Pairwise Progress Supervision

• Pairwise progress supervision is a paradigm that replaces absolute labels with comparative signals, streamlining annotation and directly encoding task progression.
• It employs graded scoring, binary judgments, and structured correspondence to enhance learning in applications like translation evaluation, robot control, and medical imaging.
• Empirical studies show that integrating pairwise comparisons with absolute anchors improves performance, though its benefits depend on task-specific geometry and calibration needs.

Pairwise progress supervision can be understood as a family of learning regimes in which supervision is supplied as a relation between two objects—translations, trajectory states, labels, datasets, images, graph nodes, or lesion candidates—rather than as an absolute annotation on a single object. In the recent literature, this pattern appears as graded quality differences in reference-free machine translation evaluation, retry-centered local preference learning for robot imitation, signed pairwise constraints for featureless graph representation learning, pairwise label comparisons in active learning, weakly supervised binary classification from similarity/dissimilarity and pairwise comparison labels, pairwise prior-order cues in multiple-unlabeled-dataset learning, pairwise comparisons for conditional image editing, pairwise comparison data for uncoupled regression, and explicit pairwise lesion correspondence supervision in mammography (Proietti et al., 25 Jan 2026, Qin et al., 23 Jun 2026, Chakraborty et al., 19 May 2026, Yona et al., 2022, Tate et al., 20 Mar 2026, Wu et al., 2023, Han et al., 2019, Xu et al., 2019, Zhao et al., 2022). Across these settings, the common structural move is to replace or augment absolute supervision with comparative information that directly specifies direction, relative order, or correspondence.

1. Formalization and scope

A recurrent formal contrast is between single-item prediction and pairwise prediction. In machine translation quality estimation, a standard single-candidate metric is written as $\hat{s}(s,mt)=g_\theta(s,mt)$, whereas PEAR reframes the task as a pairwise comparison over a source segment $s$ and two candidate translations $mt_a,mt_b$, predicting a real-valued relative score

$$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$

with sign and magnitude jointly encoding which candidate is better and by how much (Proietti et al., 25 Jan 2026). In multiclass active learning, the analogous distinction is between an argmax oracle,

$$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$

and a comparison oracle,

$$A_{\mathrm{comparisons}}^{f^\star}(x,j_1,j_2)=\mathbf 1\!\left[f^\star_{j_1}(x)>f^\star_{j_2}(x)\right],$$

which reveals a relative judgment over two labels rather than the full label identity (Yona et al., 2022).

This suggests three recurring forms of pairwise progress supervision.

Form	Representative signal	Representative paper
Graded relative scoring	Human-derived quality difference $\Delta^\star_{ab}$	PEAR (Proietti et al., 25 Jan 2026)
Binary relational judgment	Similar/different, better/worse, higher/lower prior	SD-Pcomp (Tate et al., 20 Mar 2026), MU-OPPO (Wu et al., 2023)
Structured correspondence	Cross-view lesion pairing with explicit matching supervision	CL-Net (Zhao et al., 2022)

The phrase “progress” is most literal when the relation encodes advancement or degradation along a task trajectory, as in ReTVL, where retry events induce local orderings that capture a degradation-and-recovery structure around mistakes (Qin et al., 23 Jun 2026). More broadly, the same mathematical pattern extends to settings where the relation is comparative rather than temporal: pairwise label comparisons, pairwise score differences, signed similarity constraints, prior-order relations, or pairwise correspondence.

2. Supervision sources and annotation regimes

A defining characteristic of the paradigm is that pairwise targets are often easier to elicit, cheaper to annotate, or more faithful to the underlying decision problem than direct absolute labels. In PEAR, the training target is not a human score itself but a difference of human segment-level scores,

$\Delta^\star_{ab}=s_h(s,mt_a)-s_h(s,mt_b),$

derived from DA, DA+SQM, or MQM annotations. The model therefore learns human comparative information even though the raw annotations begin as absolute segment-level judgments (Proietti et al., 25 Jan 2026).

ReTVL makes the annotation economy explicit. A retry keypoint $r_{i,j}$ is defined as “the start of the $j$-th corrective behavior in trajectory $s$0,” and only that moment is annotated; the method does not annotate the exact start of deviation and does not annotate completion of recovery. The paper reports that retry keypoints are sparse, temporally consistent, and fast to annotate, with average annotation time about 42.7 seconds per trajectory and average inter-annotator error 0.19 seconds (Qin et al., 23 Jun 2026). This yields a sparse supervision anchor from which many local pairwise comparisons can be constructed.

The same economy appears in label-space supervision. In multiclass active learning, the motivation is that when there are many classes, asking a human to identify the single best label can be cognitively expensive, whereas comparing two candidate labels is often much easier; the paper gives summary preference as an example where choosing the best summary among many may be hard but choosing between two candidate summaries is natural (Yona et al., 2022). In PC-GAN, absolute attribute scores are replaced by pairwise judgments such as $s$1 or $s$2, motivated by the claim that pairwise labels are easier than exact intensities and more accurate for subjective attributes such as age or attractiveness (Han et al., 2019).

Pairwise supervision can also be privacy-preserving or nearly cost-free. In uncoupled regression, the pairwise signal only reveals which of two unlabeled examples has the larger target value, which the paper argues does not break anonymity for sensitive targets such as income (Xu et al., 2019). In MU-OPPO, the only auxiliary information is a single pairwise numerical relationship of class priors, such as $s$3, rather than exact class priors for all unlabeled datasets; this is presented as “almost free” supervision because relative prior order is often easier to obtain than exact rates (Wu et al., 2023).

3. Core modeling patterns and objective functions

A central modeling choice is whether the learner predicts relative quantities directly or derives them by subtraction from absolute predictions. PEAR is explicit on this distinction. Given a source segment and two candidate translations, it uses a cross-encoder over

$s$4

builds source-aware candidate features, assigns each candidate a scalar utility $s$5, and models the comparison by subtraction $s$6, followed by a learned positive scale:

$s$7

Training combines regression to the human pairwise difference with an antisymmetry regularizer,

$s$8

thereby encouraging sign inversion under candidate swapping (Proietti et al., 25 Jan 2026).

ReTVL combines global absolute calibration with local pairwise preference learning. Outside retry neighborhoods, successful trajectories use the coarse progress target $s$9, terminal frames of failed trajectories use $mt_a,mt_b$0, and the value model predicts a distribution over progress bins with scalar value

$mt_a,mt_b$1

Around retry keypoints, the model defines pre-retry, retry-centered, and post-recovery windows and trains a pairwise logistic ranking loss

$mt_a,mt_b$2

with soft-window weighting

$mt_a,mt_b$3

where the appendix gives $mt_a,mt_b$4 and $mt_a,mt_b$5 (Qin et al., 23 Jun 2026). The objective is therefore hybrid rather than purely pairwise:

$mt_a,mt_b$6

In featureless graph learning, Contrastive FUSE uses partial pairwise labels encoded in a sparse signed matrix $mt_a,mt_b$7, where $mt_a,mt_b$8 denotes a positive pair, $mt_a,mt_b$9 a negative pair, and $$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$0 an unlabeled pair. Its unified spectral objective fuses graph modularity with a signed normalized Laplacian:

$$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$1

with direct projected gradient ascent and row-wise normalization of embeddings (Chakraborty et al., 19 May 2026). Pairwise supervision here does not encode temporal progress, but it still provides a comparative signal that pulls positive pairs together and pushes negative pairs apart.

Weakly supervised binary classification exhibits a second major pattern: pairwise supervision can be sufficient to reconstruct an unbiased estimator of ordinary classification risk. SD-Pcomp uses two pairwise label types—Similarity/Dissimilarity (SD) and Pairwise Comparison (Pcomp)—and proposes both a convex combination

$$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$2

and a unified unbiased estimator that models the relationship between the two label types (Tate et al., 20 Mar 2026). The essential point is that training examples are pairs, not individually labeled instances.

CL-Net instantiates pairwise supervision as explicit correspondence learning. A View-Interactive Lesion Detector allows cross-view interaction between CC and MLO candidate embeddings, while a Lesion Linker appends a dustbin embedding, uses link queries to form pairwise hypotheses, and applies Hungarian matching over predicted and ground-truth correspondence triplets under the joint loss

$$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$3

This is pairwise supervision in the form of lesion-to-lesion linkage rather than comparative ranking (Zhao et al., 2022).

4. Empirical roles across application domains

The empirical role of pairwise progress supervision depends on what the pairwise relation is meant to control. In machine translation evaluation, PEAR isolates the benefit of the pairwise formulation by comparing against strictly matched single-candidate QE baselines trained with the same data and backbones. On the WMT24 meta-evaluation benchmark, PEAR outperforms those matched baselines; PEAR-XL (3.5B) exceeds MetricX-24-Hybrid-QE-XL (3.7B) and XCOMET-QE (24B) in Avg Corr; PEAR (560M) also beats those larger models; with distilled supervision, PEAR even edges out CometKiwi-XXL while using far fewer parameters. The paper also reports that PEAR has lower segment-level pairwise-score correlation with other top metrics than those metrics have with each other, and that its approximate antisymmetry allows an MBR shortcut that computes only one triangle of the utility matrix and roughly halves the number of forward passes with nearly identical results (Proietti et al., 25 Jan 2026).

In robot imitation, ReTVL uses retry-supervised values to separate harmful local mistakes from useful recovery behavior. On four real-robot manipulation tasks, it reports competitive global calibration (VOC: 0.987, S/F Detection: 1.000) together with the best local retry metrics (Drop AUC: 0.797, Drop Prob.: 0.874, Pre $$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$4 Retry: 0.740, Post $$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$5 Retry: 0.967). When the learned value model is used to reweight chunks for behavior cloning, average success rises from 41.25% for Standard BC and 62.50% for RECAP-BC to 80.00% for ReTVL-BC, while average bad-action weight drops to 0.11 versus 1.0 for Standard BC and 0.54 for RECAP-BC (Qin et al., 23 Jun 2026).

In mammography, explicit pairwise correspondence supervision sharpens cross-view reasoning. On DDSM, CL-Net reaches $$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$6, $$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$7, $$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$8, $$\hat{\Delta}_{ab}=f_\theta(s,mt_a,mt_b)\in\mathbb{R},$$9, and $$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$0, and the paper emphasizes a 3.6-point gain over BG-RCNN at $$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$1 in the low-FPI regime. Ablation further shows Baseline Deformable DETR at $$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$2, VILD only at $$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$3, Lesion Linker only at $$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$4, and VILD + Lesion Linker at $$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$5, indicating that explicit correspondence supervision contributes beyond cross-view interaction alone (Zhao et al., 2022).

In featureless graph representation learning, partial pairwise supervision acts as a semantic correction to pure topology. Contrastive FUSE reports competitive or superior contrastive classification performance without relying on node features, remains effective on OGBN-ArXiv and OGBN-Products, and obtains large runtime gains from its approximate modularity gradient, with appendix measurements showing cosine similarity around 0.995 at initialization and speedups roughly 23×–40× over exact modularity computation (Chakraborty et al., 19 May 2026).

Other application areas show the same comparative pattern. PC-GAN learns latent continuous attribute ratings from pairwise comparisons, models uncertainty with a Bayesian Elo-style network, and reports performance comparable to fully supervised methods while outperforming unsupervised baselines on Annotated MNIST, CACD, UTKFace, SCUT-FBP, and CelebA (Han et al., 2019). MU-OPPO uses only one prior-order relation to bootstrap pseudo labeling, confident-example collection, class-prior estimation, and final classifier training; empirically, its estimated priors are accurate and downstream classification accuracy is close to the oracle-prior setting (Wu et al., 2023).

5. Theoretical guarantees and limitations

A major theme in the theory is that pairwise supervision is powerful but not uniformly superior. In active learning with label comparisons, the naive way to simulate an argmax label is a $$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$6 comparison tournament, but the paper proves that this can be significantly suboptimal. For 1D linear multiclass classifiers, comparison query complexity can be

$$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$7

whereas simulating the best argmax-based active learner would cost

$$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$8

The gain arises from the label neighborhood graph $$A_{\argmax}^{f^\star}(x)=\arg\max_{i\in[k]} f_i^\star(x),$$9, whose edges connect labels that share a decision boundary; if the graph is sparse, only neighboring labels need to be compared (Yona et al., 2022).

The same paper establishes a negative result in passive PAC learning. For multiclass linear classifiers in 1D,

$$A_{\mathrm{comparisons}}^{f^\star}(x,j_1,j_2)=\mathbf 1\!\left[f^\star_{j_1}(x)>f^\star_{j_2}(x)\right],$$0

samples are necessary whether the learner gets argmax labels or full comparison information, and more generally any PAC learner with label comparisons needs $$A_{\mathrm{comparisons}}^{f^\star}(x,j_1,j_2)=\mathbf 1\!\left[f^\star_{j_1}(x)>f^\star_{j_2}(x)\right],$$1 samples in the worst case for a class with Daniely–Shalev-Shwartz dimension $$A_{\mathrm{comparisons}}^{f^\star}(x,j_1,j_2)=\mathbf 1\!\left[f^\star_{j_1}(x)>f^\star_{j_2}(x)\right],$$2 (Yona et al., 2022). Pairwise comparisons therefore do not improve worst-case sample complexity in general; their value is geometry-dependent and especially strong in active learning.

Uncoupled regression from comparison data reaches a similar conclusion. The Risk Approximation method is consistent when the target variable is uniformly distributed, and under that condition the learned model converges to the optimal model with the optimal parametric rate. But Theorem 3 gives an impossibility result: two different joint distributions can share the same $$A_{\mathrm{comparisons}}^{f^\star}(x,j_1,j_2)=\mathbf 1\!\left[f^\star_{j_1}(x)>f^\star_{j_2}(x)\right],$$3, the same $$A_{\mathrm{comparisons}}^{f^\star}(x,j_1,j_2)=\mathbf 1\!\left[f^\star_{j_1}(x)>f^\star_{j_2}(x)\right],$$4, and the same pairwise-comparison distribution $$A_{\mathrm{comparisons}}^{f^\star}(x,j_1,j_2)=\mathbf 1\!\left[f^\star_{j_1}(x)>f^\star_{j_2}(x)\right],$$5 while having different conditional expectations $$A_{\mathrm{comparisons}}^{f^\star}(x,j_1,j_2)=\mathbf 1\!\left[f^\star_{j_1}(x)>f^\star_{j_2}(x)\right],$$6 (Xu et al., 2019). Pairwise supervision is therefore informative but not fully identifying in the general case.

In weakly supervised classification, the emphasis is on exact risk reconstruction and robustness. SD-Pcomp proves that both the convex combination estimator and the unified estimator are unbiased for the ordinary classification risk, gives a generalization bound showing statistical consistency as $$A_{\mathrm{comparisons}}^{f^\star}(x,j_1,j_2)=\mathbf 1\!\left[f^\star_{j_1}(x)>f^\star_{j_2}(x)\right],$$7, and analyzes robustness to corrupted SD labels, corrupted Pcomp labels, and class-prior estimation error (Tate et al., 20 Mar 2026). In graph learning, Contrastive FUSE proves Lipschitz smoothness of the unified objective and directional stability of the approximate modularity gradient (Chakraborty et al., 19 May 2026). In MU-OPPO, the theory quantifies how contamination in the confident positive set shifts estimated priors through

$$A_{\mathrm{comparisons}}^{f^\star}(x,j_1,j_2)=\mathbf 1\!\left[f^\star_{j_1}(x)>f^\star_{j_2}(x)\right],$$8

and upper-bounds this error under the latent-representation confidence mechanism (Wu et al., 2023).

6. Interpretation, misconceptions, and research significance

A common misconception is that pairwise supervision is synonymous with binary preference learning. The recent literature is broader. PEAR is explicitly graded rather than binary, allowing ties and preference strength; ReTVL uses pairwise supervision only locally, around retry events, while retaining absolute global calibration; SD-Pcomp combines heterogeneous pairwise labels rather than a single preference channel; CL-Net uses pairwise supervision for correspondence rather than scalar ranking (Proietti et al., 25 Jan 2026, Qin et al., 23 Jun 2026, Tate et al., 20 Mar 2026, Zhao et al., 2022). Pairwise progress supervision is therefore better viewed as a design principle—replace or augment absolute labels with structured relative information—than as one fixed loss class.

A second misconception is that pairwise supervision obviates the need for absolute anchors. ReTVL’s ablation explicitly shows otherwise: removing preference loss hurts local mistake sensitivity, but removing absolute calibration weakens global progress anchoring, dropping VOC to 0.929 and S/F Detection to 0.967 (Qin et al., 23 Jun 2026). In many problems, pairwise supervision supplies local ordering while absolute supervision supplies scale, calibration, or identifiability.

A third misconception is that pairwise information is universally better than full labels. The active-learning and uncoupled-regression results reject that stronger claim. Pairwise comparisons can dominate naive label-query simulation when the label neighborhood graph is sparse, but they do not improve worst-case PAC sample complexity; uncoupled regression attains consistency under uniform-target conditions, yet the general problem is not identifiable from pairwise comparisons alone (Yona et al., 2022, Xu et al., 2019). The effectiveness of pairwise progress supervision is thus conditional on geometry, task structure, and the relation between comparative and absolute information.

What the literature does support is a more specific conclusion: when the downstream task is inherently comparative, locally corrective, cross-view, or structurally relational, pairwise supervision can align the objective more directly with the task than one-item regression or coarse weak labels. PEAR shows this for translation comparison and MBR decoding; ReTVL shows it for local mistake sensitivity and recovery in imperfect demonstrations; Contrastive FUSE shows it for featureless graphs with partial pair labels; SD-Pcomp shows that heterogeneous pairwise judgments can reconstruct classification risk; CL-Net shows that explicit correspondence supervision can improve low-FPI mammogram detection (Proietti et al., 25 Jan 2026, Qin et al., 23 Jun 2026, Chakraborty et al., 19 May 2026, Tate et al., 20 Mar 2026, Zhao et al., 2022). This suggests that the principal research significance of pairwise progress supervision lies not in replacing all absolute supervision, but in capturing relational structure that absolute labels either obscure or encode only indirectly.