Analytic Classifier Reconstruction (ACR)

• Analytic Classifier Reconstruction (ACR) is a family of techniques that analytically recovers classifiers from compact structures (e.g., SVD-based subspaces, ridge regression solutions) without end-to-end retraining.
• ACR methods are applied in out-of-distribution detection, continual learning, and open-set recognition, utilizing explicit mathematical operations to derive decision-relevant behavior.
• These approaches enable efficient, explainable model updates by replacing traditional exemplar-based or black-box retraining with compact, analytically recoverable decision objects.

Searching arXiv for the cited papers and closely related ACR usages. Analytic Classifier Reconstruction (ACR) denotes a family of analytically specified procedures that reconstruct, reinterpret, or operationalize a classifier from compact structure rather than by end-to-end retraining. In recent arXiv literature, the phrase appears in several distinct technical senses: as a classifier-weight-induced feature reconstruction method for post-hoc out-of-distribution detection in ClaFR (Wang et al., 2 Sep 2025); as the classifier-side closed-form component of continual learning systems such as ADR and DPCR (Song et al., 3 Apr 2026, He et al., 7 Mar 2025); as an ACR-style interpretation of joint classification–reconstruction learning for open-set recognition in CROSR (Yoshihashi et al., 2018); and, in a different acronymal context, as the reconstruction of the ACR Appropriateness Criteria into an operational classifier over free-text clinical narratives (Pambudi et al., 6 Oct 2025). The common thread is the use of explicit analytic structure—most notably SVD, ridge regression, or deterministic guideline lookup—to derive decision-relevant behavior from a trained model or a codified rule system.

1. Scope of the term and recurring analytic motif

Across these works, ACR is not attached to a single canonical algorithm. Instead, it names several reconstruction paradigms in which the classifier is recovered from a compact analytic object. In ClaFR, that object is the final linear classifier weight matrix $W$, whose singular vectors define a “class-known” subspace (Wang et al., 2 Sep 2025). In ADR and DPCR, the reconstructed classifier is the closed-form solution of a ridge regression objective expressed through accumulated feature–feature and feature–label statistics, without raw old data (Song et al., 3 Apr 2026, He et al., 7 Mar 2025). In CROSR, an ACR interpretation arises because the classifier is coupled to a reconstruction-capable latent representation $[y,z]$, and open-set decisions are made by analytic modeling in that joint space (Yoshihashi et al., 2018). In the medical-imaging system, the classifier is reconstructed from an expert-authored guideline hierarchy: free text is mapped to a variant, and the final recommendation is obtained by deterministic lookup of the “Usually Appropriate” procedures (Pambudi et al., 6 Oct 2025).

This suggests a broad but technically coherent pattern. ACR methods replace direct reliance on stored exemplars, synthetic unknowns, or unconstrained parametric memory with analytically recoverable decision objects. Depending on the setting, those objects are subspaces, correlation matrices, latent manifolds, or structured guideline variants.

2. Classifier-weight reconstruction for out-of-distribution detection

The most explicit instantiation of ACR appears in "Frustratingly Easy Feature Reconstruction for Out-of-Distribution Detection" (Wang et al., 2 Sep 2025). ClaFR assumes a pretrained classifier

$g(x)=\operatorname{softmax}(W^\top f(x)),$

where $z=f(x)\in\mathbb{R}^D$ is the penultimate-layer feature and $W\in\mathbb{R}^{D\times C}$ is the final linear classifier weight matrix. The method uses only the classifier weights, with no training data, to infer a “class-known” subspace in feature space.

The key step is an SVD of the classifier: $W = U\Sigma V^\top,$ with $U_M=[u_1,\dots,u_m]$ formed from the top-$m$ left singular vectors. The subspace dimension is chosen by the explained-variance criterion

$$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$

with a typical setting $\alpha \approx 0.9$. The analytic projection operator is

$[y,z]$0

and the reconstructed feature is $[y,z]$1.

ClaFR defines reconstruction error as

$[y,z]$2

but uses the equivalent monotone score

$[y,z]$3

Large $[y,z]$4 implies small reconstruction error and more likely in-distribution behavior; small $[y,z]$5 implies a large residual in the orthogonal complement and more likely OOD behavior. Thresholding is performed with a value $[y,z]$6, for example chosen to achieve $[y,z]$7 TPR on ID validation data.

The procedure is entirely post-hoc. It requires white-box access to the feature extractor and final-layer weights, but no access to training data. This distinguishes it from Mahalanobis, KNN, ViM, Neco, and PCA-based reconstruction methods, all of which rely on training features or stored exemplars. ClaFR therefore treats the classifier weights as a compressed summary of training feature geometry and uses them to define a training-data-free analytic reconstruction operator.

Empirically, ClaFR reports leading performance on multiple OOD benchmarks (Wang et al., 2 Sep 2025). On ImageNet-1K with a ResNet-50 backbone, the average AUROC is $[y,z]$8 and the average FPR@95 is $[y,z]$9, outperforming the second-best average by $g(x)=\operatorname{softmax}(W^\top f(x)),$0 AUROC and $g(x)=\operatorname{softmax}(W^\top f(x)),$1 FPR@95. With a MobileNet backbone, the average AUROC is $g(x)=\operatorname{softmax}(W^\top f(x)),$2 and the average FPR@95 is $g(x)=\operatorname{softmax}(W^\top f(x)),$3. The method is also computationally lightweight: on ImageNet with an RTX 3090, KNN inference per sample is approximately $g(x)=\operatorname{softmax}(W^\top f(x)),$4 ms with memory approximately $g(x)=\operatorname{softmax}(W^\top f(x)),$5 GiB, whereas ClaFR inference per sample is approximately $g(x)=\operatorname{softmax}(W^\top f(x)),$6 ms with memory approximately $g(x)=\operatorname{softmax}(W^\top f(x)),$7 MiB. Ablations show stable performance for $g(x)=\operatorname{softmax}(W^\top f(x)),$8; too small $g(x)=\operatorname{softmax}(W^\top f(x)),$9 discards genuine ID structure, and too large $z=f(x)\in\mathbb{R}^D$0 retains noisy directions.

3. Closed-form classifier rebuilding in continual learning

In continual learning, ACR refers to analytic reconstruction of the classifier after the representation changes across tasks. Two recent formulations are especially explicit: ADR for non-exemplar continual graph learning and DPCR for exemplar-free class-incremental learning (Song et al., 3 Apr 2026, He et al., 7 Mar 2025).

In ADR, ACR is the classifier-side component of the Analytic Drift Resister framework. After a GNN encoder is adapted to the current task by backpropagation and then merged across tasks by Hierarchical Analytic Merging (HAM), the task-specific classifier is replaced by a new classifier $z=f(x)\in\mathbb{R}^D$1 obtained from joint ridge regression over all tasks: $z=f(x)\in\mathbb{R}^D$2 Using cumulative statistics

$z=f(x)\in\mathbb{R}^D$3

the solution becomes

$z=f(x)\in\mathbb{R}^D$4

No raw node features or adjacency matrices need to be stored; only accumulated statistics are retained. An optional random feature expansion $z=f(x)\in\mathbb{R}^D$5 enlarges linear separability, with classifier-side storage $z=f(x)\in\mathbb{R}^D$6.

The central theoretical claim in ADR is conditional. Under HAM, the merged encoder is equivalent to an encoder obtained by joint training over all tasks with a ridge objective on each layer, yielding “absolute resistance to feature drift.” Under that condition, ACR is mathematically equivalent to joint ridge regression over all tasks and thereby provides theoretically zero-forgetting class-incremental learning (Song et al., 3 Apr 2026). Empirically, ADR significantly outperforms EFC/POLO on all four benchmarks in both average and final accuracy, and the ablation “w/o HAM” causes drastic performance drops, including CS-CL $z=f(x)\in\mathbb{R}^D$7, illustrating that analytic classifier reconstruction depends on a drift-free or drift-corrected feature space.

DPCR adopts the same analytic ridge-regression core but adds explicit semantic-shift estimation. For the hypothetical all-data objective

$z=f(x)\in\mathbb{R}^D$8

the closed-form solution is

$z=f(x)\in\mathbb{R}^D$9

where $W\in\mathbb{R}^{D\times C}$0 is the uncentered covariance and

$W\in\mathbb{R}^{D\times C}$1

is the feature–label correlation derived from prototypes. Because old exemplars are unavailable, DPCR calibrates stored covariances and prototypes into the current feature space by a dual projection. Task-wise Semantic Shift Projection solves

$W\in\mathbb{R}^{D\times C}$2

and Category Information Projection restricts this global shift to class-specific row spaces: $W\in\mathbb{R}^{D\times C}$3 Old statistics are then updated by

$W\in\mathbb{R}^{D\times C}$4

The final reconstructed classifier is

$W\in\mathbb{R}^{D\times C}$5

followed by category-wise normalization of each class weight vector. DPCR stores $W\in\mathbb{R}^{D\times C}$6 parameters per class, satisfies the exemplar-free constraint, and reports that RRCR alone gives $W\in\mathbb{R}^{D\times C}$7 and $W\in\mathbb{R}^{D\times C}$8 on CIFAR-100 with $W\in\mathbb{R}^{D\times C}$9; adding TSSP, CIP, and category-wise normalization improves performance progressively (He et al., 7 Mar 2025). The method is explicitly positioned as an ACR approach that keeps the backbone plastic while analytically reconstructing a less biased classifier.

4. Classification–reconstruction learning as an ACR-style precursor

A precursor formulation appears in "Classification-Reconstruction Learning for Open-Set Recognition" (Yoshihashi et al., 2018). CROSR does not use the phrase ACR as its primary name, but it fits an ACR interpretation because the classifier is embedded in a reconstruction-capable architecture and open-set decisions are made by analytic modeling of the resulting latent space.

CROSR trains an encoder $W = U\Sigma V^\top,$0 and decoder $W = U\Sigma V^\top,$1 jointly: $W = U\Sigma V^\top,$2 with loss

$W = U\Sigma V^\top,$3

The encoder is a Deep Hierarchical Reconstruction Net (DHRNet) in which multiple compact bottlenecks $W = U\Sigma V^\top,$4 are extracted from intermediate layers and concatenated into $W = U\Sigma V^\top,$5. The classifier uses the top-level feature to produce logits $W = U\Sigma V^\top,$6, while the decoder reconstructs the input. At test time, only the encoder is used; the decoder is discarded.

Unknown detection is performed by analytic modeling in the joint space $W = U\Sigma V^\top,$7. For each known class $W = U\Sigma V^\top,$8, CROSR computes a class mean

$W = U\Sigma V^\top,$9

a distance

$U_M=[u_1,\dots,u_m]$0

and a Weibull tail model that yields a class-belongingness probability. Openmax-style recalibration then introduces an explicit unknown class. Within this interpretation, reconstruction is not merely an auxiliary loss: it creates a compact, multi-scale latent space that supports analytic open-set modeling more effectively than logits alone.

The empirical results support that interpretation (Yoshihashi et al., 2018). On MNIST with Omniglot outliers, Plain CNN + Openmax attains $U_M=[u_1,\dots,u_m]$1 macro-F1, whereas DHRNet + CROSR reaches $U_M=[u_1,\dots,u_m]$2. On MNIST-noise, the scores are $U_M=[u_1,\dots,u_m]$3 and $U_M=[u_1,\dots,u_m]$4, respectively. On CIFAR-10 with natural-image outliers, DHRNet + CROSR reports approximately $U_M=[u_1,\dots,u_m]$5–$U_M=[u_1,\dots,u_m]$6 F1 across all four outlier sets, and DenseNet + DHRNet + CROSR reaches up to $U_M=[u_1,\dots,u_m]$7 on ImageNet-resize. On the AUROC comparison against GAN-based methods, CROSR reports $U_M=[u_1,\dots,u_m]$8 on MNIST and $U_M=[u_1,\dots,u_m]$9 on TinyImageNet. The paper also shows that generic one-class detectors such as OCSVM and Isolation Forest can be applied directly to $m$0, indicating that the learned latent space possesses analytic structure beyond a single detector choice.

5. Guideline-grounded reconstruction of a decision system

A different but instructive usage appears in "Bridging Clinical Narratives and ACR Appropriateness Guidelines: A Multi-Agent RAG System for Medical Imaging Decisions" (Pambudi et al., 6 Oct 2025). Here ACR simultaneously evokes the American College of Radiology Appropriateness Criteria and the idea of reconstructing an expert-designed decision system as an operational classifier over unstructured text.

The system treats each ACR variant as a label in a structured decision space. Free-text clinical scenarios are mapped to a single variant, and that variant is mapped deterministically to the set of imaging procedures rated “Usually Appropriate.” The pipeline has three stages: a ColBERT retrieval agent returning the top-10 candidate variants; an LLM selector agent choosing one best-matching variant; and a supervisor/output module that looks up the corresponding procedures. The retrieval model is fine-tuned on $m$1 synthetic clinical scenario–recommendation pairs derived from $m$2 unique variants, with 3 semantically graduated notes and 5 synonym-enriched notes per variant. ColBERT scoring uses late interaction: $m$3

This formulation is analytic in a different sense from ClaFR or ridge-based continual learning. The rule system itself is fixed and non-parametric: the mapping from selected variant to procedures is given by the guideline database, not learned by the LLM. Hallucination is reduced because the selector is constrained to choose among retrieved candidates, and the final procedure set is a deterministic lookup from the ACR dataset.

The reported retrieval and end-to-end results are strong (Pambudi et al., 6 Oct 2025). Fine-tuned ColBERT achieves Recall@10 $m$4, Recall@5 $m$5, and Recall@1 $m$6, compared with BM25 Recall@10 $m$7 and pre-trained ColBERT Recall@10 $m$8. The GPT-4.1 RAG system reaches Exact Match $m$9 and F1 $$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$0, whereas GPT-4.1 without retrieval gives Exact Match $$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$1 and F1 $$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$2. The test set is deliberately challenging, with Jaccard similarity $$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$3 between guideline variant titles and synthetic scenarios. In this setting, ACR is the reconstruction of a structured expert policy into a classifier whose learned component is limited to mapping unstructured language onto the predefined decision space.

6. Distinctions, limitations, and conceptual synthesis

The literature surveyed here suggests that ACR is best understood as an analytic design pattern rather than a single standardized method. In OOD detection, it means reconstructing features from a classifier-induced subspace obtained directly from $$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$4 (Wang et al., 2 Sep 2025). In continual learning, it means rebuilding the classifier by closed-form ridge regression from sufficient statistics, sometimes under a zero-forgetting guarantee when the encoder-side conditions are met (Song et al., 3 Apr 2026, He et al., 7 Mar 2025). In open-set recognition, it denotes a joint classification–reconstruction architecture whose latent space is explicitly analyzable (Yoshihashi et al., 2018). In clinical decision support, it denotes reconstruction of a codified guideline hierarchy as a grounded classifier over free text (Pambudi et al., 6 Oct 2025).

Each formulation carries specific constraints. ClaFR assumes full white-box access to the feature extractor and final-layer weights, and its performance depends on the spectral truncation parameter $$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$5 (Wang et al., 2 Sep 2025). ADR’s zero-forgetting guarantee is conditional on HAM and on the ridge-regression formulation; without HAM, performance can collapse sharply (Song et al., 3 Apr 2026). DPCR reduces decision bias but stores a $$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$6 covariance per class, assumes a linear classifier, and is sensitive to the ridge parameter $$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$7 (He et al., 7 Mar 2025). CROSR relies on EVT hyperparameters such as tail_size and $$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$8, does not provide a full probabilistic generative model $$\sum_{i=1}^{m}\sigma_i > \alpha \sum_{i=1}^{D}\sigma_i,$$9, and faces scaling challenges as the number of classes grows (Yoshihashi et al., 2018). The medical-imaging system is evaluated on synthetic scenarios, is optimized for short input formats, and is explicitly positioned for human-in-the-loop deployment rather than autonomous use (Pambudi et al., 6 Oct 2025).

A plausible unifying characterization is that ACR replaces replay-based or purely black-box classifier updates with analytically reconstructible decision objects. Whether the object is a discriminative subspace, a ridge-regression solution, a reconstruction-aware latent manifold, or a guideline variant indexed by dense retrieval, the defining property is the same: the classifier’s operative structure is made explicit, compact, and recoverable by analytic means.