research/Reproduction

6,673 records · yaml · accepts undeclared fields

Fields

Rules

• reproductionNameRule
• reproductionVerdictRule
• reproductionUrlRule
• reproductionWhitespaceRule

Defects

Checked 200 of 6,673 records (the per-store cap is 200), so this is a sample, not a whole-store verdict.

Showing the first 200 of 400 defects.

Example record

/Users/eshao/.config/mnt/data/autoresearch-reproduction/repros/abbas-2026-quantumboost-a-lazy-yet-fast-quantum-algorithm-for-learning-with-weak-hypotheses--the-kristina-dimitrova.yaml

slug: abbas-2026-quantumboost-a-lazy-yet-fast-quantum-algorithm-for-learning-with-weak-hypotheses--the-kristina-dimitrova
paper:
  id: paper/abbas-2026-quantumboost-lazy-yet-fast
reproducer:
  url: https://huggingface.co/the-kristina-dimitrova
stub: true
attempts: []
claims:
- id: claim-1
  statement: QuantumBoost achieves overall runtime Õ(W/(√ϵ·γ⁴)) to reach empirical error at most ϵ, where W is the weak-learner
    runtime and γ is the edge of the weak hypotheses (Theorem 1, Section 4).
  verdict: partial
  evidence:
    sources:
    - judge-verdict
    judgeVerdict: toy
    summary: The logbook tests only the quantum amplitude estimation subroutine in isolation on a Bernoulli(p) mean-estimation
      task (slope -0.981 vs theory -1.0), not the full QuantumBoost algorithm's runtime. This verifies one ingredient of the
      speedup mechanism on a proxy task, not the overall runtime Õ(W/(√ϵ·γ⁴)).
- id: claim-2
  statement: With training set size m = Θ((d·log(d/(δϵ)) + log(1/δ))/ϵ²), QuantumBoost outputs a hypothesis with generalization
    error ≤ ϵ and success probability 1−δ (Theorem 2, Section 4).
  verdict: partial
  evidence:
    sources:
    - judge-verdict
    judgeVerdict: toy
    summary: The generalization bound is tested using classical AdaBoost (not QuantumBoost) on synthetic realizable Gaussian
      data with small dimensions (d∈{2,5,10}) and only two ε values. This verifies a standard PAC/VC-dimension sample complexity
      form on a simplified proxy, not QuantumBoost's actual generalization guarantee.
- id: claim-3
  statement: QuantumBoost uses T = O(log(1/ϵ)/γ²) boosting iterations, matching the iteration count of classical AdaBoost
    and Kale's SmoothBoost while achieving strictly better total runtime (Theorem 20, Theorem 23, Section 4).
  verdict: partial
  evidence:
    sources:
    - judge-verdict
    judgeVerdict: toy
    summary: The iteration count is tested with an idealized edge-gamma weak-learner oracle (exact 2γ edge by construction)
      on a 5×5 grid. While the γ-exponent matches well (~2.05), the log(1/ε) exponent is substantially off (1.82 vs 1.0),
      and the setup uses a synthetic oracle rather than a real weak learner.
- id: claim-4
  statement: QuantumBoost's total runtime of W/(√ϵ·γ⁴) improves on Quantum SmoothBoost's runtime of W/(ϵ^2.5γ⁴) + √d/(ϵ^3.5γ⁵)
    and on Quantum AdaBoost's runtime of W^1.5√d/(ϵγ¹¹) (Table 1).
  verdict: partial
  evidence:
    sources:
    - judge-verdict
    judgeVerdict: inconclusive
    summary: 'The check is purely algebraic: evaluating closed-form Big-O formulas across a parameter grid and confirming
      one expression is numerically smaller than another. This is trivially true by inspection of the exponents (e.g., √ϵ
      < ϵ^2.5 for small ϵ) and does not constitute experimental evidence or verify the underlying runtime analyses.'
- id: claim-5
  statement: QuantumBoost is presented as the first algorithm, classical or quantum, to combine a lazy Bregman-projection
    strategy—projecting only every K = 1/γ iterations rather than every iteration—with quantum mean estimation for boosting
    (Section 4).
  verdict: partial
  evidence:
    sources:
    - judge-verdict
    judgeVerdict: inconclusive
    summary: The core novelty claim ('first to combine lazy Bregman projection with quantum mean estimation for boosting')
      is a literature claim the logbook explicitly acknowledges it cannot test. The proxy test (K=1/γ is load-bearing) uses
      a toy setup with a modeled edge-degradation weak learner and does not address the novelty assertion.
verdict: partial
verdictSummary: The logbook provides toy-level numerical evidence for Claims 1–3 by testing simplified components (amplitude
  estimation microbenchmark, classical AdaBoost on synthetic data, simulated boosting with idealized oracle) and no meaningful
  evidence for Claims 4–5 (algebraic formula checks and an untestable novelty claim). No claim is verified at the full algorithm
  or non-trivial scale.
sources:
- id: logbook
  url: https://huggingface.co/spaces/the-kristina-dimitrova/repro-quantumboost-a-lazy-yet-fast-quantum-algorithm-for-learning-with-weak-hypotheses
  checkedOn: '2026-08-11'
  holds: Public Trackio reproduction logbook
- id: judge-verdict
  url: https://huggingface.co/datasets/ICML-2026-agent-repro/verdicts
  checkedOn: '2026-08-11'
  holds: Independent Logbook Judge verdict and per-claim evidence

Where it lives