Recency Delta: Temporal Influence

• Recency Delta is a cross-domain metric that quantifies how recent events influence current behavior, using elapsed time and rank-based measures.
• It is applied in fields like human mobility, network growth, and temporal retrieval to modulate return probability, attachment, and memory retention.
• Practical implementations include enhancing location prediction, refining search relevance, and improving biomedical recency classification through dynamic decay functions.

Searching arXiv for the cited papers to ground the article and confirm metadata. Recency Delta is a cross-domain technical notion for quantifying how strongly the present is shaped by the recent past. The term is not standardized across literatures: in human mobility it is the elapsed time since a location was last visited, in temporal network growth it is an age difference between vertices, in temporal retrieval it is a freshness gap relative to a reference time, and in several modern sequence models it appears as a decay-rate or context-sensitive recency signal rather than a literal elapsed-time variable (Barbosa et al., 2015, Prokhorenkova et al., 2014, Cao et al., 1 Sep 2025, Wang et al., 2024, Shen et al., 19 Jun 2026). A plausible synthesis is that these usages all formalize recency as a variable that modulates return probability, attachment, retrieval, memory retention, or control.

1. Terminological scope and canonical forms

Across the literature, Recency Delta is best understood as a family of related constructs rather than a single invariant definition. Some papers define it explicitly, while others introduce the closest object under names such as recency gap, age delta, recency rank, influence decay rate, or recency encoding.

Domain	Operational notion	Representative form
Human mobility	Time since most recent visit	$\Delta t_i(t)=t-t_{\mathrm{last}}(i)$
Temporal attachment	Age difference of a vertex	$\Delta_i(t)=t-i$
Temporal retrieval	Freshness distance of a document version	$$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$
State-space analysis	Exponential decay-rate proxy	$$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$
KV cache adaptation	Fractional recency-share change	$\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$
Knowledge tracing	Steps since a KC last appeared	$r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$

A useful editorial taxonomy is to separate elapsed-time deltas from rank or control deltas. Elapsed-time deltas measure how far back the last relevant event lies, as in $\Delta t_i(t)$, $\Delta_i(t)$, or $\Delta_{\mathrm{rec}}$. Rank or control deltas instead encode recency indirectly through order statistics, decay exponents, or dynamic resource reallocation. This suggests that Recency Delta is less a fixed observable than a design principle for expressing temporal proximity in probabilistic, algorithmic, or decision-theoretic systems.

2. Human mobility: recency as return bias

In human mobility, Recency Delta is formalized most directly as the time since the last visit to a location, $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$, together with a rank-based recency variable $\Delta_i(t)=t-i$0 that orders previously visited locations by how recently they were seen (Barbosa et al., 2015). The return-time distribution obeys a truncated power law,

$\Delta_i(t)=t-i$1

with fitted parameters $\Delta_i(t)=t-i$2 hours for D1 (CDR) and $\Delta_i(t)=t-i$3 hours for D2 (Brightkite). The same data exhibit pronounced peaks at $\Delta_i(t)=t-i$4 hours and weekly periodicities, indicating that recency decay is superposed with daily and weekly routines.

The paper also defines a frequency rank $\Delta_i(t)=t-i$5 and shows that both $\Delta_i(t)=t-i$6 and $\Delta_i(t)=t-i$7 are heavy-tailed, but statistically distinct. For D1, the fitted truncated-power-law parameters are $\Delta_i(t)=t-i$8 for $\Delta_i(t)=t-i$9 and $$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$0 for $$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$1; for D2 they are $$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$2 for $$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$3 and $$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$4 for $$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$5. The crucial empirical finding is that low $$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$6 increases return probability even for medium- or high-$$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$7 locations. This isolates a genuine recency effect beyond preferential return to highly visited places.

The recency-augmented mobility model extends exploration and preferential return by mixing a recency-driven exploitation channel with a frequency-driven one. Exploration is specified as

$$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$8

with $$\Delta_{\mathrm{rec}}=\lvert t_{\mathrm{ref}}(Q)-t_d\rvert$$9 and $$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$0, while exploitation uses

$$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$1

where

$$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$2

The calibrated values $$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$3 and $$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$4 imply that most exploitation steps are recency-driven. Randomization tests further show that shuffling trajectories destroys the recency signature, supporting the claim that the effect arises from genuine temporal ordering rather than from visitation frequency alone.

The significance of this formulation is methodological as well as descriptive. Frequency-only EPR reproduces Zipf-like visitation frequencies but fails to capture the observed power-law-like recency-rank distribution and the elevated revisit probability of recently discovered places. Recency Delta therefore functions here as an explicit latent variable for short-memory exploitation.

3. Age difference and temporal assortativity in network growth

In recency-based preferential attachment, the central delta is the age difference

$$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$5

where $$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$6 is the birth time of an existing vertex and $$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$7 is the current time (Prokhorenkova et al., 2014). This quantity enters the attractiveness function through a decay term. The paper analyzes two principal forms:

$$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$8

and

$$\Delta_{\mathrm{rec}}\equiv \kappa=\log(A_{\max}^{-1})$$9

The first is an exponential recency model with mean lifetime parameter $\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$0; the second is a finite-memory window.

Under i.i.d. Pareto qualities,

$\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$1

both recency constructions yield a power-law degree distribution with exponent $\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$2. For the exponential model,

$\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$3

over the stated admissible degree range, and analogous asymptotics hold for the window model. The total attractiveness in the exponential case satisfies

$\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$4

for $\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$5.

The age-sensitive quantity $\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$6, the fraction of edges whose endpoints differ in age by more than $\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$7, makes the temporal meaning of Recency Delta particularly explicit. In the exponential model,

$\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$8

while in the window model

$\Delta_r=(p_{\mathrm{new}}-p_{\mathrm{old}})/c$9

Thus, larger age differences are exponentially or linearly suppressed, depending on the recency kernel. The paper further notes that the diameter scales as

$r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$0

reflecting a chain-like temporal backbone induced by strong preference for similarly aged vertices.

Here Recency Delta is not a revisit lag but a structural age coordinate. Its main role is to encode temporal assortativity: smaller $r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$1 means a newer target, larger attractiveness, and tighter coupling between contemporaneous vertices. A plausible implication is that recency acts as a generative prior over temporal neighborhoods, while the Pareto quality distribution controls the eventual scale-free tail.

4. Temporal retrieval, search ranking, and question answering

In temporal information retrieval, the closest construct to Recency Delta is the recency gap

$r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$2

defined as the absolute day difference between a candidate document’s publication or update time $r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$3 and a scenario-dependent reference time $r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$4 (Cao et al., 1 Sep 2025). Re3 pairs this with a relevance gap,

$r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$5

and encodes both through multi-frequency Fourier features before fusing semantic and time-aware scores via

$r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$6

On Re2Bench, Re3 achieves $r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$7 on Re2-Rel, $r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$8 on Re2-Rec, and $r_t^{(k)}=t-\max\{i<t:k\in m(q_i)\}$9 on Re2-Hyb. Ablation results show that removing the gate causes Hybrid $\Delta t_i(t)$0 to fall from $\Delta t_i(t)$1 to $\Delta t_i(t)$2, indicating that the learned balance between semantic relevance and freshness is not a minor refinement but a structural component of the scoring rule.

A related web-search formulation treats recency sensitivity as a query-dependent diversification problem rather than a pure freshness score (Styskin et al., 2024). Queries receive a learned probability of recent intent, $\Delta t_i(t)$3, estimated by gradient boosted regression trees from features such as language-model probabilities over recent streams and news-click propensity. Ranking is then optimized with an intent-aware ERR-IAA objective over two intents, “recent” and “general,” using a freshness window of $\Delta t_i(t)$4 days and an abandonment parameter $\Delta t_i(t)$5. In online A/B testing over approximately $\Delta t_i(t)$6 million queries per bucket, abandonment rate decreased from $\Delta t_i(t)$7 to $\Delta t_i(t)$8, time to first click from $\Delta t_i(t)$9 to $\Delta_i(t)$0 seconds, and first-position CTR increased from $\Delta_i(t)$1 to $\Delta_i(t)$2.

RecencyQA generalizes this line of work from document freshness to answer-change frequency (Piryani et al., 17 Mar 2026). It defines recency as “the expected temporal stability of a question’s answer” and introduces $\Delta_i(t)$3 recency classes, from “An-Hour” to “Never,” together with a binary stationarity label. The dataset contains $\Delta_i(t)$4 questions, of which $\Delta_i(t)$5 are stationary and $\Delta_i(t)$6 non-stationary. Human evaluation reports recency accuracy $\Delta_i(t)$7 in the strict setting and $\Delta_i(t)$8 under a $\Delta_i(t)$9-bin tolerance, with stationarity accuracy $\Delta_{\mathrm{rec}}$0. Empirically, non-stationary questions are more difficult for LLMs, and dynamic recency transitions remain challenging: for Gemma 3 (27B), few-shot transition accuracy is $\Delta_{\mathrm{rec}}$1 even though per-context accuracies are $\Delta_{\mathrm{rec}}$2 and $\Delta_{\mathrm{rec}}$3.

These works collectively shift Recency Delta from a simple lag variable to a task-conditioned validity signal. In retrieval it measures freshness distance; in search ranking it measures uncertainty over whether freshness should matter; in QA it becomes a taxonomy of answer volatility and context dependence.

5. Recency bias, decay, and recency encoding in sequential models

For structured state space models, recency is quantified by the influence score

$\Delta_{\mathrm{rec}}$4

with the theoretical bound

$\Delta_{\mathrm{rec}}$5

under diagonal $\Delta_{\mathrm{rec}}$6 with entries in $\Delta_{\mathrm{rec}}$7 (Wang et al., 2024). The paper treats $\Delta_{\mathrm{rec}}$8 as an operational recency proxy: larger $\Delta_{\mathrm{rec}}$9 implies faster decay of distant contributions. It also shows that deeper SSMs face an over-smoothing trade-off, and proposes polarization of two channels of the transition matrix, fixing one to $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$0 and one to $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$1. On associative recall, the configuration with both polarized channels and $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$2 layers reaches average accuracy $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$3, compared with $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$4 for the default $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$5-layer model.

A mechanistic analysis of Mamba links recency to $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$6-modulated recurrence and documents a U-shaped profile of primacy and recency in structured recall tasks (Airlangga et al., 18 Jun 2025). Recent inputs gain weight through exponential decay, but this recency advantage collapses when $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$7 or $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$8 distractor tokens are inserted before the query. The same study identifies sparse long-term-memory channels, with a notable concentration in Layer $\Delta t_i(t)=t-t_{\mathrm{last}}(i)$9 of Falcon Mamba 7B, and shows by targeted ablation that these channels are causally linked to primacy. Semantic regularity also modulates the $\Delta_i(t)=t-i$00 gate: repeated relations increase forgetting of intermediate items, sharpening the lost-in-the-middle regime.

In recurrent recommendation, recency appears as a training bias toward short-horizon user interests (Chang et al., 2022). Recency dropout mitigates this by randomly removing the most recent $\Delta_i(t)=t-i$01 interactions during training and computing the state from the truncated prefix. The baseline REINFORCE model exhibits strong short-term concentration, with $\Delta_i(t)=t-i$02 and $\Delta_i(t)=t-i$03 for $\Delta_i(t)=t-i$04. Live experiments report positive shifts in overall enjoyment, DAU, and diversity: on homepage recommendations with $\Delta_i(t)=t-i$05, overall enjoyment increases by $\Delta_i(t)=t-i$06 and diversity by $\Delta_i(t)=t-i$07; on short-form content with $\Delta_i(t)=t-i$08, overall enjoyment increases by $\Delta_i(t)=t-i$09 and diversity by $\Delta_i(t)=t-i$10.

Knowledge tracing introduces an explicitly symbolic recency variable,

$\Delta_i(t)=t-i$11

defined at the knowledge-concept level on the original question sequence rather than the expanded KC sequence (Badran et al., 23 Aug 2025). This scalar is mapped to a learnable Fourier representation and added to DKT, DKT+, SAKT, and AKT embeddings. The paper couples this with a MASK-based embedding strategy to prevent label leakage in multi-KC items. The resulting recency-aware variants improve AUC across several datasets; for example, AKT-ML$\Delta_i(t)=t-i$12 reaches $\Delta_i(t)=t-i$13 on ASSISTments2009, $\Delta_i(t)=t-i$14 on Algebra2005, $\Delta_i(t)=t-i$15 on Riiid2020, and $\Delta_i(t)=t-i$16 on Duolingo2018.

Across these model classes, Recency Delta becomes a decay constant, an input gap, or an embedding feature. This suggests that recency is not merely a nuisance bias; it is also a controllable inductive bias whose usefulness depends on whether the task rewards short-horizon adaptivity, long-range retention, or both.

6. Control, verification, and biomedical extensions

In formal verification, recency is imposed as a decidability-enabling constraint rather than an empirical regularity (Abdulla et al., 2016). A $\Delta_i(t)=t-i$17-recency-bounded run in a database-manipulating system permits actions to modify only the most recent $\Delta_i(t)=t-i$18 elements in the current active domain. The paper proves that recency-bounded model checking of DMS against MSO-FO is decidable by reduction to satisfiability of MSO over nested words. Here Recency Delta is effectively the verification under-approximation parameter: increasing $\Delta_i(t)=t-i$19 monotonically enlarges the behavior space.

In online preference learning and LLM serving, recency is treated as a systems-level control variable (Nguyen, 11 Apr 2026, Shen et al., 19 Jun 2026). For normalized Kaczmarz-inspired online learners, the contribution of the $\Delta_i(t)=t-i$20-th interaction decays as $\Delta_i(t)=t-i$21; with $\Delta_i(t)=t-i$22, an interaction $\Delta_i(t)=t-i$23 swipes old has weight approximately $\Delta_i(t)=t-i$24. BlockNK removes per-step normalization and yields direction stability $\Delta_i(t)=t-i$25, versus $\Delta_i(t)=t-i$26 for NK, giving a reported Recency Delta of $\Delta_i(t)=t-i$27 at label-noise level $\Delta_i(t)=t-i$28. In KV caching, ARC-style management defines

$\Delta_i(t)=t-i$29

the fractional change in cache share assigned to recency. This adaptive split between recency and frequency improves KV-cache hit rate by up to $\Delta_i(t)=t-i$30 and reduces time to first token by up to $\Delta_i(t)=t-i$31 on synthetic document QA workloads.

Biomedical uses adopt yet another interpretation. In probabilistic HIV recency classification, Recency Delta refers to the improvement of a semi-supervised logistic model over a binary classification tree, both at individual and aggregated levels (Sheng et al., 2021). The model calibrates

$\Delta_i(t)=t-i$32

to the national recency proportion $\Delta_i(t)=t-i$33 among non-ART HIV-positive individuals in Malawi. In a complementary Bayesian framework, recency is the posterior probability that seroconversion occurred within the last $\Delta_i(t)=t-i$34 months, $\Delta_i(t)=t-i$35, estimated from longitudinal biomarker trajectories and improved by joint modeling of multiple biomarkers (Koulai et al., 2017). At the population level, cross-sectional HIV incidence estimation based on recency tests yields an algebraic Recency Delta between the FRR-adjusted and perfect-specificity estimators,

$\Delta_i(t)=t-i$36

which arises from different tail assumptions on the duration-specific recent-test probability $\Delta_i(t)=t-i$37 (Gao et al., 2021).

These extensions show that the term can denote a tunable bound, a stability gain, a cache-allocation update, or a calibration difference. The common structure is still temporal: Recency Delta measures what changes when recent evidence, recent events, or recent entities are given privileged weight.

7. Conceptual significance and recurring trade-offs

The surveyed literature converges on a small set of recurring trade-offs. Strong recency improves responsiveness to change, as in human mobility, temporal retrieval, adaptive caching, or online recommendation, but it can also suppress long-range structure, as in SSM over-smoothing or Mamba’s distractor-sensitive short-term memory (Barbosa et al., 2015, Cao et al., 1 Sep 2025, Wang et al., 2024, Airlangga et al., 18 Jun 2025). Weak recency preserves history, but may fail to track volatility, staleness, or regime shifts.

A second recurring theme is that recency is often insufficient on its own. Mobility requires recency plus visitation frequency. Retrieval requires recency plus semantic relevance and, for some questions, stationarity awareness. Knowledge tracing requires recency plus leakage-free embeddings. HIV incidence estimation requires recency-test dynamics plus explicit modeling of false-recent tails. This suggests that Recency Delta usually enters as one coordinate in a mixed mechanism rather than as a self-sufficient predictor.

A third theme is representational choice. The same substantive idea can be encoded as an elapsed-time gap, a rank, an exponential decay constant, a dynamic gate, a state-space bound, or a resource-allocation variable. That variability explains why the expression remains non-standard across fields. Even so, the underlying mathematical role is remarkably stable: Recency Delta specifies how strongly the effect of an event, item, location, document, or interaction decays as it moves away from the present.