Skip to content

FIX preserve active GCG annealing loss - #2488

Open
Roman Lutz (romanlutz) wants to merge 3 commits into
microsoft:mainfrom
romanlutz:romanlutz-daily-audit-2026-08-25
Open

FIX preserve active GCG annealing loss#2488
Roman Lutz (romanlutz) wants to merge 3 commits into
microsoft:mainfrom
romanlutz:romanlutz-daily-audit-2026-08-25

Conversation

@romanlutz

Copy link
Copy Markdown
Contributor

Description

GCG simulated annealing compared a new candidate with the previously evaluated loss, even when that candidate had been rejected. This could accept a suffix that is worse than the active suffix simply because it improves on an earlier rejected candidate.

Compare candidates with OptimizationRunState.loss, which always describes the active control, and retain candidate_loss only as the latest evaluated loss. A deterministic regression test covers consecutive rejected candidates. This also restores hidden bibliography parity by adding the existing SATA citation key to doc/bibliography.md.

Tests and Documentation

  • Full unit suite: 16,241 passed, 84 skipped
  • Dataset and Many-shot tests: 4,693 passed
  • Full GCG suite: 228 passed
  • Focused GCG state tests: 23 passed
  • Metadata assertions: 27 passed
  • YAML parsing and BibTeX/hidden citation parity checks passed
  • Ruff, type checking, changed-file hooks, and diff checks passed
  • JupyText: not run; no notebook content changed

Roman Lutz (romanlutz) and others added 3 commits August 25, 2026 01:43
Restore hidden bibliography parity for the SATA citation and ensure annealing compares each candidate with the active suffix after a rejection.

Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant