Recall: Coverage, Recovery & Retention

• Recall is a multifaceted concept defined as the degree to which relevant information is retrieved or retained, with applications in retrieval systems, cognitive science, and continual learning.
• In retrieval settings, recall measures the ratio of recovered relevant items to the total available, emphasizing search completeness, semantic relevance, and robust evaluation in high-risk domains.
• Practical applications span high-stakes patent and medical retrieval, adaptive continual learning, and memory recall in language models and games, underscoring its importance in system reliability.

Recall denotes a family of coverage, recovery, and retention notions rather than a single invariant quantity. In classical information retrieval, it measures how much of the relevant set has been retrieved; in vector search and long-form generation, recent work reframes it around semantic relevance and fact coverage; in cognitive science, it refers to recovery of stored items from memory; in extensive-form games, imperfect recall refers to forgetting one’s own prior decision history; and in continual learning, “recall” frequently names mechanisms for preserving or recovering previously learned behavior under distribution shift or sequential adaptation (Schwartz et al., 24 Dec 2025, Kuffo et al., 22 Apr 2026, Naim et al., 2019, Gimbert et al., 19 Feb 2025).

1. Formal foundations

In classical retrieval notation, if $n_p$ is the number of relevant documents among the top $K$ retrieved items and $N_p$ is the total number of relevant documents in the corpus, precision and recall are

$P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$

The same coverage idea appears in set-based form as

$$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$

where $R$ is the set of relevant items. In this formulation, recall is fundamentally about missed relevant evidence, not ranking quality per se (Schwartz et al., 24 Dec 2025, Diaz et al., 2023).

This coverage interpretation is especially salient in high-risk retrieval settings. Patent retrieval is described as recall-oriented because missing even one relevant prior-art document can undermine novelty or inventive-step assessment, while legal, patent, and medical retrieval are explicitly identified as domains in which recall is an important measure of effectiveness (Ali et al., 20 Jul 2025, Webber, 2012). This suggests that recall is best understood not as a universal scalar of quality, but as a task-dependent measure of completeness.

A further complication is that “true recall” often requires knowledge of the entire relevant set. In large or evolving knowledge bases, that quantity is often unknown “in advance for an arbitrary query,” which makes classical recall formally central but operationally elusive outside fixed benchmarks (Schwartz et al., 24 Dec 2025). That tension motivates much of the recent literature on variants, surrogates, and structured reformulations of recall.

2. Ranked retrieval, totality, and uncertainty

For ranked outputs, recall has been reinterpreted as a concern with totality: whether a user can find every relevant item, not merely some relevant items near the top. In that setting, the crucial object is the sorted vector of relevant-item positions $p=(p_1,\dots,p_m)$, where $p_m$ is the position of the lowest-ranked relevant item. The paper "Recall, Robustness, and Lexicographic Evaluation" defines Total Search Efficiency

$\tse_e(\pi,R)=e(p_m),$

so that the efficiency of finding all relevant items depends only on the last relevant item. It then defines LexiRecall by comparing the relevant-item position vectors from the bottom upward: $$\pi \succ_{LR} \pi' \leftrightarrow p_i < p'_i,\quad i=\max\{j\in[1,m]:p_j\neq p'_j\}.$$ This formalizes recall-orientation as sensitivity to the bottom-ranked relevant item and links it to worst-case robustness over users and providers (Diaz et al., 2023).

The same paper reports that LexiRecall is strongly aligned with standard recall metrics while being much less tie-prone than $K$0 or R-precision. For $K$1 and $K$2, the reported tie probabilities are $K$3, $K$4, and $K$5. This is significant because it preserves a totality-oriented semantics without collapsing most comparisons into indifference (Diaz et al., 2023).

When exhaustive judgments are unavailable, recall must be estimated. In sampled evaluation, with $K$6 relevant documents in the retrieved segment and $K$7 in the unretrieved segment, recall is written as

$K$8

"Approximate Recall Confidence Intervals" studies two-tailed interval estimation for this quantity and recommends deriving beta-binomial posteriors on retrieved and unretrieved yield with fixed hyperparameters $K$9, then using Monte Carlo to estimate the posterior distribution of recall. The paper reports that this method gives mean coverage at or near the nominal level while being balanced and stable (Webber, 2012).

A high-recall application is patent retrieval. "FullRecall: A Semantic Search-Based Ranking Approach for Maximizing Recall in Patent Retrieval" uses IPC-guided knowledge, noun-phrase extraction, and ranking over an IPC-constrained candidate set, and reports 100% recall in all five test cases. The same details block reports HRR2 recall values of 10%, 25%, 33.3%, 0%, and 14.29%, and ReQ-ReC values of 50%, 25%, 0%, 0%, and 0% on those cases (Ali et al., 20 Jul 2025). A plausible implication is that recall-preserving pruning can be operationally useful when the review set remains large but complete.

3. Semantic recall in vector search and retrieval-augmented generation

Approximate nearest neighbor search traditionally uses exact-neighbor fidelity as its target. If $N_p$0 is the set returned by the approximate method and $N_p$1 is the brute-force top-$N_p$2 set, standard recall is

$N_p$3

"Semantic Recall for Vector Search" argues that this can be a poor proxy for retrieval quality because exact nearest neighbors in an embedding space are not guaranteed to be semantically relevant. It therefore defines the semantic-neighbor set $N_p$4 as the subset of the brute-force top-$N_p$5 neighbors that are semantically relevant, and defines semantic recall as

$N_p$6

The same paper introduces Tolerant Recall for top-20 retrieval,

$N_p$7

where $N_p$8 is the largest subset that can be matched one-to-one to ground-truth items either exactly or within an $N_p$9 score tolerance (Kuffo et al., 22 Apr 2026).

The empirical gap between geometric recall and semantic recall is large on sparse-relevance queries. On MSMARCO with ScaNN and 8-bit quantization, the paper reports average top-100 values of traditional recall $P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$0, semantic recall $P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$1, and tolerant recall $P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$2; on queries with fewer than 20 semantic neighbors, the values are $P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$3, $P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$4, and $P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$5, respectively. On MIRACL, average traditional, semantic, and tolerant recall are $P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$6, $P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$7, and $P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$8. The same study reports that tuning ScaNN for 98.93% semantic recall yielded 14% lower cost while preserving semantic recall, and that targeting 95% tolerant recall instead of 95% traditional recall reduced cost by about 25% on BigANN, 5% on GloVe, and 35% on MSMARCO (Kuffo et al., 22 Apr 2026).

In RAG-style retrieval, the obstacle is often not semantic mismatch but the fact that $P=\frac{n_p}{K}, \qquad R=\frac{n_p}{N_p}.$9 is unknown. "How important is Recall for Measuring Retrieval Quality?" studies alternatives to exact recall by correlating retrieval metrics with LLM-based judgments of answer completeness. It considers the benchmark $$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$0 that depends on $$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$1, an estimated variant $$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$2, $$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$3, and a recall-free measure

$$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$4

with $$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$5. The paper states that “for many of the segments considered, the $$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$6 measure does not lose much in terms of correlations compared to $$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$7 even though it is formulated without $$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$8,” and that at ratios $$\mathrm{Recall}=\frac{|\,\mathrm{retrieved}\cap R\,|}{|R|},$$9 comparable to 1 or lower, $R$0 is better than $R$1, and for ratios below 1 it is also better than or the same as $R$2, depending on the document type (Schwartz et al., 24 Dec 2025).

4. Fact coverage, diversity, and exhaustiveness in generated or extracted text

In long-form generation, recall can be defined at the level of reference facts rather than retrieved documents. "LongRecall: A Structured Approach for Robust Recall Evaluation in Long-Form Text" defines a reference set of facts $R$3 and generated facts $R$4, and writes recall as

$R$5

where $R$6 is the set of generated facts that cover or entail $R$7. The framework decomposes answers into self-contained facts, applies lexical then semantic candidate filtering, and verifies coverage through localized entailment checks (Ardestani et al., 20 Aug 2025).

The same paper reports large gains over lexical and holistic baselines. On the standard subset, LongRecall$R$8 reaches F1 $R$9 on QAMPARI and $p=(p_1,\dots,p_m)$0 on RoMQA, compared with ARecall $p=(p_1,\dots,p_m)$1 and $p=(p_1,\dots,p_m)$2; on the challenging subset, the values are $p=(p_1,\dots,p_m)$3 and $p=(p_1,\dots,p_m)$4, compared with $p=(p_1,\dots,p_m)$5 and $p=(p_1,\dots,p_m)$6. On ExpertQA, LongRecall$p=(p_1,\dots,p_m)$7 reaches $p=(p_1,\dots,p_m)$8 on the standard subset and $p=(p_1,\dots,p_m)$9 on the challenging subset, while ERecall is reported as $p_m$0 in both conditions (Ardestani et al., 20 Aug 2025). This suggests that in long-form settings, recall is more faithfully modeled as fact-level semantic coverage than as lexical overlap.

A different decomposition is proposed in "Two Kinds of Recall." That paper distinguishes d-recall, recall as diversity, from e-recall, recall as exhaustiveness. Neural QA models are shown to have strong d-recall by answering across varied contexts, but weaker e-recall on repeated instances of the same construction. In a pattern-defined Wikipedia set of 5,000 sentences matching “was educated at,” the model returned 611 No-Answer outputs; in a PubMed set of 3654 “DISEASE was treated with CHEMICAL” instances, it returned 721 No-Answer outputs (Goldberg, 2023). The paper’s point is not that one notion replaces the other, but that aggregate recall can obscure the distinction between breadth across forms and consistency within a form.

5. Memory recall, language-model recall, and imperfect recall

In cognitive science, recall refers to recovery of studied items from memory rather than retrieval from an indexed corpus. "Fundamental Law of Memory Recall" models free recall as a deterministic walk on random graphs defined by overlaps among sparse memory representations and derives

$p_m$1

where $p_m$2 is average recall and $p_m$3 is the number of items effectively present in memory after acquisition. The same paper estimates $p_m$4 from recognition accuracy $p_m$5 using

$p_m$6

and reports that recall as a function of estimated $p_m$7 collapses onto the predicted curve across list lengths and presentation rates (Naim et al., 2019).

Autoregressive LLMs use the term in still another sense: recovery of earlier text from later text. "RECALL: Library-Like Behavior In LLMs is Enhanced by Self-Referencing Causal Cycles" argues that repeated token sequences act as “cycle tokens” that narrow the inverse search problem. The paper proposes a two-step ReCall-aware prompting strategy—first elicit broad surrounding context, then extract the predecessor—and reports 100% success on its set of key writings for GPT-4o (2024-12-23) and Llama-3.3-70B (Nwadike et al., 23 Jan 2025). In this usage, recall is neither classification coverage nor metric completeness, but backward recovery through corpus-induced causal loops.

A related line studies factual recall inside transformers. "Relation Also Knows" argues that relation information is a distinct locus of factual computation and measures relation mediation with the Indirect Effect of Relation. In the reported averages, the last relation token has AIER $p_m$8 in GPT2-XL and $p_m$9 in GPT-J, far above other token positions. The paper further reports that 68% of correctly predicted objects for GPT2-XL and 83% for GPT-J lie in the collected relation-attribute sets, and that the Spearman correlation between average negative object rank and average attributes rate is $\tse_e(\pi,R)=e(p_m),$0 for GPT2-XL and $\tse_e(\pi,R)=e(p_m),$1 for GPT-J (Liu et al., 2024). A plausible implication is that factual recall in autoregressive transformers is not exhausted by subject-centered interpretations.

In extensive-form games, recall refers to remembered decision history. "Simplifying imperfect recall games" defines perfect recall by the condition that for any information set $\tse_e(\pi,R)=e(p_m),$2, all nodes $\tse_e(\pi,R)=e(p_m),$3 satisfy $\tse_e(\pi,R)=e(p_m),$4. It then studies A-loss recall, where information loss can be traced to forgetting which own action was taken at some earlier information set, and proves that every non-absent-minded game structure has an A-loss-recall span (Gimbert et al., 19 Feb 2025). This formalizes recall as a property of information structure rather than of memory performance or retrieval evaluation.

6. Recall as retention and recovery in continual learning and adaptive systems

In continual learning, recall often denotes preservation or recovery of previously acquired behavior under sequential updating. The 2025 paper "RECALL: REpresentation-aligned Catastrophic-forgetting ALLeviation via Hierarchical Model Merging" proposes a representation-aware model merging framework for continual learning without access to historical data. According to its abstract, RECALL computes inter-model similarity from layer-wise hidden representations over clustered typical samples and performs adaptive, hierarchical parameter fusion so that shallow layers preserve domain-general features while deeper layers allow task-specific adaptation (Wang et al., 23 Oct 2025).

Several other methods use the same name for replay or retention mechanisms. "RECALL: Rehearsal-free Continual Learning for Object Classification" preserves old-category logits on new-sequence images and trains a growing multi-head network with regression-style losses. The reported final accuracies are 61.15 $\tse_e(\pi,R)=e(p_m),$5 on iCIFAR-100 and 57.83 $\tse_e(\pi,R)=e(p_m),$6 on HOWS-CL-25 for RECALL, while RECALL var. reg. reaches 71.45 $\tse_e(\pi,R)=e(p_m),$7 on CORe50 and 40.65 $\tse_e(\pi,R)=e(p_m),$8 on HOWS-CL-25 long (Knauer et al., 2022). In semantic segmentation, "RECALL: Replay-based Continual Learning in Semantic Segmentation" recreates old-class data through a pre-trained GAN or web-crawled images and uses background self-inpainting; on the VOC 10-1 overlapped setting, the reported all-class mIoU is 60.7 for RECALL (Web), compared with 15.0 for MiB and 23.2 for SDR (Maracani et al., 2021).

The same retention theme appears in reinforcement learning and robotics. "Replay-enhanced Continual Reinforcement Learning" introduces RECALL as a replay-enhanced method combining adaptive target normalization and policy distillation; on CW3, it reports average performance $\tse_e(\pi,R)=e(p_m),$9, forgetting $$\pi \succ_{LR} \pi' \leftrightarrow p_i < p'_i,\quad i=\max\{j\in[1,m]:p_j\neq p'_j\}.$$0, and forward transfer $$\pi \succ_{LR} \pi' \leftrightarrow p_i < p'_i,\quad i=\max\{j\in[1,m]:p_j\neq p'_j\}.$$1, compared with Perfect Memory $$\pi \succ_{LR} \pi' \leftrightarrow p_i < p'_i,\quad i=\max\{j\in[1,m]:p_j\neq p'_j\}.$$2 average performance and $$\pi \succ_{LR} \pi' \leftrightarrow p_i < p'_i,\quad i=\max\{j\in[1,m]:p_j\neq p'_j\}.$$3 forward transfer (Zhang et al., 2023). In vision-language-action models, "RECALL: Recovery Experience Collection for Active Lifelong Learning in Vision-Language-Action Models" reports that Strong INSIGHT online recovery with full replay reaches 72.4% overall success, versus 60.2% for passive collection and 59.8% for the original baseline, while new-only fine-tuning catastrophically forgets (Karli et al., 22 Jun 2026).

The term also extends to systems built to support personal memory. "Recall: Empowering Multimodal Embedding for Edge Devices" introduces an on-device multimodal embedding system that stores coarse-grained embeddings and refines candidates at query time. The paper reports 14.9× average throughput improvement, 13.1× average energy reduction, and <5% relative accuracy loss compared with the full multimodal embedding model (Cai et al., 2024). This suggests a broad unifying pattern: across continual learning, replay, and personal-memory systems, recall increasingly denotes controlled resistance to forgetting rather than a single metric of retrieval completeness.