acceptodds
Under review as a conference paper at ICLR 2027

OPSA: Orthogonal Phase-Scattering Adapter for Parameter-Efficient Fine-Tuning on Long-Tailed Data

Abstract

Parameter-efficient fine-tuning on a long-tailed dataset fails in a specific, measurable way. Gradients from majority classes reinforce one direction of the weight change over and over, that direction becomes the principal singular vector, and the rest are overridden, so minority-class behaviour is forgotten. We quantify this with the spectral imbalance , the ratio of the largest singular value of the update to the mean singular value, and show that alone upper-bounds how much of the update the weakest direction can retain. We then introduce the Orthogonal Phase-Scattering Adapter (OPSA), which represents every bottleneck activation as a quadrature pair. A normalized token is projected to a semantic amplitude, and its phase is generated from that amplitude by a fixed ninety-degree rotation, so the two streams are exactly orthogonal at every token and every step, with no penalty term and no extra parameters. The two streams then gate each other across the phase basin, which curves an otherwise flat response surface and applies a gain that falls as amplitude grows. We prove this contracts by exactly the ratio of the dominant gain to the spectrum-weighted mean gain. Because descent carries momentum across the basin floor and up the far wall, momentum-aware phase clipping resets that momentum on contact, which bounds the amplitude the state reaches. On a long-tailed DermaMNIST benchmark with a ViT-B/16 backbone, OPSA attains the best macro F1 against six parameter-efficient adapters, at a spectral imbalance of on the scale of Proposition 1, against – for every baseline — the lowest value in the comparison and a factor of below plain LoRA. On long-tailed 20 Newsgroups with a frozen DeBERTa-v3-base it attains the best tail and macro F1 against four adapters, leading the strongest baseline on tail F1 by points.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Parameter-efficient fine-tuning, Long-tailed learning, transfer learning, Low-rank adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Sb6XOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes OPSA, a nonlinear low-rank adapter designed to improve fine-tuning on long-tailed data. It introduces an orthogonal phase representation and momentum-aware clipping to reduce spectral concentration in adapter updates. The paper provides theoretical analyses and evaluates the method on long-tailed image and text classification tasks, with baseline comparisons and ablation studies.

Critical strengths
  1. The paper raises an interesting question about whether long-tailed data causes low-rank adapters to concentrate their updates in a few dominant directions. Using spectral imbalance to study this problem is a useful idea. OPSA also introduces a different adapter design that tries to address this issue without increasing the number of trainable parameters.

  2. OPSA reports lower spectral imbalance than the baselines and better macro F1 on both image and text classification. The improvement in tail F1 on 20 Newsgroups is particularly encouraging. Although the current experiments are not enough to establish the proposed mechanism, these results suggest that the nonlinear adapter design may be worth investigating.

Critical weaknesses
  1. The main theoretical explanation is flawed and does not match the experiments.

Theorem 3 in Appendix A.11 contains a serious problem. The proof uses the equality of two vector norms to argue that corresponding coordinates saturate together, which does not follow mathematically. A simple 2D counterexample, a=(π/2,0)a=(\pi/2,0), shows that a coordinate can reach the phase horizon while its derivative remains positive. This directly challenges the claimed gradient reversal mechanism.

More importantly, Table 2 shows that 99.7% of latent units are outside the theoretical basin by the end of training, with a maximum amplitude of 5.06 compared with the boundary of 1.57. Enforcing the basin in OPSA-v2 also leads to worse performance. Theorem 2 relies on gain-ordering assumptions that are not verified for the actual nonlinear adapter. As a result, it is unclear whether the theory explains the observed spectral improvements or the performance gains at all.

The claim of zero additional inference cost after merging is also unsupported. Unlike LoRA, OPSA contains input-dependent normalization and cosine nonlinearities, so its update cannot generally be merged exactly into a fixed weight matrix. No post-merging accuracy or inference-time results are provided.

  1. The experimental evidence is not reliable enough, and the code release does not allow full reproduction.

Appendix A.3 admits that the three-seed and key ablation experiments use a validation set overlapping with training for checkpoint selection. Even with the same imbalance ratio and seed, Tables 1 and 2 report macro F1 scores of 0.569 and 0.617. The difference is explained by different experimental protocols, but it is much larger than the main reported improvement over LoRA. This makes the ablation conclusions, especially the claimed importance of MAPC, difficult to trust without results under a consistent protocol.

Reproducibility is another major concern. The released repository does not contain the complete OPSA implementation, MAPC optimizer, training pipeline, or main checkpoints. The MAPC update rules are also inconsistent between the main text and appendix. The authors additionally state that the text checkpoints were not saved, so the central spectral metric cannot be recomputed for that benchmark.

These issues are especially concerning because the theoretical explanation is already inconsistent with the observed training behavior. Without reliable experiments and complete implementation details, it is difficult to independently verify the paper's main claims.

Questions for authors
  1. How do the authors justify the gradient reversal mechanism in Theorem 3? The equality ∥a∥=∥Ja∥\|a\|=\|Ja\| does not imply coordinate-wise saturation. For example, with a=(π/2,0)a=(\pi/2,0), the first coordinate reaches the phase horizon, but ∂z1/∂a1=1>0\partial z_1/\partial a_1=1>0. How does Theorem 3 account for this case? Also, given that nearly all latent units leave the theoretical basin during training, what evidence shows that the claimed gradient reversal and spectral contraction mechanisms actually explain the experimental results?

  2. Can the authors provide the key results under a consistent experimental protocol? In particular, does removing MAPC still cause a similar performance drop when checkpoint selection uses a validation set fully disjoint from training? How much of the difference between Tables 1 and 2 comes from the different splits and checkpoint-selection procedures? The complete OPSA and MAPC implementations would also be necessary to verify these results, especially given the inconsistent MAPC definitions in the paper.

  3. How exactly is OPSA merged for inference? Since the adapter includes input normalization and cosine nonlinearities, its output is not generally equivalent to a fixed linear weight update. Is the proposed merging exact or approximate? If it is approximate, what are the changes in prediction performance and inference latency after merging? This is important for evaluating the paper's efficiency claims.

Reviewer_PUKBOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Paper identifies a key failure mode of linear low‑rank adapters under long‑tailed data: gradients from majority classes inflate the leading singular value of update matrix ΔW\Delta W, producing large spectral imbalance I(ΔW)\mathcal I(\Delta W), which squeezes out signal for minority‑class directions. The authors formalise this phenomenon with propositions bounding information retention of weak directions. They propose OPSA, Orthogonal Phase‑Scattering Adapter. It constructs two orthogonal streams via fixed 90‑degree rotation without extra trainable parameters, applies cross‑cosine gating to attenuate over‑amplified directions, plus MAPC: a selective Adam first‑moment reset to counteract momentum punch‑through past phase horizon. Up‑projection uses norm‑fixed reparameterization avoiding manifold retraction. OPSA shares exactly the same number of trainable parameters as LoRA, and adds no inference overhead after adapter merging.

Critical strengths
  1. Most existing long‑tail remedies target loss or sampling; this paper instead analyses adapter update spectral geometry, proposes interpretable metric I(ΔW)\mathcal I(\Delta W) and provides formal propositions for how majority‑class gradients exponentially inflate singular‑value spread in factorized adapters. This is a valuable conceptual analysis for the PEFT community.
  2. Orthogonal phase‑scattering bottleneck uses a fixed skew‑symmetric rotation J to generate phase stream for free, achieving exact per‑token orthogonality without penalty terms or extra parameters. Cross‑gating design naturally attenuates large‑amplitude directions inflated by majority‑class gradients. MAPC addresses a subtle Adam momentum punch‑through failure which plain gradient clipping cannot fix.
  3. Identical parameter count to LoRA; zero overhead at inference after merging adapter weights; only moderate extra wall‑clock cost during training (9‑47%). It avoids expensive Riemannian manifold retraction used by alternatives like StelLA.
Critical weaknesses
  1. Critical theory‑practice mismatch (major conceptual concern) All core theorems and guarantees for spectral contraction, bi‑Lipschitz stability hold **only inside the phase basin. However in all main experiments, nearly all latent coordinates drift far outside this proven safe region. Even without MAPC, spectral imbalance remains low. This means the observed beneficial spectral‑imbalance reduction cannot be fully explained by the presented theory. The paper observes the effect but does not provide complete theoretical account for why it persists beyond basin boundaries. This is not a trivial footnote; it weakens the explanatory power of the theoretical framework.
  2. Limited experimental scope & reliability caveats
    • Only two benchmarks: one small medical image dataset and one text dataset. No larger‑scale LLMs, no standard long‑tail NLP benchmarks (e.g. LT‑GLUE), no multi‑modal settings. All results run on base‑size backbones; it is unknown whether spectral‑imbalance mitigation and performance gains transfer to larger model scales.
    • Text benchmark only uses single seed, and spectral imbalance metric cannot be reliably computed for text runs; all mechanism evidence comes solely from the image dataset.
    • Tail‑F1 on DermaMNIST suffers from extremely small test sample size for minority class, yielding very wide bootstrap confidence intervals. Macro‑F1 and spectral imbalance are more trustworthy, but tail‑F1 improvements are statistically noisy.
    • Missing real‑world downstream settings: continual fine‑tuning, adapter merging, LoRA‑type compositional adapters, which the authors flag only as future work.
  3. MAPC behaviour is incompletely understood Ablation shows removing MAPC causes large macro‑F1 drop (~4.7 points), yet spectral imbalance stays low. This decoupling indicates:

Spectral contraction originates from orthogonal phase‑scattering bottleneck itself, but MAPC is critical for downstream task performance for reasons not fully captured by current propositions. MAPC only erases momentum buffer, and in practice coordinates still drift well past phase horizon. The paper lacks deeper diagnostic to explain why MAPC yields large task gains even when states are outside the basin.

  1. Insufficient comparative positioning against existing non‑linear PEFT works The paper briefly cites Sinelora, CeRA, HypLoRA, but lacks systematic ablation / comparison showing what OPSA brings over those non‑linear bottleneck adapters on identical long‑tail protocols. It is hard to judge whether spectral‑imbalance reduction is uniquely delivered by OPSA or can also be partially reproduced by prior non‑linear PEFT variants.
  2. Limited analysis of failure modes. No case / error analysis showing when OPSA still fails on hardest minority samples. We do not know whether remaining errors stem from data noise, the adapter bottleneck, or optimization limits.
Questions for authors

Please see my comments in weakness.

Reviewer_w2RsOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

Paper Summary

The paper introduces OPSA—Orthogonal Phase-Scattering Adapter, a parameter-efficient fine-tuning method designed for classification on long-tailed datasets. Its central motivation is that majority-class gradients can concentrate a low-rank adapter’s update into a small number of dominant singular directions, potentially limiting adaptation for minority classes. The authors quantify this concentration using spectral imbalance,

I(ΔW)=σ11r∑i=1rσi,\mathcal I(\Delta W)=\frac{\sigma_1}{\frac{1}{r}\sum_{i=1}^{r}\sigma_i},

where σi\sigma_i are the update’s singular values and rr is the comparison dimension. For a nonzero update, larger values indicate greater spectral concentration.

Critical strengths
  1. The paper addresses a meaningful problem: how the structure of a parameter-efficient adapter affects adaptation on long-tailed data, particularly for underrepresented classes. Its use of spectral imbalance provides a concrete quantity through which to investigate concentration in the learned update. Reporting macro F1, head-class performance, and tail-class performance also makes the evaluation responsive to the stated objective. The proposed relationship between adapter geometry and class-specific behavior is a useful research direction, even though its causal interpretation requires further evidence.

  2. OPSA introduces a specific nonlinear modification to the low-rank bottleneck: a fixed quarter-turn operator generates a companion activation stream, and cross-coordinate cosine gates couple the two streams. The phase transformation introduces no additional trainable parameters, while the down- and up-projections retain the 2rd2rd parameter count of a rank-rr LoRA adapter for a square projection. The released parameter counts confirm equality between OPSA and ordinary LoRA. This makes the proposed architecture interesting as a way to change an adapter’s functional behavior within the same trainable parameter budget.

  3. The paper contains verifiable mathematical components. For even bottleneck dimension, the fixed operator guarantees that the activation and phase vectors are orthogonal and have equal norms. The pairwise output-energy identity is correct, and the local Jacobian analysis provides an explicit sufficient region for injectivity. The normalized up-projection also has a correctly derived gradient, and a residual-norm bound follows under the specified activation constraint. These results provide a useful analytical foundation and identify concrete conditions under which aspects of the nonlinear map can be characterized.

  4. The reported results provide initial support for investigating the method further. Under the corrected image protocol, OPSA improves macro F1 from 0.551 for LoRA to 0.569, while matching the best reported accuracy of 0.677. On 20 Newsgroups, it achieves macro F1 of 0.556, compared with 0.543 for rsLoRA, and tail F1 of 0.317, compared with 0.210. Observing improvements with both a vision transformer and a text transformer gives the proposal broader initial support than a single-backbone demonstration. These are encouraging point estimates, with their statistical scope limited by the experimental scale.

  5. The supporting release enables meaningful checks of the reported arithmetic. Independent recalculation from the six supplied per-seed confusion matrices reproduces the corresponding macro F1, accuracy, tail F1, pooled counts, means, and sample standard deviations in Table 5. The stored per-layer spectral values also reproduce their reported summary statistics. Providing these intermediate results improves transparency and allows readers to examine how aggregate scores arise from class-level outcomes.

  6. The paper reports task metrics alongside spectral imbalance, activation magnitudes, basin-exit fractions, and computational costs. This combination is useful for examining whether the observed behavior matches the proposed explanation. The manuscript also explicitly discloses the small image-tail test set, the older studies’ overlapping training and checkpoint-selection validation data, missing corrected text spectra, and the fact that reported activations largely leave the theoretical basin. These disclosures make the evidence easier to assess and identify concrete priorities for subsequent validation.

Critical weaknesses
  1. Theorem 2 is valid under restrictive assumptions: the nonlinear update must preserve the linear update’s singular directions, apply positive scalar gains to those directions, attenuate the leading direction most strongly, and preserve singular-value ordering. The paper does not establish that OPSA’s token-dependent, cross-coordinate cosine gate satisfies these conditions. In particular, activation amplitudes are not automatically ordered by the singular values of the complete update.

For instance, with κ=1\kappa=1, inputs (±1,±1)(\pm1,\pm1), Wdown=diag⁡(1,0.2)W_{\mathrm{down}}=\operatorname{diag}(1,0.2), and an up-projection proportional to (I+J)−1(I+J)^{-1}, every activation remains inside the proposed basin. Nevertheless, spectral imbalance increases from approximately 1.667 to 1.801, with an exact linear fit on those inputs.

Thus, the statement that basin membership is sufficient for contraction is false. The authors must either establish additional conditions connecting the actual architecture to the theorem or substantially narrow the claimed guarantee.

Theorem 3(a) claims that the transmitted derivative reverses near the phase horizon and drives activations inward. The proof incorrectly infers coordinatewise equal magnitudes from the vector identity ∥a∥=∥Ja∥\|a\|=\|Ja\|. Moreover, even when a pair saturates together, the relevant derivatives do not both reverse. For u=v=s↑H=π/2u=v=s\uparrow H=\pi/2,

\qquad \frac{\partial z_2}{\partial v}\rightarrow -H.$$ Actual optimization direction also depends on the full Jacobian and upstream loss gradient. A binary-cross-entropy example produces an outward gradient step across the horizon, contradicting universal restoration. Theorem 3(c)’s assertion that cosine factors remain negative beyond the horizon is independently false because cosine is periodic. The conditional residual bound cannot provide the training-independent protection claimed elsewhere. The gate is globally unbounded along $a=(t,0)$, where $z=(t,t)$, and Table 2 reports **99.3–100% basin-exit rates** in the measured runs. The authors acknowledge these exits, but the main mechanistic conclusions remain stronger than the applicable theory. 3. Proposition 1(b) provides an **upper bound** on the weakest-to-strongest singular-value ratio and a **lower bound** on the leading direction’s energy share: $$\frac{\sigma_r}{\sigma_1} \le \frac{r-\mathcal I}{\mathcal I(r-1)}, \qquad \frac{\sigma_1^2}{\sum_i\sigma_i^2} \ge \frac{\mathcal I}{r}.$$ The paper interprets a larger permissible upper bound as **“4.1× more retention.”** That conclusion does not follow: relaxing an upper bound does not establish an increase in the actual quantity. Similarly, at $\mathcal I=2.96$ and $r=16$, the second inequality gives a leading-energy share of **at least 18.5%**, whereas the manuscript describes OPSA as being “held to” that value. More fundamentally, update singular values contain no class labels. The paper does not establish that minority-class information occupies the weakest update directions or that reducing $\mathcal I$ preserves that information. Class-conditioned measurements are needed to support this interpretation. 4. Equation (7) sets the first moment to zero when a reset occurs, whereas Proposition 2(c) states that the resulting moment is $(1-\beta_1)g_t$. These correspond to different update orders: resetting after the current gradient is incorporated versus clearing historical momentum before incorporating it. The reset mask is also underspecified. Activations are indexed by latent coordinates and tokens, while the gradient and momentum of $W_{\mathrm{down}}$ are indexed by latent and input coordinates. The aggregation and broadcasting operation connecting these objects must be defined. Furthermore, disagreement between gradient and momentum signs does not identify which direction is outward. Resetting can remove an inward-pointing historical moment and leave an outward-pointing current gradient. Consequently, the claimed reduction in activation magnitude relative to ordinary Adam is not a general consequence of the stated rule. The withheld optimizer implementation prevents these ambiguities from being resolved directly. 5. The manuscript explicitly states that the three-seed, cost, and ablation studies select checkpoints using validation data overlapping training. Only the corrected main image study and the separately described text protocol use disjoint validation. The test set is reported as untouched, so this should not be characterized as train/test leakage. Nevertheless, the older replicated and ablation results do not directly validate the corrected protocol, and the central comparisons require clean replication. The uncertainty is substantial. The reported three-seed paired statistic $t\approx2.66$, with two degrees of freedom, corresponds to a two-sided $p$-value of approximately **0.117**. Under that model, the 95% interval for the mean macro-F1 improvement is approximately **$[-0.028,\;0.120]$**. This supports a promising estimate, but does not justify strong confidence in the effect’s magnitude. The image tail contains only **nine test records**. Pooling predictions across three seeds produces 27 predictions on those same records, not 27 independently sampled cases. The reported pooled uncertainty does not demonstrate adequate treatment of this dependence. More independent tail examples, corrected replicated runs, and uncertainty estimates preserving image and method pairing are needed. 6. The ablations vary MAPC, the bottleneck, and the reset margin, but do not isolate input normalization, fixed-radius up-projection, the cross-coordinate gate, or output scale. The statement that quadrature coupling is the only structure shared by the ablation arms is therefore inaccurate. A scale-matched linear adapter and a conventional nonlinear bottleneck are particularly important controls. The norm–frequency relationship motivating normalization is also assumed rather than measured on either backbone. Consequently, the contribution of this component remains unclear. Existing work on sine-transformed low-rank weights and nonlinearity retained at inference makes precise differentiation important. OPSA’s potential novelty lies in its particular coupling and optimization procedure, but the current experiments do not show which component accounts for the gains relative to these alternatives.
Questions for authors

Please address weaknesses

Comments on the paper

1 comment
AuthorsOpenReview ↗

We re-ran every method under one clean protocol and added ViT-L/16, HAM10000 and OL3I Heart CT. OPSA has the lowest spectral imbalance of eight adapters and is clearly ahead at the most severe imbalance ratio (150). We corrected Theorems 2 and 3, added results that explain where the contraction comes from, and withdrew the claims that did not hold. The camera-ready version will add multi-seed runs for every method and dataset, the full text ablations, and the code release.

We thank reviewers w2Rs, PUKB and Sb6X. Below: what we ran, corrected and will add (I\mathcal{I}: spectral imbalance, lower is better).

1. New experiments

  • Clean protocol. All image methods re-run on DermaMNIST with disjoint validation at ρ=100\rho=100 and 150150 (three seeds for key arms), under three checkpoint rules.
  • Published adapters. Of eight adapters (incl. HypLoRA, SineLoRA), OPSA has the lowest I\mathcal{I}: 2.632.63 vs 3.653.65–8.958.95 (2.552.55 at ρ=150\rho=150).
  • Scale. ViT-L/16: I\mathcal{I} 1.991.99 vs 7.817.81 (LoRA); macro-F1 .609.609 vs .566.566; tail F1 .471.471 vs .154.154.
  • New datasets. HAM10000 (9 methods): OPSA's rank rises 6th → 4th → 1st at ρ=50/100/150\rho=50/100/150. OL3I Heart CT (22:1): OPSA has the lowest macro-F1 (.399.399) by over-predicting the minority class (recall .811.811, precision .070.070), yet its AUC equals LoRA's (.728.728), so the failure is the operating point; OPSA-v2 is best (.566.566, AUC .761.761).
  • Severe imbalance. At ρ=150\rho=150, OPSA leads all published adapters (macro-F1 .589.589 vs LoRA .507.507, rsLoRA .519.519); paired image-clustered bootstrap +.082+.082 [.020,.146.020,.146]. At ρ=100\rho=100 the interval includes zero, and we drop the t=2.66t=2.66 statistic.
  • Component ladder (I\mathcal{I}). LoRA 8.838.83 → normalization 8.098.09 → fixed-radius linear up-projection 3.413.41 → tanh 2.582.58 / self-gated cosine 2.792.79 / OPSA gate 2.672.67.
  • MAPC. Removing it costs 4.94.9, 1.61.6, 17.917.9 macro-F1 points across seeds; training collapses by the final epoch (.611→.365.611\to.365, p=.003p=.003; also at ρ=150\rho=150); RMS amplitude 1.251.25 vs 3.263.26. Halving the learning rate avoids collapse but ends lower (.597.597).
  • Loss-matched controls. Image: OPSA .569.569 vs LoRA with weighted CE .525.525. Text: .556.556 vs .531.531 (tail .317.317 vs .166.166).
  • Diagnostics. 89% of token activations stay in the basin (81% at the final epoch); σr/σ1\sigma_r/\sigma_1 .185.185 vs .013.013 for LoRA; the Theorem 2 identity predicts I\mathcal{I} within 7–13%; 5 of the 23 official-split tail images are missed by all 36 runs of 22 methods.
  • Deployment. A least-squares merged surrogate runs at base-model speed and costs 1.41.4 macro-F1 points (3 seeds; 5.75.7 at the final epoch); unmerged OPSA is 17–59% slower than merged LoRA (compiled).

2. Theory corrections

  • Theorem 2 is now an exact identity, IOPSA=Ilin g1/gˉσ\mathcal I_{\mathrm{OPSA}}=\mathcal I_{\mathrm{lin}}\ g_1/\bar g_\sigma, reproducing w2Rs's counterexample (1.8011.801); "basin implies (H1)" is withdrawn.
  • Theorem 3: co-saturation is an explicit hypothesis; non-vanishing horizon derivative (1.571.57 vs 0.160.16 for tanh); local result; 3(c) is restricted to (π/2κ,3π/2κ)(\pi/2\kappa,3\pi/2\kappa).
  • New results: an input-uniform bound ∥Δx∥≤γ2d ∥Wdown∥2\Vert\Delta x\Vert\le\gamma\sqrt{2d}\ \Vert W_{\mathrm{down}}\Vert_2; a basin-free proposition (with a fixed-radius WupW_{\mathrm{up}}, a˙c\dot a_c is independent of aca_c, so I(t)\mathcal{I}(t) settles rather than growing to CC); a Gaussian-pair gain proposition (OPSA's cross gain never changes sign; the self-gated gain does at s=1/κs=1/\kappa).
  • MAPC: Eq. 7 corrected (mt=(1−β1)gtm_t=(1-\beta_1)g_t); one algorithm specified; Prop. 2 keeps the buffer bound and one-step erasure.
  • Retractions/disclosures: Definition 1 withdrawn; the exact-merging claim (A.4) retracted; weighted CE and the 4×4\times learning rate on WdownW_{\mathrm{down}} disclosed.

3. Points addressed

  • w2Rs: W1 (H1 counterexample), W2 (Theorem 3, bound), W3 (weak-direction retention), W4 (MAPC), W5 (statistics, recipe disclosure), W6 (source of the contraction, novelty).
  • PUKB: W1 (basin), W2 (scope: ViT-L/16, new datasets, text, tail statistics, merging), W3 (MAPC vs I\mathcal{I}), W4 (HypLoRA, SineLoRA, tanh, self-gated), W5 (failure analysis).
  • Sb6X: Q1 (Theorem 3, 99.7% exit rate, theory vs results), Q2 (consistent protocol, MAPC under disjoint validation), Q3 (merging), reproducibility.

4. Camera-ready commitments

  1. Code release in our existing repository: the MAPC optimizer as specified above, the full pipeline, configs and checkpoints.
  2. Multiple runs for every arm: ≥3 seeds for all methods and datasets (incl. ρ=150\rho=150, HAM10000, OL3I), with mean ± std.
  3. Text modality: a checkpointed re-run reporting text I\mathcal{I}, plus the full text ablations (ladder, MAPC, gate variants) over ≥3 seeds under the same disjoint-validation protocol.
  4. Recipe controls: OPSA and LoRA with swapped loss and learning-rate recipes, and loss-matched per-class confusion matrices.
  5. OL3I: validation results, with AUC reported alongside macro-F1.
  6. CeRA under the same protocol.
  7. Revisions: Theorems 2–3 and proofs, Eq. 7, corrected numbers, the merging retraction; continual and compositional settings as future work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.