Co-Readership Identity Control

• Co-readership identity control is a framework that isolates individual reading behavior by comparing readers on the same document, holding document-level factors fixed.
• It distinguishes between generic, crowd, and personal salience by measuring own-versus-other prediction gaps to uncover unique reader selection signals.
• Empirical findings suggest that while shared document salience remains consistent, individual selection exhibits a durable, trait-like signature over time.

Searching arXiv for papers on co-readership identity control and closely related co-readership methods. Co-readership identity control is a methodological framework for isolating person-specific signal in reading traces by comparing readers on the same document or within the same co-reading neighborhood, thereby holding document-level or topic-level confounds fixed while varying only the reader identity used for prediction. In the recent social-highlighting literature, the term denotes a clean personalization control: a target reader’s history is tested against another co-reader’s history on the same candidate set, and the resulting own-versus-other gap is interpreted as the recoverable individual component beyond shared salience (Nakayashiki et al., 8 Jun 2026). In older Mendeley-based network studies, the same broad idea appears in a different form: readership data are organized by self-assigned reader identities such as professional status, discipline, and country, allowing analysts to characterize or control for how identity categories structure shared reading (Haunschild et al., 2015). Taken together, these lines of work define co-readership identity control as a way to distinguish common reading structure from reader-specific behavior, either at the level of aggregate readership networks or at the level of individual prediction (Nakayashiki et al., 9 Jun 2026).

1. Conceptual definition and scope

In its most explicit formulation, co-readership identity control asks a narrow question: Does a person’s own history predict their highlights better than another co-reader’s history does, when both are judging the same document? The crucial feature is the control itself: the document is fixed, and in some analyses the broad topic is effectively fixed as well, so any performance difference between an “own” profile and an “other” profile is attributable to reader identity rather than document difficulty or topic mixture (Nakayashiki et al., 8 Jun 2026).

This framework separates three quantities that are often conflated in personalization work: generic salience, crowd salience, and personal salience. Generic salience is a content-only or document-structure proxy for what should stand out; crowd salience is what other readers actually marked in the same document; personal salience is the residual associated with a specific individual beyond those shared layers (Nakayashiki et al., 8 Jun 2026). A parallel distinction is drawn between salience and selection. Salience asks what is highlightable in a document, whereas selection asks which of the already-salient passages belong to a given reader. The main empirical consequence is that salience is mostly shared, while selection is where individuality is detectable (Nakayashiki et al., 8 Jun 2026, Nakayashiki et al., 9 Jun 2026).

In the Mendeley network literature, the notion is broader and more classificatory. There, reader identity is operationalized through self-assigned metadata such as disciplinary affiliation, professional status, and country/location, and co-readership networks are used to see whether reading differences are attributable to those identity categories or whether most groups read a common core literature (Haunschild et al., 2015). This suggests that co-readership identity control can refer either to a predictive contrast between individuals or to an analytic control over identity categories in aggregate readership structure.

2. Reader identity as metadata and network structure

A foundational use of readership-based identity control appears in the analysis of Mendeley reader statistics by constructing three networks from reader profiles: a discipline network, a status-group network, and a country network (Haunschild et al., 2015). The study used data retrieved from the Mendeley API between 11 and 23 December 2014 and examined a publication set of 1,133,224 articles and 64,960 reviews from Web of Science published in 2012, later reporting 1,137,178 papers matched via DOI (Haunschild et al., 2015).

The status-group network is particularly important for co-readership identity control because it organizes shared reading through readers’ self-assigned professional identities. In that network, the paper reports 13 status groups, average degree 13.00, density 1.00, closure 1.00, average distance 1.00, diameter 1, compactness 1.00, and modularity 0.00, meaning that the network is fully connected and exhibits no modular split among status groups (Haunschild et al., 2015). The result is not merely descriptive. It supports the interpretation that professional status does not strongly partition readership into separate communities; rather, the status network is a single dense core of interconnected reader identities (Haunschild et al., 2015).

The most central status groups are PhD students, Master’s students, and postdocs, and the paper interprets high eigenvector centrality as indicating that these groups share papers with many other groups and are connected to other well-connected groups (Haunschild et al., 2015). However, the same study explicitly notes that the centrality ranking is strongly influenced by group size, with rank correlation between eigenvector centrality and reader counts of $\rho = 0.986$ (Haunschild et al., 2015). That caveat is central to identity control: observed “importance” of an identity category may partly reflect category volume rather than a uniquely distinctive reading pattern.

The study also emphasizes that identity categories are not neutral measurement devices. Mendeley’s classification scheme contains redundant categories such as “Student PhD,” “Student Post-Graduate,” and “Doctoral Student,” includes some system-specific or US-centric labels such as “Assistant Professor,” “Lecturer,” and “Senior Lecturer,” and omits labels such as “Reader” and “Habilitand” (Haunschild et al., 2015). Accordingly, co-readership identity control at the metadata level is partly a control over the platform’s identity taxonomy itself.

3. Formalization of the modern identity-control criterion

The recent social-highlighting literature formalizes co-readership identity control as an own-versus-other comparison under matched candidate pools. In its basic form, the contrast is

$$\Delta_{\text{own-other}} = \mathrm{AP}(u_{\text{own}}) - \mathrm{AP}(u_{\text{other}})$$

where AP is average precision on the same target document and task, and the only difference is whether the scoring profile comes from the target user or from another co-reader (Nakayashiki et al., 8 Jun 2026). A positive gap indicates that the target’s own history contains recoverable identity signal beyond another reader’s history.

For within-document salience, the candidate set is all sentences, and positives are the sentences the target user highlighted (Nakayashiki et al., 8 Jun 2026). For selection, the candidate set is the union of spans highlighted by co-readers, and positives are the spans highlighted by the target user (Nakayashiki et al., 8 Jun 2026). This difference in candidate construction is not procedural detail alone; it encodes the conceptual distinction between identifying what is salient in a document and identifying which salient items belong to a specific person.

The scoring setup in "Personal Salience: Highlighting Is Social, but Individuality Lives in Selection" (Nakayashiki et al., 8 Jun 2026) is based on three functions:

• Generic score:

$$s_{\text{generic}}(x) = \cos(\mathrm{emb}(x), \mathrm{centroid}(D))$$

• Crowd score:

$$s_{\text{crowd}}(x) = \#\{\text{other co-readers who marked } x\}$$

• Personal score:

$$s_{\text{personal}}(x; u) = \cos(\mathrm{emb}(x), \mathrm{profile}(u))$$

The paper emphasizes that the crowd estimate is computed from a dense crowd, with median crowd size about 38, rather than the tiny $\le 5$ reader crowds used in earlier pilot-style settings (Nakayashiki et al., 8 Jun 2026). This is a methodological commitment: weak crowd baselines can exaggerate the appearance of personal signal.

A related document-level formalization appears in "Selection, Not Salience: The Shape and Limits of Personalization in Social Highlighting" (Nakayashiki et al., 9 Jun 2026). There, each candidate document is represented by a title embedding, each user profile by the mean embedding of highlighted spans, and the scores are

$$s_{\text{own}}(d,A) = \cos\big(\phi(d), \psi(A)\big)$$

and

$$s_{\text{other}}(d,A,B) = \cos\big(\phi(d), \psi(B)\big)$$

with evaluation conducted on the same within-community candidate pool (Nakayashiki et al., 9 Jun 2026). The formal structure is therefore consistent across altitudes: own-versus-other comparison on matched items.

4. Empirical asymmetry between salience and selection

The central empirical finding of recent co-readership identity-control work is an asymmetry: the individual signal is weak in salience and much stronger in selection. On the within-document salience task, "Personal Salience" reports crowd: 0.321 AP, generic centrality: 0.187, personal embedding profile (own): 0.172, personal embedding profile (other): 0.155, frontier twin (own): 0.247, and frontier twin (other): 0.235 (Nakayashiki et al., 8 Jun 2026). The own-versus-other gap for salience is therefore +0.017 AP for the embedding scorer, with 95% CI [0.008, 0.027], while the frontier twin shows +0.012 AP with 95% CI [-0.004, 0.029], which is not significant (Nakayashiki et al., 8 Jun 2026).

By contrast, on the selectivity task, where the candidate set is already restricted to salient spans, the same paper reports personal own profile: 0.397 AP, other reader: 0.254 AP, and an own-vs-other gap of +0.143 AP, with 95% CI [0.105, 0.183] (Nakayashiki et al., 8 Jun 2026). The paper summarizes this as the claim that the personal signal in salience is “at most a whisper,” whereas individuality “lives in selection” (Nakayashiki et al., 8 Jun 2026).

The follow-up paper "Selection, Not Salience" extends the same conclusion across reading altitudes (Nakayashiki et al., 9 Jun 2026). At the document altitude, the corrected own-versus-other gap is +0.169 with community negatives and +0.119 with topic-matched hard negatives, both highly significant (Nakayashiki et al., 9 Jun 2026). At the sentence altitude, a two-stage personalized auto-highlight system does not improve on its impersonal baseline: a personal re-ranker is beaten by the salience order even on the highest-recall candidate pool, and the own-versus-other gap remains small (Nakayashiki et al., 9 Jun 2026). The synthesis offered there is explicit: the selection signal is of comparable magnitude across altitudes, about +0.12 to +0.17, while salience remains largely shared (Nakayashiki et al., 9 Jun 2026).

A compact comparison is useful here.

Task layer	Candidate set	Reported identity signal
Within-document salience	All sentences	+0.017 AP embedding gap (Nakayashiki et al., 8 Jun 2026)
Span-level selection	Union of co-reader-highlighted spans	+0.143 AP (Nakayashiki et al., 8 Jun 2026)
Document selection	Within-community candidate pool	+0.169 community, +0.119 hard (Nakayashiki et al., 9 Jun 2026)

This suggests that co-readership identity control does not reveal a strong personal signature in what passages are intrinsically salient in context. Rather, it reveals a modest but reliable signature in which of the available salient items or candidate documents a reader selects.

5. Topic matching, hard negatives, and what the signal represents

A major concern in co-readership identity control is whether apparent individuality is merely topic preference in disguise. The recent literature addresses this directly by using topic-matched peers and hard negatives. In "Personal Salience," the selectivity gap shrinks from +0.113 against a prolific peer in the denser cohort to +0.098 against a topically matched peer (Nakayashiki et al., 8 Jun 2026). Under tertile decomposition by peer similarity, the gap collapses from +0.212 at similarity about 0.52 to +0.055 at about 0.76, and to +0.026 at about 0.90, which the paper describes as roughly an 8× reduction from weakly matched peers to near-topic twins (Nakayashiki et al., 8 Jun 2026).

The interpretation is deliberately conservative. Even the residual +0.026 AP, with 95% CI [0.016, 0.037], is not claimed to be definitively non-thematic style; it could reflect finer-grained topic or a genuine stylistic preference (Nakayashiki et al., 8 Jun 2026). This caution is essential to the concept of identity control: the method can isolate signal beyond broad topic matching, but it does not automatically identify the psychological source of that residual.

The same logic governs the document-selection experiments in "Selection, Not Salience" (Nakayashiki et al., 9 Jun 2026). There, the main condition uses community negatives, documents highlighted by other community members but not by the target user, while the stricter regime uses topic-matched hard negatives, defined as the co-reader documents most similar to the target’s profile centroid (Nakayashiki et al., 9 Jun 2026). The own-versus-other gap declines from +0.169 under community negatives to +0.119 under hard negatives, and the paper interprets roughly one-third of the community gap as attributable to coarse topic:

$0.169 - 0.119 \approx 0.05$

(Nakayashiki et al., 9 Jun 2026). A smaller content-based robustness arm using article-content centroids rather than titles still yields a positive own-versus-random contrast of +0.154 against topic-matched hard negatives, suggesting that the signal is not merely title-driven, though the own-versus-other gap is not significant at that small sample size (Nakayashiki et al., 9 Jun 2026).

The broader implication is that co-readership identity control measures a mixture in which stable thematic preference is a dominant component. A plausible implication is that “identity” in this literature should often be read behaviorally rather than psychologically: it denotes a reader-specific selection signature, much of which is topically structured.

6. Leakage, dense crowds, and evaluation pitfalls

The methodological force of co-readership identity control depends on strict leakage prevention and properly specified controls. "Personal Salience" states that naive history-conditioning can leak the target document into the personal profile in about 42% of within-document pairs and 31% of selectivity pairs, inflating apparent personalization by roughly +0.07 to +0.15 AP (Nakayashiki et al., 8 Jun 2026). The paper therefore constructs the personal profile per pair, explicitly excluding the exact target document (Nakayashiki et al., 8 Jun 2026). Without that exclusion, the control ceases to isolate reader identity.

The same paper also argues that small crowds overstate personalization. Because crowd salience is estimated from too few co-readers in pilot-style settings, the crowd baseline appears weaker than it actually is, making personal models look more competitive than they should (Nakayashiki et al., 8 Jun 2026). The use of a dense crowd, with median size about 38, is therefore part of the fairness condition of the identity-control comparison (Nakayashiki et al., 8 Jun 2026).

"Selection, Not Salience" identifies a second failure mode, the control-in-negatives bias (Nakayashiki et al., 9 Jun 2026). In an initial run, the document-level own-versus-other gap was reported as +0.227, but auditing showed that the control reader’s own documents were present in the target reader’s negative set. As a result, the other-profile scorer was penalized for ranking its own genuine positives highly, artificially depressing the “other” score and inflating the identity gap (Nakayashiki et al., 9 Jun 2026). After excluding the control member’s documents and the topic-near member’s own documents from the target’s negatives, the corrected gap became +0.169 for community negatives (Nakayashiki et al., 9 Jun 2026). The paper notes that the control scorer falling below random floor is a diagnostic symptom of this bias (Nakayashiki et al., 9 Jun 2026).

A further safeguard is the model-matched control. "Personal Salience" evaluates a proprietary gpt-5.5 twin given up to 50 of a user’s past highlights, and crucially applies the same twin to another reader’s history, yielding twin-own and twin-other (Nakayashiki et al., 8 Jun 2026). This ensures that any own-versus-other gap cannot be attributed to differences in extractor family. In effect, co-readership identity control requires not merely a personalized scorer, but a controlled counterfactual scorer built the same way from another reader.

7. Temporal durability and the status of reading identity

The temporal extension of co-readership identity control appears in "Trait, Not State: The Durability of Reading Identity in Social Highlighting" (Nakayashiki et al., 11 Jun 2026). That work asks whether the selection signature identified cross-sectionally is a transient state or a durable trait. It operationally defines trait as “a stable, person-specific selection signature under continued platform engagement” and freezes each reader’s first six months of highlighting as a profile, then evaluates its later own-versus-other advantage out to 24+ months (Nakayashiki et al., 11 Jun 2026).

The paper uses a fine-layer regime in which controls and negatives are drawn from the reader’s top-25 most profile-similar peers, along with a coarse global regime (Nakayashiki et al., 11 Jun 2026). The fine-layer anchor reproduces the earlier cross-sectional effect: +0.188 [0.160, 0.216], compared with the previous estimate of +0.169, thereby validating the harness (Nakayashiki et al., 11 Jun 2026). Across time bins, fine-layer own-versus-other advantage remains positive throughout: +0.188 at 0–1 months, +0.174 at 1–3 months, +0.176 at 3–6 months, +0.180 at 6–12 months, +0.146 at 12–24 months, and +0.136 at 24+ months, with fraction of bootstrap resamples positive = 1.00 in every bin (Nakayashiki et al., 11 Jun 2026).

The paired retention analysis is more stringent because it compares each user to themselves over time. In the primary fine layer, the paper reports 6–12 months: $d = +0.001$, 95% CI $$\Delta_{\text{own-other}} = \mathrm{AP}(u_{\text{own}}) - \mathrm{AP}(u_{\text{other}})$$0, $$\Delta_{\text{own-other}} = \mathrm{AP}(u_{\text{own}}) - \mathrm{AP}(u_{\text{other}})$$1 [0.854, 1.184], n = 212; 12–24 months: $$\Delta_{\text{own-other}} = \mathrm{AP}(u_{\text{own}}) - \mathrm{AP}(u_{\text{other}})$$2, 95% CI $$\Delta_{\text{own-other}} = \mathrm{AP}(u_{\text{own}}) - \mathrm{AP}(u_{\text{other}})$$3, $$\Delta_{\text{own-other}} = \mathrm{AP}(u_{\text{own}}) - \mathrm{AP}(u_{\text{other}})$$4 [0.705, 1.101], n = 156; and 24+ months: $$\Delta_{\text{own-other}} = \mathrm{AP}(u_{\text{own}}) - \mathrm{AP}(u_{\text{other}})$$5, 95% CI $$\Delta_{\text{own-other}} = \mathrm{AP}(u_{\text{own}}) - \mathrm{AP}(u_{\text{other}})$$6, $$\Delta_{\text{own-other}} = \mathrm{AP}(u_{\text{own}}) - \mathrm{AP}(u_{\text{other}})$$7 [0.602, 1.093], n = 65 (Nakayashiki et al., 11 Jun 2026). The paper concludes that there is no statistically detectable paired decline in the fine layer at any reported horizon (Nakayashiki et al., 11 Jun 2026).

This does not imply immutability. The same study reports slow within-person drift, with a profile built from the recent half of a reader’s history beating the old half by +0.042 [0.020, 0.064], n = 248, while the old half still retains about 91% of the recent half’s score (Nakayashiki et al., 11 Jun 2026). It also shows that held-out-domain scoring retains about 0.91–0.96 of the full advantage, indicating that the signal is not mainly due to repeated domains (Nakayashiki et al., 11 Jun 2026). The practical conclusion is that co-readership identity control can reveal a durable reading signature, but one that evolves slowly rather than remaining fixed.

In this temporal setting, co-readership identity control shifts from measuring whether individuality exists to measuring how long a fixed behavioral profile remains predictive. The evidence in this literature supports the claim that the selection signature is durable enough to function as a trait-like behavioral regularity in heavy, long-tenured readers (Nakayashiki et al., 11 Jun 2026).