Phase 4h: IRED on graph — the softmax flip is task-dependent (honest update) - #4
Merged
Conversation
…honest update) Tests whether the Phase-4g arithmetic flip (geometry beats softmax under a learned landscape) generalizes to a second task. It does not cleanly. - configs/experiments/graph_ired.toml + sweeps/graph_ired_seeds.toml: 5-seed IRED-on-graph. - eval/evaluate.py: record `objective` + `sampler` in metrics. - analysis/aggregate.py: objective-aware (objective_of, selective_rows(objective=...), headline_cell defaults to basin_center) so IRED rows never pollute the canonical basin-center T1 (fixes a fragile tie-break where IRED and basin sweeps shared K=12). - Paper (abstract/experiments/conclusion) + docs tempered to claim the flip narrowly. RESULT (5 seeds/task): arithmetic IRED beats softmax 4/5 (flip holds); graph IRED only TIES softmax (1 win / 1 loss / 3 ties, dAURC +0.004±0.014) — geometry still beats scalar energy 5/5 on both. The learned reasoner is weaker on graph (ID 0.55-0.82), so the flip is demonstrated on arithmetic, not yet graph: real but task-dependent. 92 tests green, ruff clean; paper rebuilds. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
Bugbot is not enabled for your account, so this pull request was not reviewed. Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Tests whether the Phase-4g arithmetic flip (geometry beats softmax under a learned landscape) generalizes to graph. It doesn't cleanly — an honesty update to the v0.0.2 claim.
Result (5 seeds/task, ΔAURC vs best-of-all baselines)
On graph the learned reasoner is weaker (ID 0.55–0.82), and geometry only ties softmax. So the flip is real but task-dependent — demonstrated on arithmetic, not graph. Geometry still beats scalar energy 5/5 on both.
Also
aggregate.objective_of,headline_cell(objective="basin_center")): fixes a latent bug where IRED and basin-center sweeps shared K=12 and T1 picked one by a fragile tie-break.evaluatenow recordsobjective/sampler.92 tests green, ruff clean; paper rebuilds via tectonic. Opt-in/additive; default reasoner unchanged.
🤖 Generated with Claude Code