acceptodds
Under review as a conference paper at ICLR 2027

Certified Conflict Policies for Few-Shot Class Registration

Abstract

Deployed vision systems often need to recognize new categories from only a few labeled examples while keeping the existing classifier fixed. However, freezing the classifier does not freeze its predictions. New classes compete with old classes in the same decision rule and can cause old examples to be classified as new. We study this problem as few-shot class registration and introduce Certified Arbitration for Post-training Expansion (CAPE), which provides finite-sample control of old-to-new interference over an old-class calibration frame. CAPE separates risk certification from conflict ranking, using a scorer to prioritize new-class predictions that are less likely to interfere with old classes. We find that certified utility depends jointly on calibration evidence and ranking quality, motivating Certified Conflict Policy Learning (CCPL), which learns a reusable scorer across registrations and applies it without refitting. Across 900 development episodes under a common certification protocol, CCPL retains substantially more new-class utility than two external scoring pipelines. CCPL also improves over the episode-local CAPE baseline and transfers without retraining to both held-out populations, including a 3.27 percentage-point gain on FGVC Aircraft. These results show that reusable conflict scoring can improve utility under finite-sample risk control for class registration without retraining the old classifier.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Few-shot learning, Class-incremental learning, Continual learning, Risk control, Certified prediction, Conflict ranking

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_yiv7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper points out that adding new classes can change old predictions even with frozen old parameters, and refers to this as insertion-induced class-competition interference. The paper proposes CAPE framework for few-shot class registration, which providing finite-sample control of old-to-new interference over an old-class calibration frame. It instantiates registration-specific CAPE-BF and amortized cross-registration CCPL as two scorer regimes for CAPE. Extensive experiments are conducted on 900 development episodes and two held-out populations, which supports separating learned conflict ranking from per-registration risk certification.

Critical strengths

1.The paper demonstrates a clear motivation. It points out that the fixing the model parameters does not necessarily mean that the model's decision remains unchanged when new classes are added, and refers to this as insertion-induced class-competition interference.

2.The paper is well structured, and the proposed CAPE is sound, with two scoring regimes instantiated as CAPE-BF and CCPL.

3.The paper presents a sound theoretical analysis. Extensive experiments are conducted across 900 development episodes and two held-out populations.

Critical weaknesses

1.CAPE maintains the calibration frame for old classes. As new registrations accumulate and the number of old classes grows incrementally, the calibration frame may expand, leading to increasing storage costs.

2.CAPE sets a limited number of old-class calibration anchors. As the number of old classes increases, the calibration information allocated to each old class may become fewer. Therefore, conflict between new registration and individual old classes may not be sufficiently reflected, thereby weakening the ability to protect existing old classes.

3.CAPE-BF uses a 10-D passive block, and CCPL uses the full 17-D conflict representation. Each dimension is computed based on frozen DINOv3 and CLIP features, which incurs additional computational overhead. Additionally, for CAPE-BF, this may introduce additional time costs for each registration.

4.The primary aim of paper is to control old-to-new errors. However, the experiments in the main paper focus primarily on new-class utility, and as can be seen in Figure 1(C), the proposed method results in a higher harm05 value than BiMC.

Questions for authors

1.Regarding Weakness 1 and 2, as new registrations accumulate incrementally, could the authors explain whether maintaining the calibration frame for old classes incurs increasing storage costs, and whether the potential reduction in candidate anchors for each old class might affect the ability to protect old classes?

2.Regarding Weakness 3, could the authors explain whether the conflict representation in CAPE-BF and CCPL would introduce additional computational and time costs?

3.Regarding Weakness 4, given that the primary aim is to control old-to-new errors, could the authors explain why the experiments focus primarily on new-class utility?

2 comments
AuthorsOpenReview ↗

Reply to reviewer yiv7

Thank you for the encouraging assessment and the concrete questions, which helped us make the deployment side of the paper more precise. We address them in order.

Q1 / W1 — Storage cost of the calibration frame as classes accumulate.

Storage grows linearly with the number of old classes, while each registration scores at most 530 anchors. The frame stores a fixed number of labelled feature rows per old class (27 in our protocol), so it grows with the old label set like the classifier itself. Each registration scores at most 530 anchors, drawn by stratified sampling without replacement from the frame, so the scoring cost of a registration does not grow with the frame.

Q1 / W2 — Fewer anchors per old class as classes grow.

The certificate is an aggregate guarantee over the old classes, and the design-exact budget uses the per-class design. With 530 anchors, an old class receives about 26 anchors with 20 old classes, 18 with 30, 13 with 40 and 7 with 80. The certified quantity weights every class by its share of the design, and the FP budget computes its tail class by class from that design, so thinner strata are accounted for exactly rather than approximated. A class-wise guarantee would need class-wise budgets: at ε = 2%, at least 149 anchors per class to certify zero violations.

Q2 / W3 — Computational overhead of the conflict representation.

Given features, CCPL arbitrates in 0.72 ms per registration on a Jetson Nano, about 11× faster than CAPE-BF (8.02 ms; Appendix H). This is the measured arbitration layer of vanilla CCPL; removing per-registration fitting is what makes it fast, and the serialized MLP is 46,732 bytes (634 bytes for CCPL-Linear). Aligned CCPL-FP runs the same ranker together with BiMC's proposal and margin. Most conflict dimensions use a second frozen encoder (DINOv3), which adds one forward pass per candidate query and per anchor and sits outside these latency numbers. The design-exact budget is an offline cost. It depends only on the anchor design, so it is solved once per design and cached: a first solve takes a median of 14 s (p90 144 s), and building the cache for all 303 development designs took about 20.9 h (Appendix K). Registrations with a cached design read the budget directly.

Q3 / W4 — Why focus on new-class utility, and how does old-class protection compare?

Our objective is to retain more correct new-class predictions under a fixed, certified old-to-new risk constraint. The certificate controls interference with existing classes, while new-class utility measures how effectively the registered classes are recognized under that constraint.

Our new matched-proposal comparison provides direct evidence of this benefit. Under the same BiMC proposal, final-draw anchors, and FP certification budget, aligned CCPL improves new-class utility from 87.37% to 89.76% (+2.39pp) across 900 development episodes. The empirical old-to-new migration rate remains comparable (0.53% versus 0.55%), as does the mean old-class accuracy change (+5.28pp versus +5.26pp). The descriptive harm05 measure is 15.8% versus 14.8%.

The benefit is particularly clear on held-out FGVC Aircraft. Evaluated once without further retraining, aligned CCPL improves certified new-class utility from 32.36% to 42.52% (+10.16pp), while also reducing harm05 from 10.0% to 9.3%. These results demonstrate the practical value of learned conflict ranking: more correct new-class predictions can be retained without relaxing the per-registration certification guarantee.

Thank you for your support of this work; we are glad to discuss any further questions.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer,

Thank you once again for your valuable comments on our submission. As the discussion phase is approaching its end, we would like to kindly confirm whether we have sufficiently addressed all of your concerns. Should there be any remaining questions requiring further clarification, please do not hesitate to let us know.

We sincerely look forward to your feedback.

Reviewer_DU4AOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how to add new classes to a frozen classifier from a few examples without sending old-class inputs to the new classes. The author proposes CAPE, which first inserts the new classes directly, then scores each new-class win and keeps only the wins above a threshold. The threshold is chosen for each registration so that a Clopper–Pearson upper bound on the old-to-new violation rate over a labeled calibration frame is at most 2%. The paper compares a scorer fitted for each registration (CAPE-BF) with a scorer trained once across registrations (CCPL). On 900 development episodes, CCPL keeps 83.22% new-class accuracy, against 82.79% for CAPE-BF and 86.11% for direct insertion without a certificate, and it also improves over CAPE-BF on two held-out populations.

Critical strengths
  1. The motivation is clear. The paper shows that freezing the old parameters does not freeze the old predictions, because old and new classes compete in the same argmax.
  2. The method is simple and the guarantee is correct. Ranking and certification are kept separate, so one trained scorer can be reused across registrations without refitting. I checked the proof of Theorem 1 and recomputed kmax(530) = 5 and the n ≥ 149 requirement for zero violations, and both match the paper.
  3. The experiments are comprehensive. They cover six development populations, two held-out populations, paired bootstrap intervals, and controls such as label shuffling and a fixed-margin scorer (Tables 6 and 23).
Critical weaknesses
  1. The novelty is limited. The certificate is a standard Clopper–Pearson bound used to select a threshold, as in the risk-control methods the paper cites (Sec. 2). The setting is also close to prior work. For example, few-shot learning without forgetting adds new classes from a few examples while keeping the accuracy on the base classes [1]. Calibrated stacking adds one shared offset between the scores of two class groups [2].

  2. The guarantee does not cover new queries. Theorem 1 only bounds the violation rate on the calibration frame, which has at most 27 labeled images per old class (Appendix B.4). Section 3 also says that held-out old queries are outside the guarantee. Since every image in the frame is labeled, one could just count the violations on the whole frame, and no bound would be needed. In fact, the anchors already cover almost the whole frame, such as 530 of 540 images for a tieredImageNet episode with 20 old classes, and all of it on Food-101 (Table 4). So I am not sure what the certificate adds beyond the plain error count on these images.

  3. The certificate costs a lot of new-class accuracy, and it is not clear that it is needed. Without the certificate, native BiMC already reaches 92.41% new-class accuracy, and it even improves old-class accuracy by 5.96pp over direct insertion (Table 8). Its harm05 is 23.6%, close to CCPL's 20.8%, while CCPL reaches only 83.22%. Once the certificate is applied, BiMC falls to 71.85%. The drop is even larger on Aircraft, where certified CCPL keeps 35.04% against 52.09% for direct insertion (Sec. 7.3). Table 2 therefore tells us which score works best under this certificate, rather than which method is more useful.

  4. The benefit of the shared scorer is small on the development populations. On average CCPL beats CAPE-BF by only 0.43pp (Table 1). The point estimate is negative on three of the six populations, and four of the six intervals include zero (Table 11). The two scorers also give exactly the same result in 55.4% of episodes (Appendix D). This gain is small next to the 3.32pp that the certificate itself costs CAPE-BF (Table 2). The gains on the two held-out populations are larger, at 0.88pp and 3.27pp (Table 3).

  5. The paper is hard to read. Section 3 does not say what the old classifier is or how large an episode is. The classifier is defined only in Appendix B.1, as a fixed mix of text-anchor and visual-prototype scores. The number of old and new classes in an episode is given only in Appendix C.1. The goal is also never stated as one problem, such as maximizing new-class accuracy while keeping the old-to-new rate at most 2%. In addition, the main text uses terms such as the census case and the passive block, which are explained only in the appendix.

Questions for authors
  1. Under the final calibration draw, what is the old-to-new violation rate on the held-out old queries of tieredImageNet and Aircraft, and in what fraction of episodes does it exceed 2%?

[1] Dynamic Few-Shot Visual Learning without Forgetting, CVPR 2018.

[2] An Empirical Study and Analysis of Generalized Zero-Shot Learning for Object Recognition in the Wild, ECCV 2016.

3 comments
AuthorsOpenReview ↗

Reply to reviewer DU4A (part 1 of 2)

We are grateful for the careful review, in particular for checking Theorem 1 and recomputing k_max(530) = 5 and the n ≥ 149 requirement. Your Weakness 3 quotes a BiMC value of 71.85%; with the threshold rule of Algorithm 1 applied as specified (general response, Appendix D) the value is 86.57% (BiMC here is the non-ensemble configuration described in the summary of revisions). Your reading of the theory also led to two additions in the revision: the census case in §4 and the design-exact budget.

W1 — Novelty; relation to [1] and [2].

What is new is certifying a learned conflict ranking on every registration: under the same certificate, CAPE-BF's learned scorer retains 6.46pp more new-class utility than scalar-margin calibration. Gidaris & Komodakis [1] learn a weight generator with attention over base classes and evaluate average base and novel accuracy; they do not certify per-registration interference. Calibrated stacking [2] subtracts one shared offset γ from seen-class scores and traces the seen–unseen trade-off with AUSUC; with old classes as the seen set, a new class wins exactly when its margin over the best old class exceeds −γ, a single scalar threshold. Choosing γ per registration to satisfy our certificate gives precisely the Fixed-Margin arm of Table 7 (Appendix D defines it this way). Under the same certificate and calibration take it reaches 76.23%, 6.46pp [5.89, 7.06] below CAPE-BF (82.69%); on the canonical final draw, CCPL reaches 83.22%. CAPE generalizes calibrated stacking by replacing the margin with a learned conflict score while keeping the certificate, and the gap shows how much the score matters. §2 of the revision now cites and discusses both works in these terms.

W2 — If the whole frame is labelled, why not count all of it?

In the census case counting is exact, and the revision's design-exact budget does exactly that; the sampled certificate covers the common case in which the scoring budget is smaller than the frame. With n_j = N_j in every class, the design-weighted rate equals the observed count rate, so k ≤ εn decides the constraint. The CP bound is conservative there, and the FP budget recovers the exact census value ⌊εn⌋ (§4, Appendix K). Your example, 530 of 540, is the 20-old-class case. With 27 rows per old class against a fixed 530-anchor budget, most registrations are sampled: 621 of 900 on development and every held-out registration. From the submission's design records, by sampling fraction f = n / ΣN_j:

Setfull census0.9 ≤ f < 1f < 0.9median f when f < 0.9
Development279/900 (all of Food-101)135/900486/9000.73
tieredImageNet0/30075/300225/3000.65
FGVC Aircraft0/30075/300225/3000.65

These counts come from the existing design records; no new run was needed. For the sampled designs, the FP budget uses the exact stratified law and admits more violations than CP on every design it resolves (median +3 per development episode, 6–8 instead of 5 on Aircraft). Counting and certification thus become one rule: exact where the frame is fully scored, and as tight as the design allows where it is sampled.

AuthorsOpenReview ↗

Reply to reviewer DU4A (part 2 of 2)

W3 — The certificate costs much utility; is it needed?

Certification is what keeps old classes safe, and the revision makes it cheaper at the same guarantee. Native BiMC reaches 92.41% and certified BiMC 86.57%: the certificate-only stage accounts for −4.37pp and minimum activity for −1.47pp (Table 9). Without certification, old classes pay: on held-out Aircraft, direct insertion keeps 52.09% new-class utility but routes 7.54% of held-out old queries to new classes, with harm05 96.3%; certified CCPL routes 0.92% with harm05 14.7%. The revision recovers part of the cost in two ways, both at the same ε and 95% level. The design-exact budget adds +0.80pp to BiMC over the full bank and +3.26pp on untouched DTD episodes, and a better scorer adds more: under the same budget, aligned CCPL keeps 89.76% against 87.37% for BiMC, and on Aircraft 42.52% against 32.36%.

W4 — The shared scorer brings a small benefit on development.

The shared scorer gains more on held-out Aircraft (+3.27pp) and with a stronger proposal (+2.39pp on development, +10.16pp on Aircraft). Retrained on BiMC's proposal, the ranker improves on BiMC by +2.39pp on development, positive on all six populations, and by +10.16pp on held-out Aircraft (§7.4). Without retraining, vanilla CCPL gains +0.88pp on tieredImageNet and +3.27pp on Aircraft over CAPE-BF, and on Aircraft under the matched final draw and CP budget it reaches 35.04% against 29.37% for BiMC-CP (+5.67pp; Table 2, Appendix K.6; under the earlier draw, vanilla CCPL-FP reaches 38.18%). On development its gain over CAPE-BF is +0.43pp, rising to +0.89pp on the 672 registrations where CAPE-BF can be fitted.

W5 — Readability.

The revision states the problem, the census case and the classifier where a reader first needs them. §3 now has an Objective paragraph: maximize U_new subject to the per-registration calibration-frame old-to-new constraint at ε = 2%. The census case is stated in §4, where the certificate is defined. §3 also gives the episode scale (20–35 old classes, five new classes, five shots) and the B2 score, 0.75 times the cosine to the frozen text anchor plus 0.25 times the cosine to the visual prototype; §6 names the inputs of the 10-D scorer (multi-level DINOv3 conflict margins, a consistency feature and the B2 margin) and the seven added CLIP features of the 17-D scorer (old and new evidence, prototype similarity, support dispersion and ambiguity).

Q1 — Held-out query-side old-to-new rates under the final calibration draw.

On FGVC Aircraft, both certified arms keep pooled held-out routing below 1%. Final per-registration draw, deployed threshold rule, 300 registrations, 49,500 held-out old queries:

CAPE-BFCCPL-MLP
Pooled old-to-new rate0.82%0.92%
One-sided 95% upper bound0.89%0.99%
Routed queries407454
Registrations above 2%10.0% (30/300)11.3% (34/300)
Per-registration 95th percentile2.50%2.50%
Per-registration maximum3.33%4.17%

Direct insertion routes 7.54% pooled. The pooled rates appear in the submission (§7.3, Appendix I); the per-registration statistics come from the same stored records. Registrations above 2% are what binomial granularity predicts: with 120–210 old queries per registration, if every true rate were exactly 2%, about 42% would show an observed rate above 2%; the observed 10–11% is far below that. On tieredImageNet under the final draw (deterministic CPU replay of the frozen scores and final-draw thresholds, Appendix E.1), across the 291 fitted episodes (48,090 old queries), CAPE-BF routes 0.93% (447; >2% in 8.9% = 26/291, p95 2.67%, max 4.17%) and CCPL-MLP routes 0.89% (427; >2% in 6.9% = 20/291, p95 2.50%, max 4.17%); across all 300 operational episodes (49,500 old queries, B2 pass-through for CAPE-BF on the 9 no-fit episodes), CAPE-BF routes 0.92% (456; >2% in 9.0%) and CCPL-MLP routes 0.87% (433; >2% in 6.7%), against 2.73% (56.7%) for direct insertion. Under the final draw, BiMC-FP and aligned CCPL-FP route 0.60% and 0.61% (>2% in 2.3% and 3.0% of 300 episodes). On Aircraft, aligned CCPL-FP routes 1.18% against 0.99% for BiMC-FP (Appendix K.6). These are empirical query-side measurements outside Theorem 1's calibration-frame guarantee.

Your careful reading of the theory also sharpened our comparator evaluation. The census analysis, the design-exact budget and its gains in certified utility follow directly from your questions, and we hope they support a reconsideration of your score.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer,

Thank you once again for your valuable comments on our submission. As the discussion phase is approaching its end, we would like to kindly confirm whether we have sufficiently addressed all of your concerns. Should there be any remaining questions requiring further clarification, please do not hesitate to let us know.

We sincerely look forward to your feedback.

Reviewer_GqR4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper examines old-class recognition interference during few-shot class registration. It proposes the CAPE risk certification framework and introduces the cross-registration shared scoring method CCPL. CAPE controls the risk of old-to-new misallocation. CCPL aims to preserve new-class recognition ability under risk control. The paper evaluates CCPL on 900 development evaluation episodes and two held-out populations. CCPL achieves an average improvement of 0.43 percentage points over CAPE-BF on the development populations. It also achieves improvements of 0.88 and 3.27 percentage points on tieredImageNet and FGVC Aircraft, respectively. The experimental results show some benefits of cross-registration shared scoring.

Critical strengths
  1. Clear methodological workflow. The paper separates conflict scoring from risk certification. It presents two implementations, CAPE-BF and CCPL, and describes the corresponding threshold selection and prediction decision procedures.
  2. Broad experimental coverage. The paper conducts experiments on six development populations and two held-out populations. It also provides comparative experiments, ablation studies, and risk-level sensitivity analyses.
Critical weaknesses
  1. Limited technical novelty. The scoring network and pairwise ranking loss show no clear independent advantages. The paper also provides insufficient justification for the necessity of several key designs.
  2. Limited performance gains and unstable episode-level results. CCPL achieves only a small overall improvement. Some episodes show substantial performance degradation. The minimum-activity constraint also introduces an additional loss of new-class accuracy.
  3. Limitations in the experimental design. The main comparisons change multiple experimental factors simultaneously and do not sufficiently isolate their individual effects. The development training and evaluation data also lack strict separation at the original-image level.
Questions for authors

no

3 comments
AuthorsOpenReview ↗

Reply to reviewer GqR4 (part 1 of 2)

Thank you for the review. Your three weaknesses all concern what the comparisons isolate, and the revision now answers this directly with a comparison that holds the proposal fixed (Table 3). BiMC here and below is the non-ensemble configuration described in the summary of revisions.

W1 — Limited technical novelty; necessity of the key designs.

The central design choice is where conflict ranking is estimated: within each registration, or once across registrations and reused without refitting. CAPE makes this choice possible by separating ranking from certification, so the scorer can change while the exact per-registration guarantee stays fixed. The paper separates this choice from every factor that changes with it:

FactorControlResult
CapacityCCPL-Linear vs. CCPL-MLP+0.44 vs. +0.43pp over CAPE-BF (§7.2)
Representationepisode-local arbiter given the same 17-D inputCCPL still +0.91pp [0.57, 1.25] ahead (App. D)
Certificatefrozen margin-based scorer, same certificate6.46pp below CAPE-BF (Table 7)
Supervisionshuffled training labelsreal scorer 9.05pp ahead (§7.5)
Transfershared scorer on unseen populations, no retraining+0.88pp tieredImageNet, +3.27pp FGVC Aircraft
Proposalsame BiMC proposal, learned vs. native scoring+2.39pp development; +10.16pp held-out Aircraft (Table 3)
Learned weightssame proposal and 17 features, learned vs. fixed rule+1.82pp [1.56, 2.09] (App. L)

Together these point to shared estimation as the driver of the gain, as §7.2 concludes, and §7.5 shows why: episode-local rankings become unstable as local evidence shrinks, and CCPL's advantage grows from +1.26 to +3.97pp exactly there. What is new is the decomposition that lets one reusable scorer be certified on every registration, and the revision adds a design-exact budget that certifies the same estimand with more admitted violations (§4). §6 now lists the input, loss and estimation differences between the two regimes and points to these controls.

W2 — Small gains, registration-level degradation, and the cost of the minimum-activity rule.

CCPL's advantage holds without the minimum-activity rule and grows from +0.43pp to +0.58pp. The rule prevents a certificate that is satisfied by doing nothing: when direct insertion sends sampled anchors to new classes, the certified rule must reject at least one of them. Removing it raises both arms by a similar amount (Appendix D):

With the ruleWithout
CCPL83.22%83.89%
CAPE-BF82.79%83.31%
CCPL − CAPE-BF+0.43pp+0.58pp

Where the scorer matters more, so does the benefit: on unseen populations the gain is +0.88 and +3.27pp, and on a strong proposal aligned CCPL is +2.39pp above BiMC on development, with gains on all six populations, and +10.16pp on held-out Aircraft, better on 272 of 300 episodes. The development gain of vanilla CCPL is +0.43pp; the two methods give identical results on 55.4% of registrations, and Appendix D reports the full per-registration distribution.

AuthorsOpenReview ↗

Reply to reviewer GqR4 (part 2 of 2)

W3a — The main comparisons change several factors at once.

Table 3 now changes one factor at a time: holding the proposal fixed, learned conflict ranking improves certified new-class utility by 2.39 percentage points. It keeps BiMC's proposal, the episodes, anchors and threshold rule fixed and swaps either the scorer (native margin vs. aligned CCPL: +2.55pp CP, +2.39pp FP) or the budget (CP vs. FP: +0.80pp native, +0.64pp aligned); the interaction is small (−0.16pp). The aligned scorer retrains the ranker on BiMC's candidates, pair labels and proposal-score features and combines it with BiMC's margin, so it is a retrained ranker combined with BiMC's margin on the same proposal. Capacity, objective, optimizer and the BiMC configuration are unchanged.

In Table 1, all arms share the registrations, the frozen old classifier, the calibration anchors (one per-registration draw), the threshold rule, the activity budget and the certificate. What differs:

ArmCandidatesScorerFitted on
CAPE-BFdirect-insertion new-class winsclass-balanced logistic, 10-D passive blockthat registration only (support pseudo-queries, fit-frame negatives)
CCPL-MLPsamepairwise MLP, 17-Ddevelopment-training registrations 150–299, once; no refitting
CCPL-Linearsamelinear, same items as CCPL-MLPas CCPL-MLP (isolates capacity)
Representation-matchedsamecross-registration scorer on the 10-D inputas CCPL-MLP (isolates the input; Appendix D)
LP++-NEW-onlyjoint argmax of frozen old-class evidence and LP++ new-class scores (positive affine bridge)joint-score margin5-shot support, per registration
BiMCits own native old ∪ new score matrixtraining-free; per-population β by its query-blind rulerejected wins fall back to BiMC's own old-class argmax
Aligned CCPLBiMC's native proposalBiMC margin + retrained ranker (Eq. 5)development-training registrations, once; λ = 1 chosen on their validation split

W3b — Development training and evaluation are not separated at the image level.

The held-out populations carry the transfer claims, and they are class-disjoint from development. Training registrations (150–299) and evaluation registrations (0–149) are disjoint (Appendix C), but registrations of one development population draw from the same per-class image pools, so an image can occur in both. This is exactly why the transfer claims rest on tieredImageNet and FGVC Aircraft, which were not used for training, selection or tuning of vanilla CCPL; its gain over CAPE-BF is larger there (+0.88 and +3.27pp vs. +0.43pp), so the gain does not depend on shared images. The advantage also persists when the evaluated development population is left out of training (Appendix J). For aligned CCPL, Aircraft was not touched by any aligned development step; the frozen method was evaluated once there and gains +10.16pp. As stated in Appendix B.4, the canonical results use a per-registration calibration draw that is independent of scorer training.

Your question on what the comparisons isolate led directly to Table 3. The fixed-proposal comparison, the matched-feature control and the held-out Aircraft evaluation answer it, and we hope they support a reconsideration of your score.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer,

Thank you once again for your valuable comments on our submission. As the discussion phase is approaching its end, we would like to kindly confirm whether we have sufficiently addressed all of your concerns. Should there be any remaining questions requiring further clarification, please do not hesitate to let us know.

We sincerely look forward to your feedback.

Reviewer_ZteyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates few-shot class registration, where new classes are introduced into a pretrained classifier without updating the existing classifier parameters. The authors investigate how newly registered classes compete with existing classes under a shared decision rule. To address this, they propose Certified Arbitration for Post-training Expansion (CAPE), which uses old-class calibration samples to select a threshold that controls old-to-new prediction interference under a finite-sample risk guarantee. The framework separates conflict scoring from risk certification and investigates two implementations: CAPE-BF, which learns a conflict scorer independently for each registration, and Certified Conflict Policy Learning (CCPL), which learns a reusable scoring policy across multiple registration episodes.

Critical strengths
  • The paper addresses a seemingly interesting and practical problem in few-shot class registration.
  • The separation between conflict scoring and risk certification is intuitive. By decoupling the ranking of candidate new-class predictions from the selection of threshold, the framework allows different scoring strategies to be evaluated under a common risk-control mechanism.
  • The paper tries to analyse and provide a theoretical formulation to account for class-stratified calibration sampling without replacement.
Critical weaknesses
  • Firstly, the introduction lacks clarity in presenting the problem setting and motivating the proposed framework. The paper introduces the notion of class registration without first providing a sufficiently clear definition or distinguishing it from conventional incremental learning (Lines 36–37). Also the explanation of insertion-induced class-competition interference introduces unnecessarily complex terminology for the relatively straightforward phenomenon of newly added classes competing with existing classes under a shared prediction rule. The introduction needs more concise and intuitive explanation of the registration setting.

  • The central motivation lacks clarity and is difficult to follow. The motivation for prioritizing old-to-new interference is not sufficiently established. The paper emphasizes that registering new classes can redirect old-class examples toward newly introduced classes, even when the original classifier is frozen. While this phenomenon is possible, the opposite failure mode is equally intuitive. Typically, a pretrained classifier may favor existing classes and incorrectly assign new-class samples to them, particularly in few-shot settings. The paper does not sufficiently explain why old-to-new interference should be the primary concern or characterize the relative prevalence of these two error types. An analysis of both error directions, together with the comparability of old- and new-class scores, would provide stronger justification for the proposed framework.

  • The main theoretical result in Theorem 1 establishes finite-sample control of the design-weighted old-to-new violation rate over a fixed calibration frame. However, this guarantee does not directly extend to unseen old-class queries or the underlying data distribution. Similarly, a certified violation rate on the calibration frame does not necessarily imply comparable protection under distributional variations or when the calibration frame is not sufficiently representative. These need a more detailed analysis to study the relationship between the certified finite-frame risk and the observed interference on held-out old-class.

  • Although the proposed framework is motivated by few-shot class registration, the finite-sample certification mechanism relies on a separate collection of labeled old-class calibration examples. The paper clarifies that the 5-shot restriction applies to the newly registered classes rather than the old-class calibration system. This distinction is important for the practical applicability of the method. In realistic incremental-learning scenarios, access to source labeled data may be restricted. While Section 5 and the appendix investigate calibration power, comparisons under substantially smaller old-class calibration budgets would help establish the practicality of the proposed framework in genuinely data-constrained registration settings.

  • A central contribution of CCPL is learning a reusable conflict-scoring policy across registration episodes rather than fitting a separate scorer for each episode. However, the scoring mechanism itself relies on conventional supervised ranking and relatively lightweight feature-based prediction models. The theoretical certification mechanism is also independent of the particular scorer used. Consequently, the main methodological distinction appears to be the amortization of scorer estimation across registration episodes rather than a fundamentally new scoring or learning formulation. We can see that even in the the experimental results where CCPL-Linear performs almost identically to the proposed CCPL-MLP. This raises questions about the necessity of the proposed learning architecture and the extent of the algorithmic contribution.

  • Considering the results, the improvements over CAPE-BF are relatively small in several experimental settings. For example, Table 1 reports negligible improvement (+0.43 points) across the six developments. Also we can see CCPL actually falls short Caltech-A (−0.22), Food-101 (−0.17) and SUN397 (−0.39). These results indicate that cross-registration scoring does not consistently improve performance across all populations.

  • The evaluation scope is limited. Main experiments use 5 shots per newly registered class. However, the quality of the registered class representations and the resulting competition with existing classes can depend strongly on the number of shots. For instance, with only one or two samples, new-class representations may be considerably noisier making conflict ranking more difficult. Conversely, with larger support sets, direct insertion may already provide better separation between old and new classes. It is key to evaluate across different shot settings, particularly 1, 2, 4, 8, and 16 shots to determine whether the proposed approach consistently improves class registration or is particularly effective under the selected 5-shot configuration.

Questions for authors
  • Can the authors clarify how the proposed few-shot class registration setting differs from conventional class-incremental learning? In particular, what are the practical assumptions regarding the frozen classifier, access to old-class data, and the registration of new classes?

  • In few-shot settings, a pretrained classifier may favor existing classes and incorrectly classify new-class examples as old classes. Can the authors provide an empirical comparison of these error directions and explain why old-to-new interference is prioritized? Additionally, to what extent does the observed interference arise from differences in old- and new-class score distributions.

  • Can the authors clarify the specific methodological contribution of CCPL beyond pooling supervision across registration episodes? Would simpler pooled or regularised scoring approaches provide comparable performance?

  • Why are the main experiments restricted to 5-shot class registration? Can the authors provide results for additional settings, particularly 1, 2, 4, 8, and 16 shots, to demonstrate whether the proposed approach remains effective under different levels of new-class supervision?

3 comments
AuthorsOpenReview ↗

Reply to reviewer Ztey (part 1 of 2)

Thank you for the detailed review. Your questions on the setting and the objective led us to state the problem much more directly in the revision, and we address them in order in this comment and the next.

Q1 / W1 — How does class registration differ from class-incremental learning?

Registration keeps the old classifier frozen and gives each registration its own finite-sample guarantee. The two settings differ in what is updated, what is kept and what is reported:

Class-incremental learning (typical)Certified class registration (ours)
Old model at a new sessionupdated (backbone, head or both)frozen: no backbone or old-head parameter changes
How new classes entertrained into the modelfew-shot support sets join the same shared argmax
Data kept from old classesexemplars or none, for traininga labelled calibration frame, for certification
What is reportedaverage accuracy over a sequenceper registration: a 95% bound on the old-to-new rate, plus new-class utility

Each episode has 20–35 old classes and five new classes with five shots each, and B2, the reference rule, scores a class by 0.75 times the cosine to its frozen text anchor plus 0.25 times the cosine to its visual prototype (§3, Appendix B.1). Even with everything frozen, adding classes changes old predictions, because all classes compete in one decision rule: direct insertion lowers old-class accuracy by more than 0.5pp in 57.3% of development episodes. Our question is therefore not how to learn new classes, but whether each registration can carry a finite-sample statement about how many old-class inputs it redirects. The revision adds an Objective paragraph to §3 that states this directly: for each registration, maximize new-class utility subject to the calibration-frame old-to-new constraint.

Q2 / W2 — Why constrain old-to-new errors, and what about new-to-old errors?

New-to-old errors are in the objective, and old-to-new errors are under a certified constraint. U_new is accuracy on new-class queries, so every new-to-old error lowers it directly, and every method is compared on it. Old-to-new errors are held under the certificate because they change predictions that existing users of the old classes rely on; a deployer can accept a new class that is sometimes missed, but not one that silently takes over old inputs. Registration is therefore a constrained optimization problem, not a single accuracy trade-off. New-to-old errors are also the dominant error on new-class queries. For the native, uncertified BiMC proposal (macro over the six development populations, 95% bootstrap intervals):

New-class queriesShare
correct92.41%
routed to an old class6.46% [6.12, 6.81]
routed to a wrong new class1.13% [0.98, 1.30]

New-to-old errors are thus 85.1% of that proposal's new-query errors. The certified rule then decides which of the proposal's new-class wins to keep, and a better conflict scorer raises certified utility by keeping more of the correct ones that a weaker scorer would reject. New-to-old errors made by the proposal itself are addressed by the proposal, which is why Table 3 pairs the learned ranker with BiMC's stronger proposal.

As for the score distributions, a single shared offset between old- and new-class scores is exactly the Fixed-Margin control (Appendix D). Under the same certificate it reaches 76.23% against 82.69% for CAPE-BF, a 6.46pp gap [5.89, 7.06], so per-candidate conflict ranking retains considerably more utility than a global score shift. The decomposition comes from the stored scores of the BiMC proposal; nothing was rerun.

W3 — The certificate covers the calibration frame, not new queries.

The guarantee is on the frame, and on held-out queries the control carries over well. Theorem 1 bounds the design-weighted old-to-new rate on the finite calibration frame; held-out rates are empirical, and §4 now states this explicitly. On FGVC Aircraft, CAPE-BF and CCPL route 0.82% and 0.92% of held-out old queries to new classes, against 7.54% for direct insertion (§7.3); our reply to Reviewer DU4A (Q1) gives the per-registration distribution. The FP budget and the aligned scorer keep exactly this constraint: same estimand, ε = 2% and 95% level (Corollary 2).

AuthorsOpenReview ↗

Reply to reviewer Ztey (part 2 of 2)

W4 — Smaller old-class calibration budgets.

The budget needed for each guarantee follows in closed form from the exact bound. At α = 0.05, the Clopper–Pearson rule of Eq. (1) needs n ≥ 299, 149 and 59 anchors to certify a registration with zero sampled violations at ε = 1%, 2% and 5%. At ε = 2%, the admitted violations k_max grow with the budget as follows; these values are arithmetic from Eq. (1) and involve no experiment:

Anchors n149–235236400530 (ours)8001,000
k_max (CP)0135912

The design-exact FP budget of the revision uses the actual stratified design and admits more at the same guarantee: 6–8 instead of 5 on the Aircraft designs and a median of +3 per exactly resolved development episode, so the same anchors buy more certified utility. A fixed budget also keeps the labelling and scoring cost of each registration bounded as old classes accumulate; "few-shot" refers to the new classes.

Q3 / W5 — What does CCPL add beyond pooling supervision across registrations?

The contribution is a conflict ranking learned across registrations and certified separately on every new one, and the controls show that shared estimation is what improves retained utility. CAPE-BF must estimate its scorer from the few examples of one registration; CCPL estimates it once from many registrations and certifies it on each new one. The controls separate this choice from the factors that change with it:

Control (same certificate)ResultWhere
CCPL-Linear, same cross-registration itemsmatches CCPL-MLP: +0.44 vs. +0.43pp over CAPE-BF§7.2
Episode-local arbiter given the full 17-D inputCCPL still +0.91pp [0.57, 1.25] ahead (earlier take)App. D
Frozen margin, no learning6.46pp [5.89, 7.06] below CAPE-BFTable 7
Shuffled training labelsreal scorer 9.05pp [8.49, 9.62] ahead§7.5
Same BiMC proposal, learned vs. fixed weights+1.82pp [1.56, 2.09]App. L

So the gain needs real cross-registration supervision, does not need network capacity, and transfers to a stronger proposal.

W6 — The gains are small and inconsistent across populations.

The gain grows on unseen populations, and on a strong proposal it is large and consistent. On development, vanilla CCPL improves over CAPE-BF by +0.43pp [0.16, 0.71], and by +0.89pp on the 672 registrations where CAPE-BF can be fitted; the three negative population estimates are small and their intervals include zero. Without retraining it gains more on held-out populations: +0.88pp on tieredImageNet and +3.27pp on FGVC Aircraft, and the advantage rises from +1.26 to +3.97pp as local fitting evidence shrinks (§7.5). On Aircraft under the matched final draw and CP budget, frozen vanilla CCPL reaches 35.04% against 29.37% for BiMC-CP (+5.67pp; Table 2, Appendix K.6; under the earlier draw, vanilla CCPL-FP reaches 38.18%). With the proposal held fixed (Table 3), aligned CCPL improves on BiMC by +2.39pp [2.16, 2.63], positive on all six development populations, and by +10.16pp [9.24, 11.08] on held-out Aircraft (§7.4).

Q4 / W7 — Behaviour at other shot counts.

We measured it: the frozen vanilla CCPL scorer runs unchanged at 1, 2, 4 and 5 shots on FGVC Aircraft, and its certified new-class utility increases across the tested shot counts. Nothing is retrained: the 300 episodes, old-class anchors and CP certificate are the same, and only the number of support images per new class changes. Over all 300 episodes, new-class utility is 22.43% at 1 shot, 30.92% at 2, 34.65% at 4 and 35.53% at 5 (95% bootstrap CIs within ±1.7pp). Paired on the same episodes, 2 shots add +8.48pp over 1 shot [7.64, 9.31], 4 add +3.73pp over 2 [3.03, 4.46], and 5 add +0.88pp over 4 [0.41, 1.34]. The certificate is met in all 300 episodes at every shot count, and held-out old queries routed to new classes are 0.41%, 0.79%, 0.98% and 0.95% of 49,500. The 5-shot run reproduces the stored records of this evaluation in all 300 episodes. CAPE-BF's leave-one-out fit has no same-class support image at one shot and succeeds in 0/300 episodes (272/300 at two shots, 300/300 at four and five).

These runs use the original G1 calibration draw, which explains the difference between the 5-shot result (35.53%) and the final-draw result in Table 2 (35.04%). They evaluate the frozen CCPL scorer; we also record CAPE-BF fitting feasibility but do not report per-shot certified comparisons against CAPE-BF or BiMC. The 8- and 16-shot evaluations were not completed.

The Objective paragraph, the held-out routing results and the aligned results now state the setting and its value directly. We hope they make the contribution clear and that you will take them into account in your assessment.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer,

Thank you once again for your valuable comments on our submission. As the discussion phase is approaching its end, we would like to kindly confirm whether we have sufficiently addressed all of your concerns. Should there be any remaining questions requiring further clarification, please do not hesitate to let us know.

We sincerely look forward to your feedback.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response (part 1 of 2)

We thank all reviewers, and Reviewer DU4A in particular for checking Theorem 1 and recomputing k_max(530) = 5 and the n ≥ 149 requirement. Four questions came up in several reviews: how CCPL compares with the strongest pipeline we evaluated, whether the certificate costs too much utility, what CCPL adds beyond a classical certificate and a small scorer, and why old-to-new errors should be certified per registration. This response answers them.

Under the same strong proposal, learned conflict ranking adds a clear certified gain.

In a post-submission audit we found that the external-comparator threshold for episodes with no sampled old-to-new intrusion (k_B2 = 0) used τ = +∞ instead of the τ = −∞ that Algorithm 1 prescribes (134 BiMC and 64 LP++-NEW-only episodes; Appendix D). With the rule applied as specified, BiMC's certified U_new on development is 86.57% (previously 71.85%) and LP++-NEW-only's 68.42% (previously 61.44%); all CAPE-family results are unchanged. The corrected BiMC baseline provides a stronger test of our central contribution: does learned conflict ranking improve certified utility when the proposal and the certification budget are held fixed? BiMC's native insertion is already 6.30pp above direct insertion before certification (§7.2), and Table 3 answers the question directly. Table 3 holds BiMC's proposal, the episodes, the anchors and the threshold rule fixed, and changes only the scorer and the budget. Proposal-aligned CCPL retrains the same ranker, with the same capacity, objective and optimizer, on BiMC's candidates, pair labels and proposal-score features, and combines it with BiMC's own margin (Eq. 5). On the 900 development episodes:

Scorer (same BiMC proposal)CP budgetFP budget
BiMC native margin86.57%87.37%
Aligned CCPL, linear (18 parameters)88.81%89.45%
Aligned CCPL (5,377 parameters)89.12%89.76%
Aligned − native+2.55pp [2.32, 2.79]+2.39pp [2.16, 2.63]

The gain is positive on all six populations; on Caltech-B, where the proposal is already strong, it is +1.00pp [0.53, 1.59], and all 90 newly retained crossings there are correct. The gain comes from the learned weights: a fixed equal-weight rule on the same 17 features and the same composition recovers only +0.57pp, and the learned scorer stays 1.82pp ahead [1.56, 2.09] (Appendix L). Capacity matters little, since the 18-parameter linear ranker keeps +2.08pp [1.89, 2.28] of the gain.

The aligned protocol was selected on development-training episodes, so the development bank evaluates the final method after selection. The independent check is FGVC Aircraft, which no aligned development step had used. We froze the scorer, the composition weight, the standardizers and the features, fixed a shared BiMC proposal for both arms in advance (β = 0.8, the median development temperature), and evaluated once on 300 episodes. Aligned CCPL-FP raises U_new from 32.36% to 42.52% (+10.16pp, 95% CI [9.24, 11.08]). It is better on 272 episodes, worse on 18 and unchanged on 10; it newly retains 2,294 new-class predictions, 2,121 of them correct. Harm05 also falls, from 10.0% to 9.3%.

The design-exact budget reduces the utility cost of certification under the same guarantee.

The certificate's cost depends on how many violations a budget admits. The CP budget treats the anchors as one pooled binomial sample, but they are drawn class by class without replacement, which carries more information. The FP budget uses that design exactly: it computes the stratified hypergeometric tail, maximizes it over every unsafe allocation of violations across classes, and keeps the largest count whose tail stays below 5%. It certifies the same estimand at the same ε and 95% level (Corollary 2), and on a census design it returns the exact count ⌊εn⌋. On development it admits more violations than CP on all 753 episodes whose designs it resolves exactly (165 distinct designs; median +3), and on Aircraft 6–8 instead of 5 on every design. The other 147 development episodes, from partially censused Caltech designs that reach the solver caps, keep the CP budget. The result is more certified utility for every scorer: +0.80pp for BiMC, +0.64pp for aligned CCPL and +0.55pp for vanilla CCPL over the full bank, and +3.26pp [2.74, 3.78] on untouched DTD episodes. Budgets depend only on the design, so they are solved offline once and cached.

AuthorsOpenReview ↗

General response (part 2 of 2)

CCPL's core contribution is a conflict scorer learned once and certified on every registration.

CAPE separates conflict ranking from risk certification, so any scorer plugs into the same exact per-registration rule. CCPL estimates the scorer across registrations and reuses it without refitting, whereas CAPE-BF refits it inside every registration. The controls show that where the scorer is estimated determines how much utility survives certification. A linear scorer trained on the same cross-registration items matches the MLP (+0.44 vs. +0.43pp over CAPE-BF), so capacity is not the driver. Giving the episode-local arbiter the full 17-D input still leaves CCPL ahead (+0.91pp [0.57, 1.25] on development, +5.03pp on Aircraft; earlier-take diagnostic, Appendix D). Shuffling the training labels removes the advantage (real scorer 9.05pp ahead, [8.49, 9.62]), and it persists when the evaluated population is left out of training. The mechanism is estimation stability: reducing local fitting evidence from 1/2 to 1/8 makes episode-local rankings less stable, and on the 514 episodes fitted at every fraction, CCPL's advantage rises monotonically from +1.26 to +3.97pp (§7.5).

The gains are largest where reuse matters most, and per-registration certification keeps old classes safe.

On development, vanilla CCPL improves over CAPE-BF by +0.43pp [0.16, 0.71], and by +0.89pp on the 672 episodes where both arms are fitted. On unseen populations the gain grows without retraining: +0.88pp [0.44, 1.33] on tieredImageNet and +3.27pp [2.56, 3.99] on FGVC Aircraft (under the earlier calibration draw, vanilla CCPL-FP also reaches 38.18% on Aircraft). Given features, CCPL arbitrates in 0.72 ms per registration on a Jetson Nano (CAPE-BF: 8.02 ms). Certification is what makes registration safe: freezing the old classifier does not freeze its predictions, and direct insertion lowers old-class accuracy by more than 0.5pp in 57.3% of development episodes. On Aircraft it routes 7.54% of held-out old queries to new classes, with harm05 at 96.3%; under the certificate, CCPL routes 0.92% and CAPE-BF 0.82%, and CCPL's harm05 falls to 14.7%.

For reference, the key numbers behind this response are summarized below.

QuestionKey resultWhere
Learned scoring, same BiMC proposal, developmentaligned CCPL-FP 89.76% vs. BiMC-FP 87.37%: +2.39pp [2.16, 2.63]; positive on 6/6 populationsTable 3, App. K.5
Held-out check of the frozen aligned methodFGVC Aircraft 42.52% vs. 32.36%: +10.16pp [9.24, 11.08]; 272 better, 18 worse; harm05 10.0% → 9.3%Table 3, App. K.6
Learned vs. fixed weights, same features+1.82pp [1.56, 2.09] (FP); linear aligned ranker 89.45%App. K.5, L
Design-exact budgetmore admitted violations on 753/753 exactly resolved episodes (165 designs); +0.80 / +0.64 / +0.55pp for BiMC / aligned / vanilla CCPL; DTD +3.26pp§4, §7.4, App. K
Reusable vs. per-registration scoring, development+0.43pp [0.16, 0.71]; +0.89pp on the 672 episodes where both arms are fitted§7.2, App. D
Transfer without retrainingtieredImageNet +0.88pp [0.44, 1.33]; FGVC Aircraft +3.27pp [2.56, 3.99]Table 2
Capacity and inputCCPL-Linear +0.44pp vs. CCPL-MLP +0.43pp over CAPE-BF; 17-D episode-local arbiter still +0.91pp behind CCPL§7.2, App. D
Supervision and mechanismlabel-shuffled scorer 9.05pp [8.49, 9.62] below the real one; advantage +1.26 → +3.97pp as local evidence shrinks§7.5
Old-class protection, Aircraftheld-out old-to-new routing 7.54% (direct insertion) → 0.92% (CCPL); harm05 96.3% → 14.7%§7.3
Arbitration latency, Jetson Nano, given features0.72 ms (CCPL-MLP) vs. 8.02 ms (CAPE-BF)App. H
External comparators, shared threshold ruleBiMC 86.57%; LP++-NEW-only 68.42%Table 1, App. D

To summarize, CAPE gives every registration a finite-sample guarantee on its old-to-new rate, and it lets the scorer and the budget be improved separately. A conflict scorer learned across registrations improves on per-registration fitting on development and on two unseen populations. Aligned to the strongest proposal we evaluated and certified with the design-exact budget, it improves on that proposal by 2.39pp on development and by 10.16pp on held-out Aircraft, under the same guarantee. The reviewers' questions led to the fixed-proposal comparison, the design-exact budget and the held-out Aircraft evaluation, and these now carry the paper.

AuthorsOpenReview ↗

Closing Summary of Contributions and Practical Significance

Dear Area Chair and Reviewers,

Thank you for your careful reviews. We would like to close the discussion by highlighting the central contributions of our work and the new evidence demonstrating its practical value.

A new approach to reliable class registration. Freezing a classifier's parameters does not freeze its predictions. Adding new classes can redirect existing predictions even when the original classifier remains unchanged. We address this problem through certified class registration: maximizing new-class recognition utility while controlling old-to-new interference. Our CAPE framework provides a finite-sample certificate for each registration, keeping the calibration-frame old-to-new risk within 2% with 95% confidence.

Learn conflict ranking once, certify every registration. CAPE separates conflict ranking from statistical certification. Building on this framework, CCPL learns a reusable conflict scorer across registrations rather than fitting a new scorer each time. The learned ranking identifies promising new-class predictions, while registration-specific certification determines which can be retained. Our new design-exact finite-population (FP) budget further exploits stratified sampling to retain more useful predictions under the same statistical guarantee.

Consistent gains under a fixed proposal (+2.39pp dev, +10.16pp Aircraft). Our new controlled comparison isolates what learned conflict ranking adds when both the proposal and the certification budget are held fixed. With the protocol-adapted BiMC proposal and the same design-exact FP budget, aligned CCPL raises certified new-class utility from 87.37% to 89.76% (+2.39 percentage points, 95% CI [2.16, 2.63]) across 900 development episodes, with improvements on all six populations, recovering nearly half of the 5.04pp drop from BiMC's uncertified ceiling (92.41%). On 300 held-out FGVC Aircraft episodes (untouched by any aligned development step), the frozen aligned method jumps from 32.36% to 42.52% (+10.16 points, 95% CI [9.24, 11.08]), outperforming BiMC on 272 of 300 episodes.

Generalization, protection, and efficiency. The original CCPL scorer also transfers without retraining, improving over per-registration CAPE-BF by +0.88 points on tieredImageNet and +3.27 points on FGVC Aircraft. On Aircraft, the measured old-to-new routing rate on held-out queries falls from 7.54% under direct insertion to 0.92% under CCPL. Given precomputed visual features, vanilla CCPL's arbitration layer takes a median of 0.72 ms per registration on a Jetson Nano, compared with 8.02 ms for CAPE-BF.

Broader significance. CCPL establishes a practical way to combine conflict ranking learned across tasks with statistical guarantees tailored to each individual task. CAPE's modular design allows ranking and certification budgets to improve independently, as shown by our aligned ranker and finite-population analysis, giving pretrained recognition systems higher new-class utility without relaxing control of old-class interference.

We appreciate your consideration of our revised work and the opportunity to present these additional findings to the ICLR community.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.