Full-Match Learning Explained

• Full-Match Learning is a comprehensive paradigm that explicitly matches entire target signals, replacing partial or heuristic methods.
• It is applied across varied domains—from software vulnerability detection and stereo vision to domain adaptation and video analysis—using domain-specific matching functions.
• By learning explicit matching functions such as prototype libraries, pairwise comparisons, or gradient matching, it significantly improves performance metrics and research outcomes.

Full-Match Learning denotes a family of learning paradigms in which the central operation is not merely classification from latent features, but an explicit attempt to match an entire target structure: learned vulnerability prototypes against program statements, candidate pixels across images, source and target distributions in representation space, unlabeled samples under pseudo-label constraints, subset gradients against full-data gradients, or even the gain of full-parameter reinforcement learning through a minimal subset of trainable layers. Across the cited works, the phrase is used in several related but non-identical senses. This suggests that Full-Match Learning is best understood as a recurring methodological pattern—comprehensive, data-driven matching of a target signal—rather than a single standardized algorithm (Fu et al., 2023, Ladický et al., 2015, Yu et al., 2020, Chen et al., 2023, Killamsetty et al., 2021, Li et al., 2019, Ding et al., 26 Mar 2026, Zhang et al., 1 Jul 2026).

1. Terminological scope and recurring structure

A consistent theme across the literature is the replacement of partial, heuristic, or implicit matching by an explicitly learned matching operator applied comprehensively at training time, inference time, or both. In software security, OptiMatch learns a codebook of vulnerability patterns and matches all learned centroids against a function during inference (Fu et al., 2023). In stereo and optical flow, the matcher itself is learned as a discriminative function over contextual feature differences rather than fixed by Euclidean or histogram distances (Ladický et al., 2015). In domain adaptation, the matching loss is learned by a meta-network instead of being fixed as MMD or an adversarial discrepancy (Yu et al., 2020). In semi-supervised learning, FullMatch uses Entropy Meaning Loss and Adaptive Negative Learning so that all unlabeled examples contribute to optimization, including low-confidence cases (Chen et al., 2023).

Domain	What is matched	Operational meaning of “full-match”
Vulnerability localization	Vulnerable scopes to learned centroids	Iterate over all codebook patterns at inference
Stereo / optical flow	Candidate pixel pairs across views or time	Learn the matching function directly
Domain adaptation	Cross-domain source–target distributions	Learn the matching loss rather than fix it
Semi-supervised learning	Unlabeled data under positive and negative pseudo-labels	Exploit all unlabeled data
Data subset selection	Weighted subset gradients to full-data or validation gradients	Match the full learning signal
Search ads retrieval	Student fast matcher to annotator scores on logs	Learn deployable matching from weak annotations
Full-match video analysis	Complete rallies and match timelines	Model full matches rather than isolated clips

This breadth matters because the term can otherwise be misread as denoting only exhaustive nearest-neighbor retrieval or only prototype matching. The cited works show a broader pattern: the object being “matched” changes by domain, but the design principle remains the same—replace incomplete surrogates with a learned matching mechanism that covers the relevant signal more completely.

2. Explicit matching functions and prototype libraries

One of the clearest formulations appears in statement-level vulnerability localization. OptiMatch treats a function as a sequence of statements $X=[x_1,\dots,x_n]$, constructs statement embeddings with an RNN over BPE tokens, summarizes vulnerable scopes into vectors, and then learns a vulnerability codebook $C=\{c_k\}$ by clustering vulnerable-scope representations with Optimal Transport. During inference it performs full matching: for every centroid $c_j\in C$, it concatenates the projected centroid with statement embeddings, runs the encoder, obtains function- and statement-level probabilities, and aggregates them as

$$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$

A function is labeled benign if $P_i^{func}<\tau_{func}$ with $\tau_{func}=0.5$; otherwise statement probabilities are thresholded with $\tau_{stmt}=0.5$. The approach uses $n=155$ statements and $r=20$ tokens per statement, learns centroid selection with cross-attention, and uses a straight-through estimator through the non-differentiable argmax (Fu et al., 2023).

The significance of this design is not just architectural. It addresses the spatial variability of vulnerable scopes, which may occur in different locations and formats even for the same vulnerability type. OptiMatch makes the learned patterns explicit at inference rather than leaving them implicit inside a single discriminative boundary. This full-pattern matching is the defining property of the method.

A related but older formulation appears in dense visual correspondence. "Learning the Matching Function" treats stereo reconstruction and optical flow as binary decisions over candidate pairs $(p,q)$, builds a high-dimensional contextual descriptor

$C=\{c_k\}$0

and learns the score

$C=\{c_k\}$1

with AdaBoost decision stumps. Stereo then selects

$C=\{c_k\}$2

and optical flow analogously selects $C=\{c_k\}$3. Here “full-match” means direct learning of the matching function over contextual, multi-scale evidence rather than hand-designed distances plus a downstream regularizer (Ladický et al., 2015).

These two cases illustrate two major variants of explicit full matching. One variant uses a learned prototype library and exhaustive matching against all prototypes; the other learns the pairwise matcher itself so that the score directly represents match compatibility under realistic nuisance variation.

3. Learned matching losses, weak supervision, and exhaustive data use

In unsupervised domain adaptation, the matching problem shifts from instances to distributions. Learning to Match decomposes the predictor as $C=\{c_k\}$4 and optimizes

$C=\{c_k\}$5

where the matching loss is itself learned:

$C=\{c_k\}$6

The meta-network $C=\{c_k\}$7 is an MLP with structure $C=\{c_k\}$8-$C=\{c_k\}$9-$c_j\in C$0-$c_j\in C$1 and receives task-independent features such as embeddings or logits, optionally concatenated with human-designed features such as MMD- and adversarial-based discrepancies. Its parameters are updated by a meta-objective defined on target-domain pseudo-labeled meta-data, with high-confidence pseudo labels selected at softmax probability $c_j\in C$2 (Yu et al., 2020).

This formulation is important because many domain adaptation methods only match marginals or conditionals through fixed priors. L2M instead learns how matching should be measured. A plausible implication is that “full-match” here refers to an attempt to approximate joint distribution alignment without hard-coding the discrepancy.

A different use of the term arises in semi-supervised learning. FullMatch extends FixMatch by adding Entropy Meaning Loss and Adaptive Negative Learning. For unlabeled data, FixMatch uses a high confidence threshold $c_j\in C$3, typically $c_j\in C$4, and ignores samples with $c_j\in C$5. FullMatch keeps the standard unsupervised term

$c_j\in C$6

but augments it with

$c_j\in C$7

EML regularizes non-target probabilities for pseudo-labeled samples, while ANL assigns negative pseudo-labels to all unlabeled samples by choosing the minimal $c_j\in C$8 such that batchwise top-$c_j\in C$9 accuracy between weak and strong predictions is $$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$0 (Chen et al., 2023).

In this setting, Full-Match Learning means exhaustive use of unlabeled evidence rather than exhaustive prototype enumeration. The shift is conceptual but consistent: samples that would otherwise be discarded remain part of the matching problem.

Search advertising offers another weakly supervised interpretation. A deployable CDSSM fast matcher is trained from stronger but undeployable annotators such as Deep Crossing and Decision Tree ensembles. Human labels are enumerated, auxiliary tasks are constructed by pivoting those labels, annotators score both labeled and unlabeled query–ad pairs, the student is pretrained on scored unlabeled data, and then fine-tuned on scored labeled data with the label-aware weighted loss

$$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$1

where the asymmetry depends on the binary human label and the sign of $$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$2 through a discount factor $$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$3 (Li et al., 2019). Here full matching is tied to large-scale recall matching under severe latency constraints, but the training principle remains the same: match a deployable model to a richer supervisory signal than direct labels alone would provide.

4. Matching full training signals and full-parameter gains

A further generalization appears when the object being matched is not a sample or distribution but the optimization signal itself. GRAD-MATCH defines per-example gradients $$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$4, the full gradient $$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$5, and seeks a weighted subset $$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$6 such that

$$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$7

The paper solves this approximately with Orthogonal Matching Pursuit, gives weak-submodularity guarantees, and extends the target from training gradients to validation gradients under class imbalance or label noise (Killamsetty et al., 2021).

This is an especially strong version of the full-match idea because the matched quantity is the update direction that determines training dynamics. A subset is considered good only if its weighted gradient reproduces the behavior of the full dataset closely enough.

A related notion appears in reinforcement-learning post-training of LLMs. "Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training" defines the layer contribution of layer $$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$8 as

$$P_i^{func}=\max_{j=1..k} P_{ij}^{func}, \qquad P_i^{stmt}=\frac{1}{k}\sum_{j=1}^k P_{ij}^{stmt}.$$9

where $P_i^{func}<\tau_{func}$0 is the metric after RL updates only layer $P_i^{func}<\tau_{func}$1, and $P_i^{func}<\tau_{func}$2 is the metric after full-parameter RL. In this usage, Full-Match Learning means that a minimal update set—sometimes a single layer—matches the gain of full-parameter RL. Across seven models, high-contribution layers concentrate in the middle of the transformer stack, and guided strategies that train only the top-$P_i^{func}<\tau_{func}$3 or middle layers often exceed full-parameter RL (Zhang et al., 1 Jul 2026).

The conceptual continuity with GRAD-MATCH is striking. In one case a subset of data is selected so that its gradients match the full-data gradient; in the other, a subset of parameters is selected so that its trainable subspace matches the gain of full-parameter training. Both are forms of signal compression under a full-match criterion.

5. Full-match as whole-sequence or whole-event modeling

In sequential domains, “full-match” can refer to the temporal extent of the modeled object. BFMD introduces the first Badminton Full Match Dense dataset, with 19 broadcast matches covering 20.32 hours of play, 1,687 rallies, and 16,751 hit events, each annotated with a shot caption. The dataset includes hierarchical annotations for match segments, rally events, shot types, shuttle trajectories, player pose keypoints, and captions. The captioning model uses VideoMAE-base, a Token Refiner, multimodal fusion over positions, pose, and shuttle trajectories, and a Semantic Feedback mechanism that predicts shot category, trajectory intensity, court region, and tactical intent, trained with

$P_i^{func}<\tau_{func}$4

The operational meaning of full-match learning here is analysis of complete broadcasts rather than isolated clips (Ding et al., 26 Mar 2026).

The full-match framing matters because tactical interpretation depends on rally continuity, interruptions such as replays and Hawk-Eye segments, and temporal evolution across the match. The BFMD paper explicitly uses sliding-window tactical pattern detection over complete matches and smooths occurrences over time to obtain intensity curves.

A looser but related use appears in cricket analytics, where the literature distinguishes pre-match full-match prediction from intra-match sequence prediction. Team-level models use features such as ICC points, home advantage, venue average score, day/night conditions, player statistics, participation ratios, and AHP-derived attributes such as consistency, recent form, and venue points. Some pipelines first predict batsmen and bowlers separately and then aggregate outputs to team batting and bowling scores before predicting the match winner; other approaches update score or win probability over 5-over blocks (Mittal et al., 2021). In this literature, Full-Match Learning refers to the match as the prediction unit rather than to a particular matching algorithm.

This temporal interpretation broadens the term further. Matching can be exhaustive not only over prototypes or unlabeled examples, but also over the temporal span of the event being modeled.

6. Empirical behavior, misconceptions, and open problems

The empirical record is heterogeneous but consistently favorable within each formulation. On Big-Vul, OptiMatch was evaluated on 188k C/C++ functions with 3,754 vulnerabilities across 91 CWE types and 348 GitHub projects, achieving function-level $P_i^{func}<\tau_{func}$5 and statement-level $P_i^{func}<\tau_{func}$6, corresponding to gains of $P_i^{func}<\tau_{func}$7 and $P_i^{func}<\tau_{func}$8 over the previous best. Removing the vulnerability codebook and matching drops function $P_i^{func}<\tau_{func}$9 from $\tau_{func}=0.5$0 to $\tau_{func}=0.5$1 and statement $\tau_{func}=0.5$2 from $\tau_{func}=0.5$3 to $\tau_{func}=0.5$4, indicating that the full-match step is not incidental (Fu et al., 2023).

In dense correspondence, the learned matcher on KITTI reported $\tau_{func}=0.5$5 error at the 3-pixel threshold and $\tau_{func}=0.5$6 at the 5-pixel threshold, while on the introduced TimeLapse change detection dataset it reported $\tau_{func}=0.5$7 pixel accuracy, average recall per class $\tau_{func}=0.5$8, and average precision $\tau_{func}=0.5$9 (Ladický et al., 2015). In domain adaptation, L2M reported average accuracies of $\tau_{stmt}=0.5$0 on Office-Home, $\tau_{stmt}=0.5$1 on ImageCLEF-DA, $\tau_{stmt}=0.5$2 on VisDA-2017, and $\tau_{stmt}=0.5$3 on Office-31; on pneumonia-to-COVID chest X-ray transfer it reported Precision $\tau_{stmt}=0.5$4, Recall $\tau_{stmt}=0.5$5, and $\tau_{stmt}=0.5$6 (Yu et al., 2020).

In semi-supervised learning, FullMatch on CIFAR-100 with 10k labels selected about $\tau_{stmt}=0.5$7 of unlabeled examples by 200k iterations and about $\tau_{stmt}=0.5$8 by the end under the FixMatch mask ratio, while EML increased the fraction of low-entropy predictions with entropy $\tau_{stmt}=0.5$9 from about $n=155$0 to about $n=155$1. On CIFAR-100 with 400 labels, the combined FullMatch formulation improved top-1 by $n=155$2 over baseline, and on ImageNet with 100 labels per class, FullMatch and FullFlex improved top-1 by more than $n=155$3 over their corresponding baselines (Chen et al., 2023). In Search Ads, the weak-annotation framework improved the fast matcher by $n=155$4 ROC AUC and $n=155$5 PR AUC over the labeled-only baseline, while dispensing with $n=155$6 labeled samples (Li et al., 2019). GRAD-MATCH reported, for CIFAR-10 with ResNet-18, about $n=155$7 speedup at the 30% subset with about $n=155$8 accuracy drop, about $n=155$9 speedup at the 20% subset with about $r=20$0 accuracy drop, and about $r=20$1 speedup at the 10% subset with about $r=20$2 accuracy drop (Killamsetty et al., 2021). BFMD’s full multimodal captioning model reported BLEU-4 $r=20$3, METEOR $r=20$4, ROUGE-L $r=20$5, and CIDEr $r=20$6, improving over the RGB-only variant (Ding et al., 26 Mar 2026). In RL post-training, Qwen3-1.7B-Base reached best single-layer $r=20$7 at Layer 10, and 5 of 28 layers achieved $r=20$8; Qwen3-4B-Base and Qwen3-8B-Base both reached their best values at Layer 16, with guided layer-selective strategies exceeding full-parameter RL (Zhang et al., 1 Jul 2026).

Several misconceptions follow naturally from the diversity of these results. First, Full-Match Learning is not a single algorithm. It may denote prototype matching, learned pairwise matching, learned distribution discrepancy, exhaustive pseudo-label utilization, gradient matching, weak-annotation distillation, or whole-match temporal modeling. Second, it is not synonymous with brute-force search. Some methods do iterate over all centroids or candidates, but others learn a match loss, a subset, or a trainable layer whose effect matches a fuller target. Third, it does not guarantee robustness. The cited limitations include obfuscated code and rare vulnerability types in OptiMatch, memory stress under large flow ranges in learned optical flow, noisy pseudo labels or severe domain shift in L2M, class imbalance and calibration issues in FullMatch-style SSL and sports prediction, and modest dataset scale in BFMD (Fu et al., 2023, Ladický et al., 2015, Yu et al., 2020, Chen et al., 2023, Ding et al., 26 Mar 2026).

Open directions are correspondingly domain-specific. The vulnerability-localization work proposes adaptive codebook updates, automated selection of codebook size $r=20$9, multimodal program signals, and approximate nearest-neighbor search over centroids (Fu et al., 2023). L2M suggests broader learned matching losses for other distribution-matching problems (Yu et al., 2020). BFMD points toward longer temporal models, calibrated court coordinates, and semi-supervised use of unlabeled broadcasts (Ding et al., 26 Mar 2026). The RL work points toward dynamic layer selection and subcomponent ablations within high-contribution layers (Zhang et al., 1 Jul 2026). The common research question is whether the target of interest—pattern library, discrepancy, supervisory signal, optimization trajectory, or temporal event structure—can be represented and matched more directly than with the partial surrogates used by earlier methods.