Skip to content

Nine sources read in full, and what reading them changed

# research(2026-09). The Stage 1 report carried ten \needcite badges because the intake staged every entry as a search-level characterization. All but one are now cited: each source was read against arXiv, the ACM Digital Library, the ACL Anthology or Springer, and four of the claims built on them were wrong in a way a reviewer would have caught.

  • The conditional branch moves from rank 64 to rank 256. Biderman et al. (TMLR 2024) tested ranks 16, 64 and 256 only, on Llama-2-7B with alpha = 2r, and found code instruction tuning ordered by rank from the first epoch: HumanEval 0.358, 0.417, 0.498, against 0.497 for full fine-tuning. Their recommendation is a rank of 256 "since ranks 16-64 tend not to suffice for code tasks". The branch existed to separate a null from insufficient capacity, and at rank 64 it would not have. In continued pretraining no rank closes the gap, which is a second reason this study's instruction-tuning setting is the one the paper speaks to.
  • LORA stays at rank 32, and its comment no longer claims performance flattens there. The gate reads a difference between two adapters at the same rank, and the positive controls show this rank adapts. What changes is the justification: rank 32 is a compromise the branch covers, not a point the evidence identifies as sufficient.
  • Code2LoRA shows repository conventions by example, not by measurement. Its quantitative results are exact match on assertion completion; conventions appear in an appendix of qualitative cases. The background section now says premise where it said finding.
  • Project acceptance history is secondary. In the 66,329-interaction study the project's own acceptance ratio contributes 0.012 accuracy in the ablation, against 0.065 for the developer's history and 0.074 for the IDE version. The claim was true and the emphasis was not.
  • The 92.6% authorship figure is a ten-author closed set, and at 636 authors the same setting gives 27.9%. On coursework, 0.2% over 690 authors sits just above its own 0.14% chance baseline; only the 812-author open-assignment set is strictly below chance.
  • Gold and Krinke's finding is scoped. Their Gerrit dataset needed no ethics application under their local rules, on accessible data with no profiling, and they note review data carries higher risk than version control data because reviewers express opinions about work and its authors. The ethics section now carries both halves.
  • The COLING contamination paper does not cover Min-K%++. It benchmarks five detectors, including its own Local Order Quiz, and its finding this study relies on is that contamination introduced by instruction fine-tuning escaped every method but one. Min-K%++ belongs to this study's battery.
  • CodeReviewer is a model name. The ESEC/FSE 2022 paper is "Automating Code Review Activities by Large-Scale Pre-training"; the other title is the arXiv v1. The task shape and the exact-match metric are confirmed, and the paper reports BLEU beside EM.
  • The LLM-as-judge exclusion rests on the right paper. arXiv:2604.24525 is about judging review comments, not about exact match on generated code, which is exactly what that paragraph claims: 0.44 to 0.62 agreement on 2,604 industrial comments across three frontier models, EASE 2026.

The one badge left is the window assignment rule in section 5, which is a decision rather than a citation.