not the source of the reported gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Behavioral-science evidence that screen-based reading is scanning-oriented; grounds the three-step scan-title-click behavior chain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Entity-aware contrastive learning of sentence embeddings that the signature entity-title pretraining adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The temperature-scaled contrastive loss used in the pretraining objective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Multi-head self-attention news and user encoder that supplies the base architecture for the dual towers and is a primary baseline."},{"cited_title":"PerCoNet: News Recommendation with Explicit Persona and Contrastive Learning","cited_arxiv_id":"2304.07923","evidence_quote":"Entity-based persona and contrastive learning baseline whose reported results are used for comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MIND dataset and the AUC, MRR, nDCG@5, nDCG@10 evaluation protocol."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Adressa-1week dataset used as the second benchmark."}],"review_version":1}