OT Score: An OT based Confidence Score for Prototype-Assisted Source Free Unsupervised Domain Adaptation

Alex Cloninger; Sitong Liu; Yiming Zhang

arxiv: 2505.11669 · v3 · submitted 2025-05-16 · 💻 cs.LG · cs.AI

OT Score: An OT based Confidence Score for Prototype-Assisted Source Free Unsupervised Domain Adaptation

Yiming Zhang , Sitong Liu , Alex Cloninger This is my paper

Pith reviewed 2026-05-22 14:09 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords optimal transportsource-free unsupervised domain adaptationconfidence scoreprototype-assisted alignmentpseudo-label uncertaintydomain adaptationsemi-discrete OTreweighting

0 comments

The pith

The OT score from semi-discrete optimal transport gives principled uncertainty estimates for target pseudo-labels in source-free domain adaptation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper addresses computational and theoretical shortcomings of distributional alignment methods that rely on source class-mean prototypes for source-free unsupervised domain adaptation. It develops the OT score through a new analysis that uses the flexible decision boundaries produced by semi-discrete optimal transport to generate tractable quantities reflecting the alignment process. This yields an interpretable confidence metric that supplies uncertainty estimates for any collection of target pseudo-labels in the complete absence of target labels. Readers would care because existing frameworks often produce intractable results that do not capture the actual alignment algorithm, limiting reliable performance estimation and training improvements in unlabeled target domains. The work shows the score exceeds prior confidence measures and raises adaptation accuracy when used to reweight training examples.

Core claim

By exploiting the flexibility of decision boundaries induced by semi-discrete optimal transport alignment, a novel theoretical analysis produces computationally tractable quantities that form the OT score. This score is both intuitively interpretable and theoretically rigorous, supplying principled uncertainty estimates for any given set of target pseudo-labels and serving as a label-free proxy for model performance in source-free unsupervised domain adaptation.

What carries the argument

The OT score, a confidence metric derived from theoretical analysis of semi-discrete optimal transport alignment that exploits flexible decision boundaries to yield tractable uncertainty quantities reflecting alignment properties.

Load-bearing premise

The flexibility of decision boundaries induced by semi-discrete optimal transport alignment enables a novel theoretical analysis that produces computationally tractable quantities reflecting the alignment algorithm properties.

What would settle it

If experiments demonstrate that the OT score neither correlates more strongly with actual target accuracy than existing scores nor improves adaptation performance when used for reweighting, the central claim would be falsified.

read the original abstract

We address the computational and theoretical limitations of current distributional alignment methods for source-free unsupervised domain adaptation (SFUDA) using source class-mean features. In particular, we focus on estimating classification performance and confidence in the absence of target labels. Current theoretical frameworks for these methods often yield computationally intractable quantities and fail to adequately reflect the properties of the alignment algorithms employed. To overcome these challenges, we introduce the Optimal Transport (OT) score, a confidence metric derived from a novel theoretical analysis that exploits the flexibility of decision boundaries induced by Semi-Discrete Optimal Transport alignment. The proposed OT score is intuitively interpretable and theoretically rigorous. It provides principled uncertainty estimates for any given set of target pseudo-labels. Experimental results demonstrate that OT score outperforms existing confidence scores. Moreover, it improves SFUDA performance through training-time reweighting and provides a reliable, label-free proxy for model performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper introduces an OT-derived confidence score for source-free domain adaptation that ties directly to semi-discrete transport properties and shows gains via reweighting, but the derivations need close checking to confirm they are independent and tractable.

read the letter

The main takeaway is that this work proposes an OT score as a confidence measure for prototype-assisted SFUDA, derived from a theoretical analysis of semi-discrete optimal transport alignment. It positions the score as interpretable, label-free, and useful both for uncertainty estimation and for reweighting target samples during training to boost adaptation performance. Experiments reportedly show it beating prior confidence heuristics and improving overall results, which addresses a practical pain point in the subfield where source data is unavailable and performance estimation is tricky without labels. That framing is the clearest new angle here. The attempt to ground the score in properties of the alignment algorithm rather than purely empirical tweaks is a step forward from many ad-hoc methods in SFUDA literature. If the analysis holds, it could give a more principled proxy for model quality on the target domain. The soft spots sit mostly in the theory and verification. The abstract claims the score comes from exploiting flexible decision boundaries in semi-discrete OT, but it is not obvious from the high-level description whether the resulting quantity stays independent of the fitted prototypes or ends up circular with the alignment itself. Without the explicit derivations or assumptions laid out, it is hard to judge computational tractability or generality across different OT solvers. On the experimental side, the outperformance claims would benefit from tighter controls on baseline implementations and more datasets to rule out setup-specific effects. This is squarely for researchers in unsupervised domain adaptation who care about source-free settings and need better ways to handle pseudo-label uncertainty. A reader already working with OT or prototype methods would find the reweighting application and the label-free performance proxy directly relevant. The work shows enough engagement with the core limitations to merit a serious referee, even if revisions will likely focus on fleshing out the proofs and strengthening the empirical comparisons. I would send it to review rather than desk reject.

Referee Report

2 major / 2 minor

Summary. The paper introduces the OT Score, a confidence metric for prototype-assisted Source Free Unsupervised Domain Adaptation (SFUDA). Derived from a novel theoretical analysis of Semi-Discrete Optimal Transport alignment exploiting decision boundary flexibility, it claims to deliver intuitively interpretable and theoretically rigorous principled uncertainty estimates for target pseudo-labels. Experiments show it outperforms existing confidence scores, improves SFUDA performance through training-time reweighting, and serves as a reliable label-free proxy for model performance.

Significance. If the theoretical derivation and experimental results hold, this could meaningfully advance SFUDA by supplying a tractable, OT-grounded alternative to heuristic confidence measures that better reflect alignment algorithm properties. The focus on computational tractability and interpretability addresses real limitations in current distributional alignment approaches for label-free settings.

major comments (2)

§3.2, Eq. (10): The derivation of the OT score must explicitly show that the resulting quantity does not reduce to a direct function of the fitted transport plan or cost matrix; otherwise the claim of providing independent principled uncertainty estimates is at risk of circularity with the SDOT alignment step itself.
Table 4, reweighting rows: The reported accuracy gains lack error bars or results over multiple random seeds; without this, it is difficult to assess whether the improvements are stable or could be explained by variance in pseudo-label quality.

minor comments (2)

The abstract would benefit from naming the specific datasets and baseline SFUDA methods used in the experiments to give readers immediate context.
Figure 2: Axis labels and the color scale for the OT score visualization should be clarified to make the decision boundary flexibility more immediately readable.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We sincerely thank the referee for their constructive and insightful comments on our manuscript. We have carefully addressed each major comment below and describe the revisions we will incorporate to strengthen the paper's clarity, rigor, and empirical robustness.

read point-by-point responses

Referee: §3.2, Eq. (10): The derivation of the OT score must explicitly show that the resulting quantity does not reduce to a direct function of the fitted transport plan or cost matrix; otherwise the claim of providing independent principled uncertainty estimates is at risk of circularity with the SDOT alignment step itself.

Authors: We thank the referee for raising this important concern about potential circularity. In the derivation of the OT score (Section 3.2, Eq. (10)), the score is obtained from the dual formulation of the semi-discrete OT problem and specifically quantifies uncertainty via the distance to the flexible decision boundary induced by the prototype alignment; it is not a direct algebraic function of individual entries in the transport plan or the cost matrix. To eliminate any ambiguity, we will revise the manuscript by adding an explicit paragraph immediately following Eq. (10) that formally demonstrates the OT score cannot be reduced to a function of the fitted plan or cost matrix alone, thereby confirming its status as an independent uncertainty measure. revision: yes
Referee: Table 4, reweighting rows: The reported accuracy gains lack error bars or results over multiple random seeds; without this, it is difficult to assess whether the improvements are stable or could be explained by variance in pseudo-label quality.

Authors: We agree with the referee that the absence of error bars and multi-seed results makes it harder to judge stability. The numbers in Table 4 were generated with a fixed seed for exact reproducibility. In the revised manuscript we will re-run the OT-score reweighting experiments across five independent random seeds, report mean accuracies with standard-deviation error bars, and add a short discussion confirming that the observed gains remain consistent and are not explained by pseudo-label variance alone. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained

full rationale

The paper presents the OT score as the output of a novel theoretical analysis that exploits properties of semi-discrete OT alignment to produce tractable uncertainty quantities for pseudo-labels. No equation or step in the abstract or stated claims reduces the derived score to a fitted parameter, a self-citation chain, or an input by construction. The central derivation is described as independent of the target performance metrics, with experimental reweighting results offered as separate validation. The manuscript is treated as self-contained against external benchmarks, yielding no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that semi-discrete OT produces flexible decision boundaries suitable for deriving a tractable confidence score; no free parameters or invented entities are explicitly detailed in the abstract.

axioms (1)

domain assumption Semi-discrete optimal transport alignment induces flexible decision boundaries that support a novel theoretical analysis yielding tractable quantities.
Invoked as the basis for overcoming limitations of current theoretical frameworks in the abstract.

pith-pipeline@v0.9.0 · 5683 in / 1047 out tokens · 32322 ms · 2026-05-22T14:09:21.774840+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Theorem 6 ... sup max min ~d_w(x,y) - ~d_w(x,z) ... g value gap reflects flexibility of classification boundary induced by semi-discrete OT
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Algorithm 1 ... entropically smoothed indicator χ^ε_j ... Laguerre cells

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.