Autoresearch#2
Open
thibautbar wants to merge 38 commits into
Open
Conversation
added 16 commits
May 27, 2026 09:17
…6 subset) Full train: sr = 0.9845, 100% validity (56102/56102) Bench commit: 3eb0f69 # iter #1 — finer-grained early termination in the Laporta level loop ## Mutation One edit in `experiments/fire7_src/functions.cpp`, in the per-sector Laporta driver (`reduce_in_sector`, the loop that was previously `for (unsigned int current_sum = 2; ...)`). Vanilla groups **all seeding levels of equal total degree** `(p+m)` into a single batch, submits them to the level-reduction threads to
…set) Full train: sr = 0.9756, 100% validity (56102/56102) Bench commit: 3bb1ba3 # iter #2 — pure-dot symmetry seeding at all dot levels ## Mutation One edit in `experiments/fire7_src/functions.cpp`, in the per-sector Laporta driver (`forward_stage`), in the symmetry-writing block that runs at the top of each level before the IBP pass. Vanilla (and the iter #1 lineage tip) writes internal symmetries **only at the `(pos <= pos_pref, neg == 1)` levels**. With the default `p
…bset) Full train: sr = 0.9740, 100% validity (56102/56102) Bench commit: 2c699d3 # iter #3 — extend pure-dot symmetry seeding to one-numerator levels ## Mutation One edit in `experiments/fire7_src/functions.cpp`, in the per-sector Laporta driver (`forward_stage`), inside the `pos_pref > 0` symmetry-seeding block. iter #2 changed the seeding condition to fire on every **pure-dot** level (`current_level.second == 1`). iter #3 widens it by one numerator level: ```cpp - if
…30 subset) Full train: sr = 0.9735, 100% validity (56102/56102) Bench commit: aedfb4b # iter #4 — extend pure-symmetry seeding to two-numerator levels ## Mutation One edit in `experiments/fire7_src/functions.cpp`, in the per-sector Laporta driver (`forward_stage`), at the symmetry-seeding guard inside the level loop: ```diff - if (current_level.second <= 2) { + if (current_level.second <= 3) { symmetrie
…(0.9830 subset) Full train: sr = 0.9735, 100% validity (56102/56102) Bench commit: 044477c # iter #5 — drop the numerator cap on symmetry seeding (the limit of the iter #2→#4 ramp) ## Mutation One edit in `experiments/fire7_src/functions.cpp`, in the per-sector Laporta driver (`forward_stage`), in the pure-symmetry pre-seeding block that runs before each level's IBP tasks are dispatched. ```diff - if (current_level.second <= 3) { -
Full train: sr = 0.9683, 100% validity (56102/56102) Bench commit: 3132daf # iter #6 — pre-IBP early-exit (probe `done` *before* the level's IBP pass) ## Result - Full train: `macro_validity = 1.0`, `mean_step_ratio = 0.96831` (prev best iter #5 = 0.97352). **Improvement ≈ 0.0052, 100% valid.** - Subset (mini_eval, 5D/bl2em + 9D/banana3L): sr 0.9821 vs iter #5's 0.9830. ## Mutation One edit in `experiments/fire7_src/functions.cpp`, in the per-sector Laporta dri
…bset) Full train: sr = 0.7977, 100% validity (56102/56102) Bench commit: e35feb3 # iter #8 — front-load symmetry seeding for *all* levels before the IBP loop ## Mutation One edit in `experiments/fire7_src/functions.cpp`, in the per-sector Laporta driver, right after the level-worker threads are spawned and before the level loop (`auto itr = levels.begin();`). I added a single pre-pass that calls `write_symmetries(...)` for **every** pending level in `levels` up front, usin
…0.8519 subset) Full train: sr = 0.7934, 100% validity (56102/56102) Bench commit: 4037714 # iter #10 — recover the \"sending-higher\" internal-symmetry relations ## Result - **Full train**: `macro_validity = 1.0`, `mean_step_ratio = 0.79341` (prev best iter #8 = 0.79767). **Improvement: −0.0043, validity preserved.** - Subset (mini_eval): `sr = 0.8519` (vs iter #8's 0.855), 100% valid. ## Mutation One edit in `experiments/fire7_src/functions.cpp`, in `write_symmetries` (the
… train (0.6995 subset) Full train: sr = 0.6128, 100% validity (56102/56102) Bench commit: 3fab1ca # iter #11 — front-load symmetry seeding for the upper-neighbour levels too ## Result - **Full train**: `macro_validity = 1.0`, `mean_step_ratio = 0.613` (prev best iter #10 = 0.79341). **Improvement: −0.180, validity preserved.** All 13 topologies 100% valid, full coverage (56102/56102). - Subset (mini_eval): `sr = 0.6995` (vs iter #10's 0.8519), 100% valid (5D/bl2em 0.522, 9D/banana3
…0.6836 subset) Full train: sr = 0.6081, 100% validity (56102/56102) Bench commit: fe46398 # iter #12 — seed the diagonal upper-neighbour level (p+1, n+1) ## Result - **Full train**: `macro_validity = 1.0`, `mean_step_ratio = 0.608` (prev best iter #11 = 0.6128). **Improvement: −0.005, validity preserved.** All 13 topologies 100% valid, full coverage (56102/56102). - Subset (mini_eval): `sr = 0.6836` (vs iter #11's 0.6995), 100% valid (5D/bl2em 0.491 ← 0.522, 9D/banana3L 0
Full train: sr = 0.4618, 100% validity (56102/56102) Bench commit: de0d07e # iter #24 — within-level early-exit, re-landed via a file-scope hook + the merge-blocker fix ## TL;DR Two things, one root cause for the stuck loop: 1. **Diagnosed and fixed the infra blocker** that discarded the two best results so far. iter #16 and iter #23 both reached `sr = 0.4618`, `validity = 1.0` on the full train split — a **−0.146** improvement over the current best (it
…1815) Full train: sr = 0.4618, 100% validity (56102/56102) Bench commit: bf65723 # iter #25 — within-GROUP early-exit (finer than iter #24's group-boundary probe) ## TL;DR iter #24 re-probed the \"all needed targets resolved\" condition only at **group boundaries** of the final level's IBP pass — after a whole group of equations sharing the same highest member had been worked. But a single group writes **more than one pivot**: the first equation reduces the shared high
Full train: sr = 0.4618, 100% validity (56102/56102) Bench commit: f900fe5 # iter #25 — within-GROUP early-exit (finer than iter #24's group-boundary probe) ## TL;DR iter #24 re-probed the \"all needed targets resolved\" condition only at **group boundaries** of the final level's IBP pass — after a whole group of equations sharing the same highest member had been worked. But a single group writes **more than one pivot**: the first equation reduces the shared high
… 0.46137 full train Full train: sr = 0.4614, 100% validity (56102/56102) Bench commit: d6b8039 # iter #33 — depth-2 symmetry composition in the seeding (revisit resolved points) ## TL;DR One edit in `experiments/fire7_src/functions.cpp`, inside `write_symmetries` (the internal-symmetry seeding routine). It lets the seeder apply a point's internal symmetries **even after that point already has a rule**, but *only* to write the iter #10 \"sending-higher\" rule for a strictly-higher, sti
Full train: sr = 0.4497, 100% validity (56102/56102) Bench commit: 9155e42 # iter #37 — cross-group closure segregation (defer every off-closure group) ## TL;DR One edit in `experiments/fire7_src/functions.cpp`, in `reduce_in_level`, just before the per-level equation-group loop. Before working the level's equation groups, **stable_partition** them into `[groups that pivot a needed-closure point]` ++ `[off-closure groups]`, each kept in its original ascending (bottom
Full train: sr = 0.4491, 100% validity (56102/56102) Bench commit: 123197b # iter #39 — multi-wave closure re-segregation (deepen iter #37) ## TL;DR One change in `experiments/fire7_src/functions.cpp`, in `reduce_in_level`, inside the per-level equation-group loop. iter #37 floats the needed-target **closure** groups ahead of the off-closure tail with a **single** `stable_partition` computed *before* the loop. The fix: re-segregate the still-unworked suffix at **ever
added 13 commits
May 27, 2026 09:48
…idity) Bench commit: ab58f3397fe95a307c8d0658a9e5638ad5c45264 Branch: claude/exp/exp-0044
…lidity) Bench commit: 6c02829f68a5669d8c562a43643cbbe10aa5f871 Branch: claude/exp/exp-0045
…lidity) Bench commit: 7398f6152b6640b172ded74cd998a16fe4a480c6 Branch: claude/exp/exp-0056
…dity) Bench commit: f3c84ca5711498ab615b9257707170ead71984b2 Branch: claude/exp/exp-0057
…lidity) Bench commit: 7d2c19a847a8c7d80d66a338bc1f19ee910ce339 Branch: claude/exp/exp-0059
…idity) Bench commit: 58d8c0f778ad62d6ff4a0b93140a020910703050 Branch: claude/exp/exp-0075
…idity) Bench commit: 055cfe72998287a41507be6a67383c7d71ee6857 Branch: claude/exp/exp-0092
…idity) Bench commit: 94a3c5f2884577cc3f10ef72be77f4ae416991b7 Branch: claude/exp/exp-0098
…alidity) Bench commit: e8159fb857f282bb46f8f94d67b8432b2f3911bd Branch: claude/exp/exp-0118
…lidity) Bench commit: 508b0fdfe3cdc7d841e86a45a91a797d0b4f05ef Branch: claude/exp/exp-0119
…alidity) Bench commit: 302315687495136485e2dc469f16acfed5023ac3 Branch: claude/exp/exp-0121
…idity) Bench commit: 55977e795839db945098092c308e20e22fc6ecf2 Branch: claude/exp/exp-0126
…lidity) Bench commit: a7da5449cdbbb26f6216ac50c90e50f56fbe7fe7 Branch: claude/exp/exp-0127
added 9 commits
May 28, 2026 08:06
…lidity) Bench commit: 97eeaa0b798f3030ae9aedaa6c526cb30c5e8849 Branch: claude/exp/exp-0130
…alidity) Bench commit: f260e658af6ea3cf0ca2e5c6346b3db4c2953971 Branch: claude/exp/exp-0155
…lidity) Bench commit: bca8b18171517f3bad49533287869dc7beb82f63 Branch: claude/exp/exp-0157
…alidity) Bench commit: fc069996623612a0f9a2d089b8a426882e2f3a04 Branch: claude/exp/exp-0166
…lidity) Bench commit: ed51c7ab1f1ab9430d67de136f61c0a13a2d8ca5 Branch: claude/exp/exp-0210
…lidity) Bench commit: c461bc200b104bbf496e3314fb15a400cea68914 Branch: claude/exp/exp-0245
…lidity) Bench commit: b991290eb4428e3d84a9200172e47162d4f40433 Branch: claude/exp/exp-0360
…alidity) Bench commit: 61fa1a87acf941ed961603ddd3f0de204f69e60d Branch: claude/exp/exp-0371
…lidity) Bench commit: 43ccbf8c0eb45530fb884ef924885580f9b4f478 Branch: claude/exp/exp-0424
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.
Uh oh!
There was an error while loading. Please reload this page.