Phase 4i: strong IRED graph reasoner still ties softmax — the flip is task-specific, not accuracy-gated - #5
Merged
Conversation
…k-specific, not accuracy-gated Tests the "the graph tie is just a weak reasoner" hypothesis by scaling the IRED graph model (hidden 256 / latent 48) to ID ~0.83 — as strong as arithmetic, low variance. - configs/experiments/graph_ired.toml: stronger model (h256/l48, e110, n_train 12000), n_eval 700 (a distinct (K, n_test) group from the Phase-4h weak run). RESULT (5 seeds, strong IRED graph, ID 0.831±0.012): geometry still beats scalar energy 5/5 (dAURC +0.092±0.029) but only TIES the best softmax baseline (dAURC +0.001±0.003, 0/5 seeds clear 0). So the graph tie is NOT a weak-reasoner artifact — the hypothesis is refuted. The arithmetic flip is genuinely task-specific: on graph a well-calibrated softmax already captures what descent geometry would add, even under a strong learned landscape. Paper (abstract/experiments/conclusion) + docs updated to state this boundary plainly; the open question becomes *when* geometry adds over softmax. 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Refutes the natural explanation for the graph tie (a weak IRED reasoner) by scaling it up — an honest negative result.
Test
The Phase-4h IRED graph reasoner capped at ID ~0.58 and only tied softmax. Hypothesis: with a stronger reasoner it would flip (like arithmetic). We scaled the model (hidden 256 / latent 48, more training) to ID ~0.83 — as strong as arithmetic, low variance.
Result (5 seeds, strong IRED graph)
Geometry still beats scalar energy 5/5 but still only ties softmax — so the graph tie is not a weak-reasoner artifact. The arithmetic flip (4/5) is genuinely task-specific, not a function of accuracy nor a general property of learned landscapes: on graph a well-calibrated softmax already captures what descent geometry would add.
Paper (abstract/experiments/conclusion) + docs updated to state this boundary plainly; the open question becomes when descent geometry adds over softmax. 92 tests green, ruff clean; paper rebuilds. Default reasoner unchanged; T1 stays canonical basin-center.
🤖 Generated with Claude Code