1.59 MB
53 files
Updated about 1 month ago
Name
Size
evidence
outputs
poster
reproduction
source
PROTOCOL.md3.56 kB
xet
PROVENANCE.md1.59 kB
xet
README.md2.48 kB
xet
requirements.txt63 Bytes
xet
README.md

Parameter-free Dynamic Regret — exact 6/6 CPU candidate

Source-locked reproduction of OpenReview xnj5CyzVVW / arXiv 2602.06902, “Parameter-free Dynamic Regret: Time-varying Movement Costs, Delayed Feedback, and Memory.” The conservative forecast is 6/6; this is not an official verdict.

  • Algorithms 1–3 are independently transcribed from the paper's displayed updates. The closed-form radial Composite Mirror Descent step passes 2,048 KKT/subgradient checks with maximum residual below 1.3e-13.
  • The exact Algorithm 2 exponential grid contains 8–11 parallel instances across the operative matrix. Algorithm 3 makes 200 source-prescribed threshold updates rather than 20,000 per-round base updates in the batching stress test.
  • The movement-cost matrix covers 1,092,000 CPU action rounds, 37 configurations, dimensions 2/5/9, and a 420,000-round long trace. All traces are finite and comparator-independent.
  • Theorem 4.1's first-order term is separated from the quadratic predecessor by 999x when gradients are small relative to movement costs. Direct scale certificates recover square-root exponents in horizon, comparator path, and cumulative movement.
  • With every lambda_t=0, Algorithm 3 is bit-identical to direct Algorithm 2 for 64,000 rounds; all 64,000 updates trigger, movement cost is exactly zero, and both static (P_T=0) and dynamic (P_T>0) comparator cases use the same action trace.
  • The delay wrapper processes genuinely out-of-order feedback, passes the exact Lemma 5.1 reduction and 2*d_tot certificate, and exhibits an 89.8x separation between d_max*T and d_tot.
  • The memory wrapper uses the exact time-varying xi_t, passes the Lemma 5.6 reduction, verifies xi_t<=B^2 and sum xi_t<=sum b_t^2, and exhibits a 66.4x adaptive-memory separation from B^2*T.
  • Nine destructive controls, 53/53 science gates, 85/85 independent tests, and a full byte-exact replay pass.
  • Independent NumPy paper-equation implementation; CPU only; no GPU/MPS.

Target: neonforestmist/parameter-free-dynamic-regret-repro. Exact tags: icml2026-repro, paper-xnj5CyzVVW. Resolved artifact target: https://huggingface.co/buckets/neonforestmist/parameter-free-dynamic-regret-repro-artifacts.

Remote publication remains fail-closed until the coordinator assigns a rolling Space slot and reconfirms live target/tag/verdict absence.

Total size
1.59 MB
Files
53
Last updated
Jul 18
Pre-warmed CDN
US EU US EU

Contributors