Early Termination by Recall

• Early Termination by Recall is a family of stopping mechanisms that halts computation using target recall metrics, recovered structural information, or surrogate continuation values.
• It applies to diverse areas such as approximate nearest neighbor search, clustering, and reinforcement learning to balance computational efficiency and accuracy.
• Practical implementations demonstrate significant speedups and controlled error trade-offs through methods like learned predictors, graph traversal heuristics, or analytical continuation payoffs.

Early termination by recall denotes a family of stopping mechanisms in which a process is halted before its natural endpoint using information tied to target recall, recovered structure, recalled feasible plans, or analytically recalled continuation values. The phrase is not standardized across the literature represented here. In approximate nearest neighbor search, recall is explicit and declarative: a target recall $R_t$ is requested, a runtime predictor estimates current recall $R_p$, and search terminates when $R_p \ge R_t$ (Chatzakis et al., 25 May 2025). In large-scale clustering, early stopping and recall are deliberately separated: the algorithm first stops neighbor traversal to protect precision and then performs edge recall to recover missed positives (Liu, 2024). In control and reinforcement learning, related designs replace continued rollout by a recalled feasible plan or an analytically computed continuation payoff rather than by an explicit recall metric (Amiri et al., 1 Jul 2025, Das et al., 16 Mar 2026).

1. Conceptual scope and terminological variants

Across the cited work, “recall” enters early termination in several distinct ways. In one regime, recall is the formal target of the stopping rule. DARTH overlays an approximate nearest neighbor search index with a learned online recall estimator and exposes a declarative API $ANNS(q,G,k,R_t)$, where the user specifies the desired recall and the system decides how much search effort to spend (Chatzakis et al., 25 May 2025). In a second regime, recall is not the stopping signal but the quantity that must be repaired after stopping. FC-ES halts neighbor exploration at the first weak graph-consistency event, while FC-ESER later recalls missed edge connections through a separate supervised stage (Liu, 2024). In a third regime, recall is only an interpretive analogy: some methods stop on proxies for recoverability, decision stability, or continuation value rather than on an explicit recall statistic (Gümüş et al., 4 Jul 2025, Albayrak et al., 2016, Das et al., 16 Mar 2026).

A compact comparison is useful.

Regime	Stopping signal	Role of recall
Declarative recall control	$R_p \ge R_t$	Recall is the objective
Early stopping plus repair	$\widetilde p_{ij}<\theta$ in traversal	Recall is recovered afterward
Recoverability proxy	Reliability decrease or sign stability	Recall is indirect
Recalled continuation or plan	Analytical payoff or shifted feasible sequence	Recall supplies the substitute continuation

The same vocabulary also has domain-specific meanings that are not metric-oriented. In instruction-tuned diffusion LLMs, early termination can arise from a “suffix contamination” mechanism in which repeated $\langle eos\rangle$ padding is effectively reinstated during denoising; the paper explicitly relates this to a recall-like effect from the padded suffix rather than to a stopping policy (Kim et al., 4 Oct 2025). In bilateral credit valuation adjustment, a recall or break clause is a contractual right to terminate a transaction at predefined dates, priced as Bermudan-style optionality under counterparty risk rather than as a quality-control device (Giada et al., 2012). These cases widen the semantic field of “early termination by recall,” but they do not collapse to a single technical definition.

2. Declarative recall as a runtime stopping objective

The clearest explicit formulation appears in “DARTH: Declarative Recall Through Early Termination for Approximate Nearest Neighbor Search” (Chatzakis et al., 25 May 2025). DARTH assumes a high-recall underlying index and learns, during search, to estimate the current recall of the ongoing query from online search-state signals. The runtime decision rule is deliberately simple: if the predicted recall $R_p$ reaches the requested target $R_t$, search stops and the current result set is returned. The paper defines per-query error as $\text{error} = |R_t - R_q|$, assumes target attainability through $R_p$0, and uses an adaptive predictor-invocation interval

$R_p$1

The predictor is a Gradient Boosting Decision Tree implemented with LightGBM, using 100 estimators. It is trained from logged search traces whose features summarize both progress and the evolving geometry of the current result set: $R_p$2, $R_p$3, $R_p$4, $R_p$5, $R_p$6, $R_p$7, $R_p$8, $R_p$9, $R_p \ge R_t$0, $R_p \ge R_t$1, and $R_p \ge R_t$2. This makes recall estimation query-adaptive rather than a static translation from target recall to a backend parameter such as HNSW’s $R_p \ge R_t$3 or IVF’s $R_p \ge R_t$4.

The empirical claims are unusually specific. For HNSW, DARTH is reported to always meet or exceed the target on average, with speedups up to 14.6x (average: 6.8x; median: 5.7x) faster than search without early termination. For IVF, it also always meets and exceeds the target on average and reaches speedups up to 41.8x (average: 13.6x; median: 8.1x). On HNSW, DARTH performs, on average, only 5% more distance calculations than an oracle that knows the exact earliest point at which the target recall is first reached. The paper is careful, however, not to claim a per-query guarantee: on SIFT100M with $R_p \ge R_t$5, roughly 15% of queries may fall below the target for a given setting even though the workload average exceeds $R_p \ge R_t$6. That distinction is central. DARTH is a learned runtime controller for recall, not a certified lower-bound method.

3. Early stopping followed by explicit recall in clustering

“Face Clustering via Early Stopping and Edge Recall” develops a different architecture in which stopping and recall are separated by design (Liu, 2024). FC-ES constructs a KNN graph over normalized embeddings and replaces raw cosine similarity with a neighbor-based edge probability $R_p \ge R_t$7 derived from common-neighbor structure. Traversal then proceeds in the original cosine-ranked KNN order. For each neighbor $R_p \ge R_t$8, the edge $R_p \ge R_t$9 is kept if $ANNS(q,G,k,R_t)$0; the first time $ANNS(q,G,k,R_t)$1, traversal stops immediately for node $ANNS(q,G,k,R_t)$2. The paper emphasizes that $ANNS(q,G,k,R_t)$3 is not sorted before this test. That detail is the mechanism: the first low-probability event in the cosine-ordered list is treated as the boundary where neighbors begin to look like weak links or negatives.

This stopping rule is intentionally conservative. It protects precision and graph cleanliness, but it also misses some true edges after the stopping point. FC-ESER therefore adds a separate edge recall stage. After the first break, later neighbors with score above a looser threshold $ANNS(q,G,k,R_t)$4 are gathered into a hard connection set $ANNS(q,G,k,R_t)$5. A supervised linkage predictor, PCENet in the experiments, is then trained on $ANNS(q,G,k,R_t)$6 and recalled edges are accepted when the predictor score $ANNS(q,G,k,R_t)$7. In the method’s own terms, early stopping does not optimize recall directly; recall is first preserved indirectly by the robustness of $ANNS(q,G,k,R_t)$8 and then recovered explicitly afterward.

The experimental pattern matches that interpretation. Removing early stopping (“w/o ES”) decreases average $ANNS(q,G,k,R_t)$9 from 90.31 to 88.05 on MS1M, from 72.33 to 71.92 on MSMT17, and from 73.12 to 72.92 on VeRi776. Adding edge recall increases $R_p \ge R_t$0 from 90.31 to 91.37 on MS1M, from 72.33 to 73.19 on MSMT17, and from 73.12 to 74.57 on VeRi776. On MSMT17, FC-ES has pairwise precision 85.04 and pairwise recall 58.60; FC-ESER lowers precision slightly to 82.66 but raises recall to 61.78, increasing pairwise $R_p \ge R_t$1 from 69.39 to 70.71. The method is therefore not “stop when recall is high enough.” It is “stop early to avoid false positives, then recall difficult positives in a second stage.”

4. Recoverability proxies and decision consistency in iterative decoders

A substantial part of the literature close to this topic uses no explicit recall metric at all. Instead, early termination is driven by a proxy for likely recoverability or by inter-iteration decision stability.

In “Early Termination of Low-Density Parity-Check Codes for Continuous-Variable Quantum Key Distribution,” the stopping rule is variable node reliability early termination (VNR-ET) (Gümüş et al., 4 Jul 2025). After each belief-propagation iteration $R_p \ge R_t$2, the decoder computes

$R_p \ge R_t$3

where $R_p \ge R_t$4 excludes degree-1 variable nodes. Parity-check early termination remains active, but VNR-ET adds a second criterion: if $R_p \ge R_t$5, decoding stops. This is not successful-recovery stopping, because it does not certify a valid codeword. It is a confidence-based proxy: once aggregate posterior reliability over informative variable nodes begins to decrease, further iterations are treated as unlikely to recover the frame. The paper reports increases in decoded secret key rate throughput of up to 182% in simulation and experiment, while also stressing the corresponding limitation: VNR-ET incurs a frame-error-rate penalty because some decodable frames are terminated early.

A related but technically distinct mechanism appears in “Least reliable messages based early termination method for LT soft decoder” (Albayrak et al., 2016). The proposed LRM method selects the $R_p \ge R_t$6 messages with the smallest $R_p \ge R_t$7 after a determination condition DC-LRM and then monitors only their sign alterations across iterations. Decoding stops when no sign changes occur in this least reliable subset for $R_p \ge R_t$8 consecutive iterations. In the reported settings, LRM uses $R_p \ge R_t$9 and $\widetilde p_{ij}<\theta$0, causes no BER performance degradation, and reduces the early-termination-section computation time by roughly 92% relative to CSR. Conceptually, this is a decision-consistency criterion: the decoder terminates when the weakest messages keep reproducing the same hard-decision sign.

These two decoders illustrate an important distinction. VNR-ET uses a trajectory of aggregate reliability; LRM uses persistence of sign decisions among least reliable messages. Neither paper defines recall in the information-retrieval sense. A plausible implication is that “early termination by recall” in such settings is best interpreted as stopping by estimated recoverability or by repeated local decision agreement rather than by a recovered-fraction statistic.

5. Recalled continuations and recalled feasible plans in sequential decision systems

Sequential decision systems introduce another form of recall: a process can be terminated early if the remaining trajectory is replaced by a recalled continuation value or by a recalled feasible control plan.

In “Game-Theory-Assisted Reinforcement Learning for Border Defense: Early Termination based on Analytical Solutions,” reinforcement-learning episodes terminate at the first sensing time

$\widetilde p_{ij}<\theta$1

and the remainder of the pursuit phase is replaced by the analytically computed Nash equilibrium payoff $\widetilde p_{ij}<\theta$2 (Das et al., 16 Mar 2026). In the GT-assisted formulation, the terminal set is

$\widetilde p_{ij}<\theta$3

and the terminal reward is assigned immediately at detection. The reported effect is a 10% improvement in the homogeneous 1v1 case and an 18.9% improvement in the homogeneous 3v1 case, together with faster and more stable convergence. In 1,000 evaluation episodes for homogeneous 3v1, the sensing rate rises from 45.2% to 67.1%, and the median analytical Nash payoff at sensing rises from 0.701 to 0.768. Here recall is best understood as recall of an analytically derived continuation solution.

“REAP-T: A MATLAB Toolbox for Implementing Robust-to-Early Termination Model Predictive Control” uses recall differently (Amiri et al., 1 Jul 2025). REAP warm-starts each new optimization by recalling the previously accepted feasible control sequence, shifting it by one step, and appending the terminal control law $\widetilde p_{ij}<\theta$4. The optimizer is a virtual primal-dual dynamical system whose primal iterate remains feasible for all internal computation steps $\widetilde p_{ij}<\theta$5, and an acceptance/rejection mechanism ensures that a newly computed iterate is used only if its MPC cost is lower than that of the recalled initial feasible candidate. Early termination is therefore tolerated because the controller can always fall back to the recalled feasible plan. This is not recall as a quality metric; it is recall as remembered feasibility.

Predictive pruning in systems optimization supplies a further contrast. Cello terminates a running sample when the predicted final outcome $\widetilde p_{ij}<\theta$6 exceeds the current best threshold $\widetilde p_{ij}<\theta$7, that is, when $\widetilde p_{ij}<\theta$8 (Ding et al., 2022). The paper explicitly states that it does not report recall or other confusion-matrix metrics, even though the mechanism can be interpreted as a binary screening problem over bad and good runs. Predictive Hyperparameter Optimisation likewise stops most configurations after $\widetilde p_{ij}<\theta$9 iterations, predicts final performance by linear regression, and fully trains only the top $\langle eos\rangle$0 predicted candidates; it explicitly notes that “any other metric can also be used” (Marinov et al., 2019). A plausible implication is that these predictive-pruning frameworks can host recall-oriented stopping only if early recall is itself stable and predictive, a condition not demonstrated in those papers.

6. Trade-offs, guarantees, and common misconceptions

A recurrent misconception is that early termination by recall always means stopping when a formal recall statistic crosses a threshold. The evidence does not support that generalization. In DARTH, recall is the stopping target (Chatzakis et al., 25 May 2025). In FC-ESER, recall is repaired after stopping (Liu, 2024). In VNR-ET and LRM, the operative quantities are reliability decline and sign stability, not recall (Gümüş et al., 4 Jul 2025, Albayrak et al., 2016). In REAP, the critical object is a recalled feasible sequence (Amiri et al., 1 Jul 2025). In GT-assisted border defense RL, the recalled object is an analytical continuation payoff (Das et al., 16 Mar 2026).

Another misconception is that early termination is necessarily aligned with reliability. “On the Role of Early-Termination for Age of Information in Tree-Based Random Access Protocols” shows the opposite trade-off (Munari et al., 5 Jun 2025). In CTM-ET, a contention resolution interval is truncated after $\langle eos\rangle$1 slots and unresolved packets are dropped. Plain CTM has $\langle eos\rangle$2, whereas CTM-ET deliberately sacrifices delivery probability to improve freshness. For low load, $\langle eos\rangle$3 worsens AoI by more than 15% compared with plain CTM; for high load, the same setting improves AoI by roughly 30%. Early termination here is beneficial precisely because stale packets are abandoned.

The literature also contains negative or divergent uses of the same vocabulary. In instruction-tuned diffusion LLMs, early termination is not a control policy but a failure mode: increasing max_length can make outputs shorter because repeated $\langle eos\rangle$4 padding creates a position-dependent prior that propagates backward during any-order denoising (Kim et al., 4 Oct 2025). Rainbow Padding addresses that failure by decoupling true termination from padding, and the paper reports that as few as seven padding tokens are sufficient to prevent early termination. In derivatives, an optional early termination or recall clause is valued through bilateral CVA and may have non-negligible inception impact under default-free close-out; its exercise rule is

$\langle eos\rangle$5

which makes it a Bermudan-style option on future bilateral credit adjustment rather than a computational stopping criterion (Giada et al., 2012).

Finally, not every ostensibly relevant paper can be used as evidence. For arXiv (Liu et al., 16 Jun 2025), relevance to the topic is explicitly undetermined from the provided material because the supplied content does not expose the method, stopping rule, mathematical formulation, or experiments (Liu et al., 16 Jun 2025). That limitation is itself informative: “early termination by recall” is currently better treated as a cross-domain pattern than as a single settled method family.

Taken together, the cited literature suggests a precise but plural picture. Early termination by recall can mean declarative stopping at a requested recall level, early stopping followed by recovery of missed structure, truncation justified by recoverability proxies, replacement of future dynamics by a recalled continuation value, or reuse of a recalled feasible plan. What unifies these mechanisms is not one metric but one design principle: further computation is avoided once some remembered, predicted, or reconstructed surrogate is judged sufficient for the objective at hand.