Beyond Logit Adjustment: A Residual Decomposition Framework for Long-Tailed Reranking

Dimensions overall score 7.0

• Existing logit adjustment methods using fixed classwise offsets cannot adapt to input-dependent variation in the correction needed for optimal reranking.limitation 90%
  Evidencepartial

  the correction required to restore the relative ranking of two classes need not be constant across inputs, and a fixed offset cannot adapt to such variation

  Implicationpartial

  Directly stated in abstract with clear contrast to proposed method

  Verificationpartial

  partial

• The residual correction between optimal score and base score decomposes into a classwise component constant within each class and a pairwise component dependent on input and competing labels.technical 90%
  Evidencepartial

  The gap between the optimal score and the base score, the residual correction, decomposes into a classwise component that is constant within each class, and a pairwise component that depends on the input and competing labels

  Implicationpartial

  Explicitly stated as a core theoretical finding in the abstract

  Verificationpartial

  partial

• When the same label pair induces incompatible ordering constraints across contexts, no fixed offset can achieve Bayes-optimal ordering recovery.technical 85%
  Evidencepartial

  when the same label pair induces incompatible ordering constraints across contexts, no fixed offset can achieve this recovery

  Implicationpartial

  Directly stated theoretical result with clear implication for method limitations

  Verificationpartial

  partial

• REPAIR (Reranking via Pairwise residual correction) combines shrinkage-stabilized classwise term with linear pairwise term driven by competition features on the shortlist.method 90%
  Evidencepartial

  We develop REPAIR (Reranking via Pairwise residual correction), a lightweight post-hoc reranker that combines a shrinkage-stabilized classwise term with a linear pairwise term driven by competition features on the shortlist

  Implicationpartial

  Direct description of the proposed method in abstract

  Verificationpartial

  partial

• Experiments on five benchmarks spanning image classification, species recognition, scene recognition, and rare disease diagnosis confirm the decomposition explains where pairwise correction helps.result 85%
  Evidencepartial

  Experiments on five benchmarks spanning image classification, species recognition, scene recognition, and rare disease diagnosis confirm that the decomposition explains where pairwise correction helps and where classwise correction alone suffices

  Implicationpartial

  Direct statement of experimental validation across multiple domains

  Verificationpartial

  partial

• When residual is purely classwise, a fixed offset suffices to recover Bayes-optimal ordering.technical 85%
  Evidencepartial

  When the residual is purely classwise, a fixed offset suffices to recover the Bayes-optimal ordering

  Implicationpartial

  Direct theoretical statement about conditions where simple methods work

  Verificationpartial

  partial

• The decomposition leads to testable predictions regarding when pairwise correction can improve performance and when it cannot.result 80%
  Evidencepartial

  This decomposition leads to testable predictions regarding when pairwise correction can improve performance and when cannot

  Implicationpartial

  Direct statement about predictive power of the theoretical framework

  Verificationpartial

  partial

Competitive landscape

A lightweight post-hoc reranker that decomposes and corrects for biases in long-tailed classification, improving accuracy on rare classes.