Skip to content

Latest commit

 

History

History
68 lines (57 loc) · 3.87 KB

File metadata and controls

68 lines (57 loc) · 3.87 KB

ICLR 2027 submission plan

Updated 2026-08-05. Official deadlines: abstract registration 2026-09-18 AoE; full paper 2026-09-25 AoE. The official template imposes a strict 9-page initial-submission main-text limit; references and appendices follow.

Current decision

Submit to ICLR 2027 after the confirmatory package is complete. The paper is now framed around placebo-calibrated predictive compression, not immediate two-part MDL consolidation. On current evidence it is a credible borderline accept rather than the earlier weak reject: the held-out split, direct ConsistencyGate-style baseline, hierarchical intervals, contamination control, and generated-candidate loop resolve the largest evaluation objections.

Non-negotiable claims

  • Primary evidence is the frozen P12 G1--G3 test: 90 pairs over five unseen generation seeds.
  • z is a calibrated predictive-compression signal, not a two-part MDL score.
  • Full textual descriptions do not repay their code cost in six observations.
  • The 1.5B development threshold transfers across unseen items and prompt phrasings, but not unchanged across model scales.
  • G4 is a failed scope diagnostic and remains reported.
  • P13 is a templated interface stress test, not a natural-conversation result.
  • The first P13 decoding seed is frozen; later seeds are explicitly post-hoc.

Completion gates

  • Frozen development/test protocol committed before confirmatory runs.
  • Qwen2.5-1.5B and 3B confirmatory runs.
  • Direct support, entropy, surprise, simplicity, and raw-gain baselines.
  • Hierarchical bootstrap, paired sign tests, and all-negative FPR.
  • Pseudoword zero-shot contamination controls on Qwen models.
  • Natural-utterance extraction and generated-candidate admission loop.
  • Two post-hoc decoding-seed robustness checks.
  • SmolLM2-1.7B held-out-family run and contamination control.
  • Remove every pending placeholder after the SmolLM run.
  • Regenerate figures and result tables from raw JSON.
  • Rebuild and visually inspect every final PDF page.
  • Independent human audit of citations, code, and LLM disclosure.
  • Confirm every author has a current OpenReview profile before abstract registration; freeze title, abstract, author list, and order.

Main reviewer risks and responses

risk evidence now in the paper still needed
synthetic benchmark construction isolates confident misreading; five unseen seeds; generated-candidate loop one real or naturally sourced user-convention benchmark would materially strengthen the paper
only one model family two Qwen scales plus frozen SmolLM2 family report SmolLM regardless of outcome
baseline fairness same candidate pairs; direct support probability; entropy, surprise, simplicity avoid claiming an exact reimplementation when a released baseline detail differs
MDL overclaim explicit code-length experiment and zero text-code payback keep title/abstract free of immediate MDL language
threshold overfitting dev-only threshold; unseen FPR; no test ranks require per-model calibration in deployment claims
weak end-to-end induction write-all is harmful; gate yields positive admitted gain preserve low coverage and malformed examples; do not cherry-pick generations

Schedule

  • By 2026-08-08: finish cross-family run, regenerate all results, and close numerical placeholders.
  • By 2026-08-15: complete a human citation/method audit and decide whether a naturally sourced convention dataset can be added without changing the confirmatory claim.
  • By 2026-08-29: external mock reviews, ablation requests, and one response cycle; freeze the experiment suite afterward.
  • By 2026-09-12: final anonymous PDF, repository snapshot, and OpenReview metadata rehearsal.
  • 2026-09-18 AoE: submit genuine abstract and final author metadata.
  • 2026-09-25 AoE: submit paper and supplement at least 24 hours early.