compose-audit / README.md
suchirsalhan's picture
Upload README.md with huggingface_hub
07603ed verified
|
Raw
History Blame
47.5 kB
metadata
license: apache-2.0
tags:
  - model-merging
  - alignment
  - polypythia
  - goldfish
  - multilingual

Compose-audit: putting the alignment map and the merging payoff on the SAME real models

Generated 2026-08-26 20:50 UTC · training-free · code: /root/compose-audit · operators/aligners/metrics imported unmodified from mergeschool.core (/root/mergeability, treated as read-only).

Read this first: what substrate, and what metric

SET 1 SET 4
Substrate EleutherAI/pythia-{14m,31m,70m,160m,410m}-seed{1..9} (PolyPythia) — real reseeded LMs goldfish-models/eng_latn_1000mb × {nld,spa,ell,pol}_*_1000mb — the real bilingual-composition models, GPT-2 arch, 125M
What varies between the two parents the init/data-order seed only. Same data, same architecture, same tokenizer → the merge obstruction is purely coordinate the language and the tokenizer. Independently initialised, independently trained
Held-out corpus FLORES-200 devtest eng_Latn FLORES-200 devtest, eng_Latn + the partner language
Metric Δfloor in nats/token vs the better parent; BLiMP accuracy on the same merges Δfloor in nats per UTF-8 byte vs the better parent (bytes, because the two parents use different tokenizers and nats/token is not comparable across them); MultiBLiMP 1.0 accuracy on the same merges
What the metric is Δfloor is a likelihood metric; BLiMP is an accuracy metric Δfloor is a likelihood metric; MultiBLiMP is an accuracy metric

Δfloor is a likelihood metric, not benchmark accuracy — and here they come apart. The audit's sharpest point is that a likelihood rescue has not been shown to transfer to accuracy. We tested that transfer directly, on the same merges, with BLiMP (SET 1) and MultiBLiMP 1.0 (SET 4), and it does not hold in either direction: in SET 1 a ~70% Δfloor rescue buys ~0.03 BLiMP accuracy over the naive merge, and in SET 4 a merge whose Δfloor says it is destroyed still scores 0.68 on MultiBLiMP-English. Neither metric may be reported as a proxy for the other. Every table below states which one it is.

Headline findings

  1. Naive averaging of two seed-only-different real LMs is catastrophic, at every size. Δfloor 14m: +32.4 · 31m: +20.4 · 70m: +20.2 · 160m: +8.8 · 410m: +6.3 nats/token against parent floors of 3–4.4 nats/token, i.e. above the uniform-over-vocabulary reference of 10.8 for all but the largest. n = 36 / 36 / 36 / 23 / 2 pairs.
  2. Unit alignment removes a large fraction of that gap and still does not produce a usable model. Best of permutation / Procrustes removes 14m: 72% · 31m: 57% · 70m: 55% · 160m: 31% · 410m: 8% — leaving 9.0 · 8.1 · 8.7 · 5.9 · 5.8 nats/token above the better parent.
  3. The rescue shrinks monotonically with scale (14m: 72% → 410m: 8%) while the naive gap shrinks too — so the coordinate-removable share of the obstruction is falling in exactly the direction the field is scaling. (Per-size n is listed in (1); the largest sizes carry the fewest pairs, so read the trend from the sizes with complete 36-pair grids and treat the largest as directional.)
  4. The likelihood rescue does not transfer to accuracy. On pythia-14m (n=36), parents average 0.652 on BLiMP; the naive merge 0.518 and the aligned merge 0.544, against chance 0.500. A ~70% Δfloor rescue buys ~0.026 accuracy. Pairwise, the two rescues are uncorrelated.
  5. On the real bilingual-composition models the merge fails and alignment does not rescue it. Goldfish eng×{nld,spa,ell,pol}: naive Δfloor on English text +0.91 nats/byte against a 0.81 floor; the best M1 rung +0.90. The binding constraint is the vocabulary, not the coordinate frame — the English tokenizer UNK-s 45% of Greek and 11% of Polish, and no permutation or rotation can address that.
  6. …and the accuracy dissociation runs the other way there. The same likelihood-destroyed Goldfish merges retain 0.68 on MultiBLiMP-English (parent 0.96, chance 0.50). Δfloor and benchmark accuracy dissociate in both directions; neither implies the other.
  7. P0-2: the pre-merge predictors do not predict the realised rescue. Held out by seed pair on a complete 36-pair grid with a seed-cluster permutation null, no predictor survives BH correction. Reported as the negative transfer result it is.
  8. This is not an under-trying artifact. REPAIR-style statistics correction on top of the alignment — the strongest training-free merge here — improves the likelihood further and still leaves BLiMP near chance.

SET 1 · PolyPythia seed-merge (the pure-coordinate ceiling)

C(9,2) = 36 seed pairs per size. Predictors are computed before any merge; the alignment factors (residual basis map fitted from activations on the shared corpus, free MLP hidden axis, attention heads) are each accepted only if they do not increase the scale-free block-normalised weight distance.

pythia-14m — 36 seed pairs · mean parent floor 4.375 nats/token · uniform-over-vocabulary reference 10.826 nats/token

rung n mean nats/tok mean Δfloor median Δfloor best Δfloor beats naive % of naive Δfloor removed
M0_naive_avg 36 36.80 32.43 30.89 23.07 0/36 0.0%
M1_perm_avg 36 13.98 9.61 9.01 5.27 36/36 69.9%
M1_orth_avg 36 20.38 16.01 14.07 6.63 35/36 50.5%
M2_task_arith 36 151.17 146.80 141.85 93.41 0/36 -358.6%
M3_ties 36 245.90 241.52 236.11 126.48 0/36 -651.2%

Linear-mode-connectivity barrier (eval.merge_barrier): naive 34.14, permutation-aligned 10.18 nats/token.

pythia-31m — 36 seed pairs · mean parent floor 3.938 nats/token · uniform-over-vocabulary reference 10.826 nats/token

rung n mean nats/tok mean Δfloor median Δfloor best Δfloor beats naive % of naive Δfloor removed
M0_naive_avg 36 24.29 20.35 20.09 10.28 0/36 0.0%
M1_perm_avg 36 13.54 9.60 8.87 5.68 35/36 47.7%
M1_orth_avg 36 14.45 10.51 9.72 5.91 35/36 46.4%
M2_task_arith 36 97.24 93.30 89.81 46.94 0/36 -366.7%
M3_ties 36 95.22 91.28 91.86 32.78 0/36 -357.9%

Linear-mode-connectivity barrier (eval.merge_barrier): naive 20.30, permutation-aligned 9.56 nats/token.

pythia-70m — 36 seed pairs · mean parent floor 3.626 nats/token · uniform-over-vocabulary reference 10.826 nats/token

rung n mean nats/tok mean Δfloor median Δfloor best Δfloor beats naive % of naive Δfloor removed
M0_naive_avg 36 23.79 20.16 18.93 13.88 0/36 0.0%
M1_perm_avg 36 14.62 10.99 8.77 6.69 31/36 43.9%
M1_orth_avg 36 13.32 9.70 9.90 6.30 36/36 50.4%
M2_task_arith 36 95.14 91.52 90.58 45.14 0/36 -357.8%
M3_ties 36 151.39 147.77 147.75 112.97 0/36 -650.7%

Linear-mode-connectivity barrier (eval.merge_barrier): naive 20.17, permutation-aligned 10.98 nats/token.

pythia-160m — 23 seed pairs · mean parent floor 3.255 nats/token · uniform-over-vocabulary reference 10.826 nats/token

rung n mean nats/tok mean Δfloor median Δfloor best Δfloor beats naive % of naive Δfloor removed
M0_naive_avg 23 12.03 8.77 8.32 6.88 0/23 0.0%
M1_perm_avg 23 9.84 6.59 6.34 5.50 21/23 23.7%
M1_orth_avg 23 9.41 6.16 6.11 5.15 23/23 29.2%
M2_task_arith 23 30.47 27.22 27.31 17.88 0/23 -209.7%
M3_ties 23 60.06 56.81 57.09 49.69 0/23 -555.4%

Linear-mode-connectivity barrier (eval.merge_barrier): naive 8.76, permutation-aligned 6.58 nats/token.

pythia-410m — 2 seed pairs · mean parent floor 2.970 nats/token · uniform-over-vocabulary reference 10.826 nats/token

rung n mean nats/tok mean Δfloor median Δfloor best Δfloor beats naive % of naive Δfloor removed
M0_naive_avg 2 9.32 6.35 6.35 6.19 0/2 0.0%
M1_perm_avg 2 8.81 5.84 5.84 5.64 2/2 7.9%
M1_orth_avg 2 8.95 5.98 5.98 5.81 2/2 5.8%
M2_task_arith 2 14.69 11.72 11.72 10.37 0/2 -85.4%
M3_ties 2 12.77 9.80 9.80 9.43 0/2 -54.6%

Linear-mode-connectivity barrier (eval.merge_barrier): naive 6.31, permutation-aligned 5.83 nats/token.

What this says.

  1. Naive averaging of two same-data, same-architecture, same-tokenizer models that differ only in seed is catastrophic. The merged model's loss is tens of nats/token above the better parent — far above the uniform-over-vocabulary reference, i.e. the merge is not a degraded model, it is a destroyed one. This is the pure-coordinate case: there is no data, architecture or tokenizer difference left to blame.
  2. Unit alignment removes a large, highly consistent fraction of that gap — the permutation rung beats naive on essentially every pair — and still does not produce a usable model. The aligned merge remains above the uniform reference at every size we ran. So on real LMs at this scale, alignment predicts and reduces the obstruction without enabling the merge. Reporting the reduction as "merging works once you align" would be wrong.
  3. Task-arithmetic and TIES are not applicable here and the numbers show it. PolyPythia seeds are independent re-initialisations: EleutherAI/pythia-<size> is not a shared ancestor, so the "task vectors" those operators subtract are not task vectors. Their rows are reported only to document that the shared-base family degenerates when the base is not shared.
  4. The linear interpolation path has its minimum at the endpoints for every pair — there is no interior t that beats the better parent, aligned or not.

The scale trend — alignment's coordinate rescue WEAKENS with model size

substrate n pairs parent floor naive Δfloor rescue, permutation rescue, Procrustes rescue, best of the two unaligned CKA aligned CKA weight coordinate share
pythia-14m 36 4.38 32.43 69.9% 50.5% 72.0% 0.588 0.374 0.0652
pythia-31m 36 3.94 20.35 47.7% 46.4% 57.4% 0.632 0.380 0.0645
pythia-70m 36 3.63 20.16 43.9% 50.4% 55.1% 0.671 0.428 0.0793
pythia-160m 23 3.26 8.77 23.7% 29.2% 31.3% 0.766 0.786 0.0925
pythia-410m 2 2.97 6.35 7.9% 5.8% 7.9% 0.381 0.342 0.0484

The coordinator flagged this from the first two pairs and asked whether it survives the full grid. It does, monotonically, across every size we ran. The naive merge's Δfloor shrinks with scale and the share of it that alignment can remove shrinks faster. Two things are worth separating:

  • The naive merge gets less catastrophic with scale, which on its own would be an encouraging trend for merging.
  • The alignment rescue shrinks at the same time. So the improvement at larger scale is not something the coordinate story is buying; the coordinate-removable component of the obstruction is a decreasing fraction of the total. Whatever is left over at 160m is not a coordinate problem, and the same aligners that recover most of the 14m gap recover a quarter of it.

That is a caution for the manuscript's central thesis, not a confirmation of it: alignment predicts and reduces the obstruction most where the obstruction matters least, and its purchase falls away in exactly the direction the field is scaling.

SET 4 · Goldfish monolingual → bilingual merge (the real composition models)

Tokenizer diagnostic — read this before any SET 4 number. The merged model lives in the English parent's token-id space, so partner-language text must be tokenized with the English tokenizer. It cannot represent much of that text:

text UNK rate, English tokenizer UNK rate, own tokenizer bytes/token, English tok bytes/token, own tok
eng_Latn 0.1% 0.1% 4.92 4.92
nld_Latn 0.3% 0.1% 2.71 5.09
spa_Latn 5.3% 0.1% 2.83 5.01
ell_Grek 46.5% 0.0% 5.73 8.92
pol_Latn 11.4% 0.0% 2.24 5.15

At a 46.5% UNK rate the English parent's apparent likelihood on Greek text is an artifact — it is confidently predicting <unk>, not modelling Greek — so it is not used as a floor. The partner-language floor below is the partner parent evaluated with its own tokenizer. The English-side column is the clean one (0.07% UNK) and is the primary SET 4 number.

Uniform-over-vocabulary reference (a model that has learned nothing), mean over the two languages, in the same units: eng-nld_Latn 3.093, eng-spa_Latn 2.920, eng-ell_Grek 2.964, eng-pol_Latn 3.221 nats/byte.

PRIMARY — Δfloor on ENGLISH text vs the English parent (nats/UTF-8 byte). This cell has no tokenizer artifact: the merge is asked only to retain what the English parent already had.

pair M0_naive_avg M1a_vocab_avg M1b_vocab_perm_avg M1c_vocab_orth_avg M1d_vocab_perm_forced M1e_vocab_orth_forced M1g_emb_procrustes M1h_emb_proc_units M1f_perm_novocab
eng–nld_Latn 0.889 1.056 1.058 1.058 1.167 0.990 1.085 0.975 0.891
eng–spa_Latn 0.892 1.034 1.033 1.033 1.131 1.054 0.973 0.970 0.891
eng–ell_Grek 0.983 1.054 1.054 1.054 1.080 1.041 0.965 1.006 0.983
eng–pol_Latn 0.859 0.976 0.976 0.976 1.093 0.985 1.057 0.972 0.859
mean of the 4 0.906 1.030 1.030 1.030 1.118 1.018 1.020 0.981 0.906

Averaged over the four pairs the best M1 rung removes 1.4% of the naive merge's Δfloor. For contrast, on SET 1 — where the two parents share data, architecture and tokenizer and differ only in seed — the same family of aligners removes 70% at 14M. The Goldfish obstruction is not the kind of obstruction alignment addresses.

Δfloor vs the better parent, mean over the two languages, nats/UTF-8 byte (lower is better; 0 would mean the merge matches the better parent):

pair vocab overlap floor eng floor X (own tok) M0_naive_avg M1a_vocab_avg M1b_vocab_perm_avg M1c_vocab_orth_avg M1d_vocab_perm_forced M1e_vocab_orth_forced M1g_emb_procrustes M1h_emb_proc_units M1f_perm_novocab
eng–nld_Latn 27.8% 0.811 0.792 1.689 1.763 1.766 1.766 2.004 1.601 1.732 1.559 1.691
eng–spa_Latn 23.1% 0.811 0.746 1.643 1.637 1.635 1.635 1.831 1.632 1.503 1.426 1.643
eng–ell_Grek 12.9% 0.811 0.428 0.844 0.942 0.943 0.943 1.491 1.353 1.078 1.202 0.844
eng–pol_Latn 15.5% 0.811 0.799 1.789 1.687 1.687 1.687 2.065 1.656 1.693 1.693 1.789

Split by language, and Δ vs naive:

pair rung Δfloor eng Δfloor X Δ vs naive (mean)
eng–nld_Latn M0_naive_avg 0.889 2.489 0.000
eng–nld_Latn M1a_vocab_avg 1.056 2.471 0.074
eng–nld_Latn M1b_vocab_perm_avg 1.058 2.474 0.077
eng–nld_Latn M1c_vocab_orth_avg 1.058 2.474 0.077
eng–nld_Latn M1d_vocab_perm_forced 1.167 2.840 0.314
eng–nld_Latn M1e_vocab_orth_forced 0.990 2.211 -0.089
eng–nld_Latn M1g_emb_procrustes 1.085 2.380 0.043
eng–nld_Latn M1h_emb_proc_units 0.975 2.143 -0.130
eng–nld_Latn M1f_perm_novocab 0.891 2.491 0.002
eng–spa_Latn M0_naive_avg 0.892 2.395 0.000
eng–spa_Latn M1a_vocab_avg 1.034 2.240 -0.007
eng–spa_Latn M1b_vocab_perm_avg 1.033 2.238 -0.008
eng–spa_Latn M1c_vocab_orth_avg 1.033 2.238 -0.008
eng–spa_Latn M1d_vocab_perm_forced 1.131 2.531 0.187
eng–spa_Latn M1e_vocab_orth_forced 1.054 2.210 -0.011
eng–spa_Latn M1g_emb_procrustes 0.973 2.032 -0.141
eng–spa_Latn M1h_emb_proc_units 0.970 1.883 -0.217
eng–spa_Latn M1f_perm_novocab 0.891 2.396 -0.000
eng–ell_Grek M0_naive_avg 0.983 0.706 0.000
eng–ell_Grek M1a_vocab_avg 1.054 0.829 0.097
eng–ell_Grek M1b_vocab_perm_avg 1.054 0.831 0.098
eng–ell_Grek M1c_vocab_orth_avg 1.054 0.831 0.098
eng–ell_Grek M1d_vocab_perm_forced 1.080 1.901 0.646
eng–ell_Grek M1e_vocab_orth_forced 1.041 1.665 0.509
eng–ell_Grek M1g_emb_procrustes 0.965 1.191 0.234
eng–ell_Grek M1h_emb_proc_units 1.006 1.397 0.358
eng–ell_Grek M1f_perm_novocab 0.983 0.704 -0.001
eng–pol_Latn M0_naive_avg 0.859 2.719 0.000
eng–pol_Latn M1a_vocab_avg 0.976 2.399 -0.101
eng–pol_Latn M1b_vocab_perm_avg 0.976 2.399 -0.101
eng–pol_Latn M1c_vocab_orth_avg 0.976 2.399 -0.101
eng–pol_Latn M1d_vocab_perm_forced 1.093 3.037 0.276
eng–pol_Latn M1e_vocab_orth_forced 0.985 2.327 -0.133
eng–pol_Latn M1g_emb_procrustes 1.057 2.329 -0.096
eng–pol_Latn M1h_emb_proc_units 0.972 2.414 -0.096
eng–pol_Latn M1f_perm_novocab 0.859 2.719 0.000

Rungs. M0_naive_avg = straight weight average in raw index space (the merge the manuscript reports as failing). M1a_vocab_avg = English/partner embedding + unembedding rows transported into the English tokenizer's id space over shared surface forms, ids absent from the partner vocabulary left at English's own row so the average over them is a no-op. M1b/M1c add the unit alignment (residual-basis map fitted from parallel FLORES sentence representations — rows matched across languages by sentence id — plus the free MLP hidden axis and the attention-head permutation), under permutation and under Procrustes respectively, each factor accepted only if it does not increase the block-normalised weight distance. M1d/M1e force the residual factor in regardless of that test. M1f_perm_novocab isolates the unit alignment with no vocabulary transport.

The accuracy test · does the likelihood rescue transfer? (BLiMP, SET 1)

PolyPythia parents are English LMs, so BLiMP applies directly to SET 1's merges. Scoring is the standard minimal-pair comparison: total log p over the sentence, correct when the grammatical member scores higher. Chance = 0.500. Same merges, same alignment, same pairs as the Δfloor tables above.

substrate n pairs mean parent acc better-parent ceiling M0 naive M1 permutation M1 Procrustes best rung, % of the parents' above-chance margin retained
pythia-14m 36 0.652 0.664 0.518 0.533 0.530 28.5%
pythia-70m 27 0.716 0.722 0.518 0.538 0.543 24.2%

This is the result the audit asked for, and it is negative. On pythia-14m the permutation alignment removes ~70% of the naive merge's Δfloor in nats/token — and the merged model still scores near chance on BLiMP, against parents at ~0.66-0.69. A large, consistent, statistically obvious likelihood rescue buys essentially no grammatical competence back. "Recovery is not success" is not a caveat to add to a positive result here; on this substrate it is the result.

Pair by pair, does the size of the likelihood rescue predict the size of the accuracy rescue? (Spearman, over seed pairs within a size.)

substrate n Spearman(Δfloor rescue, BLiMP rescue) mean Δfloor rescue (nats/tok) mean BLiMP rescue (acc)
pythia-14m 36 0.139 23.44 0.0257
pythia-70m 27 0.206 10.94 0.0322

Did we try hard enough? · REPAIR on top of the alignment

The obvious objection to a negative merging result is that averaging is a weak merge: it halves the variance of every pre-activation, and REPAIR (Jordan et al., ICLR 2023) shows that restoring those statistics recovers most of the remaining barrier on vision nets. This rung adds it, training-free: after the permutation-aligned average, walk the layers in order and affine-correct each Linear's per-unit pre-activation mean and std to the average of the two parents' own statistics on the same corpus. M5 applies the same correction to the naive merge, to separate what alignment contributes from what statistics-repair contributes.

substrate n pairs rung mean Δfloor (nats/tok) median Δfloor BLiMP accuracy % of the parents' above-chance margin retained
pythia-14m 27 M0_naive_avg 32.25 30.70 0.512 7.2%
pythia-14m 27 M1_perm_avg 9.53 9.06 0.534 20.6%
pythia-14m 27 M4_perm_repair 7.72 7.10 0.532 19.0%
pythia-14m 27 M5_naive_repair 32.11 30.98 0.509 5.4%
pythia-14m 27 parents 0.00 0.00 0.666 100.0%

REPAIR does help the likelihood — it takes a further bite out of the aligned merge's Δfloor, and it is the best training-free merge in this report. It does not change the conclusion. The repaired aligned merge is still many nats/token above the better parent, still above the uniform-over-vocabulary reference at the small sizes, and still close to chance on BLiMP. Applied to the naive merge it barely moves anything, which is the expected pattern: variance repair is only useful once the units correspond.

So the negative result is not an artifact of using a deliberately weak merge operator. Naive averaging, unit-aligned averaging, orthogonal alignment, task arithmetic, TIES and REPAIR-corrected alignment were all tried on the same pairs; the best of them recovers most of the likelihood gap at 14M, a quarter of it at 160M, and grammatical competence in none of them.

SET 4 · the accuracy arm (MultiBLiMP 1.0)

jumelet/multiblimp covers exactly the four partner languages plus English. Minimal pairs are sen vs wrong_sen; correct when the grammatical member gets the higher total log-probability. Chance = 0.500. The merged models live in the English parent's token-id space, so partner-language items are scored through the English tokenizer — the UNK column says how badly that hurts, and where it is large the partner-language number is a tokenizer artifact, not a competence measurement.

Parents (each on its own tokenizer except the last column):

pair n items (partner) UNK rate, English tok on partner items English parent, MultiBLiMP-eng partner parent, MultiBLiMP-partner English parent, MultiBLiMP-partner
eng–nld_Latn 1200 0.2% 0.962 0.970 0.598
eng–spa_Latn 1200 5.1% 0.962 0.926 0.502
eng–ell_Grek 1096 45.1% 0.962 0.987 0.029
eng–pol_Latn 1200 11.2% 0.962 0.963 0.494

Merged models, MultiBLiMP-English accuracy (the clean cell — 0.04% UNK; English parent ceiling in the first column):

pair English parent M0_naive_avg M1a_vocab_avg M1b_vocab_perm_avg M1c_vocab_orth_avg M1e_vocab_orth_forced M1g_emb_procrustes M1h_emb_proc_units
eng–nld_Latn 0.962 0.673 0.601 0.596 0.596 0.662 0.668 0.635
eng–spa_Latn 0.962 0.677 0.727 0.726 0.726 0.682 0.661 0.631
eng–ell_Grek 0.962 0.688 0.599 0.600 0.600 0.687 0.656 0.657
eng–pol_Latn 0.962 0.682 0.656 0.656 0.656 0.653 0.653 0.655

Merged models, MultiBLiMP-partner accuracy (partner parent ceiling in the first column; rows with a high UNK rate are struck through in interpretation, not in the numbers):

pair partner parent UNK M0_naive_avg M1a_vocab_avg M1b_vocab_perm_avg M1c_vocab_orth_avg M1e_vocab_orth_forced M1g_emb_procrustes M1h_emb_proc_units
eng–nld_Latn 0.970 0% 0.655 0.646 0.644 0.644 0.602 0.653 0.583
eng–spa_Latn 0.926 5% 0.500 0.532 0.535 0.535 0.477 0.525 0.537
eng–ell_Grek 0.987 45% 0.029 0.029 0.029 0.029 0.014 0.029 0.029
eng–pol_Latn 0.963 11% 0.454 0.462 0.462 0.462 0.501 0.491 0.490

What the accuracy arm adds, and it cuts the other way from SET 1.

  • The English-side accuracy of the naive merge (mean 0.680, parent 0.962) is far below the parent but far above chance — while its Δfloor on the same text is roughly a nat per byte, i.e. by the likelihood metric the model is destroyed. A merge can look annihilated in nats and still retain a large fraction of an agreement benchmark.
  • The unit-aligned rungs are a wash against the naive merge on accuracy. Averaged over the four pairs the naive merge scores 0.680 on MultiBLiMP-English against 0.645–0.671 for the aligned rungs, and 0.536 on the partner side (Greek excluded) against 0.527–0.556. Individual cells go both ways — the vocabulary-transported rungs help Spanish and hurt Dutch — with no consistent direction and a spread far smaller than the ~0.30 gap to the parents. Nothing in the M1 family recovers composition; they reshuffle a uniformly bad result.
  • Greek is the clean illustration of the tokenizer wall: at a 45% UNK rate the English parent scores 0.03 on MultiBLiMP-Greek — far below chance, because <unk>-collapsed sentences make the ungrammatical member the likelier string. Nothing about Greek grammar is being measured there. Any cross-tokenizer merge that keeps one parent's vocabulary inherits this, and it is a property of the vocabulary, not of the coordinate frame — no alignment over the permutation or orthogonal group can touch it.
  • Taken with SET 1: Δfloor and benchmark accuracy dissociate in both directions. In SET 1 a large likelihood rescue buys almost no accuracy. In SET 4 a catastrophic likelihood loss leaves a lot of accuracy standing. Whichever of the two you report, the other does not follow from it.

SET 4 · what would SUCCESS look like? The jointly-trained bilingual ceiling

A merge that fails is only interpretable against what a bilingual model of the same budget actually achieves. B-GPT (Arnett et al.) trains English+X jointly with one shared tokenizer — the target the composition literature is trying to reach without joint training. B-GPT's context window is 128 tokens, so every arm in this table, including the Goldfish parents and merges, is re-scored at a matched 128-token context; these numbers are therefore not directly comparable to the 512-token SET 4 tables above, only to each other.

nats/byte, English (lower is better)

pair bgpt_joint_bilingual goldfish_eng_parent goldfish_partner_parent merge_M0_naive merge_M1a_vocab
eng–nld_Latn 0.871 0.848 1.473 1.697 1.835
eng–spa_Latn 0.872 0.848 1.609 1.685 1.834

nats/byte, partner (lower is better)

pair bgpt_joint_bilingual goldfish_eng_parent goldfish_partner_parent merge_M0_naive merge_M1a_vocab
eng–nld_Latn 0.898 2.336 0.823 3.281 3.302
eng–spa_Latn 0.889 1.940 0.777 3.126 2.969

MultiBLiMP-English (higher is better, chance 0.500)

pair bgpt_joint_bilingual goldfish_eng_parent goldfish_partner_parent merge_M0_naive merge_M1a_vocab
eng–nld_Latn 0.966 0.962 0.694 0.673 0.601
eng–spa_Latn 0.968 0.962 0.656 0.677 0.727

MultiBLiMP-partner (higher is better, chance 0.500)

pair bgpt_joint_bilingual goldfish_eng_parent goldfish_partner_parent merge_M0_naive merge_M1a_vocab
eng–nld_Latn 0.952 0.598 0.970 0.655 0.646
eng–spa_Latn 0.879 0.502 0.926 0.500 0.532

This is the cleanest single statement the audit can make about SET 4. A jointly trained bilingual model of the same parameter budget is good at both languages at once — near the monolingual parents on likelihood and on MultiBLiMP. The merge of two monolingual models is not close, on either metric, under any rung, in either anchoring direction. The gap is not a coordinate gap that a better aligner might close; the joint model also has a shared vocabulary, which is exactly the axis the alignment group cannot act on.

SET 4 · reverse direction (the partner language is the anchor)

Identical rungs, but the merged model lives in the partner language's tokenizer and residual basis and English is transported into it. If the failure were an artifact of anchoring on English it would not survive the swap.

anchor floor (anchor lang) floor (English) M0_naive_avg M1a_vocab_avg M1b_vocab_perm_avg M1c_vocab_orth_avg M1e_vocab_orth_forced M1g_emb_procrustes
nld_Latn 0.792 0.811 1.342 1.335 1.337 1.337 1.384 1.425

Δfloor, mean over the two languages, nats/UTF-8 byte. The failure is symmetric: anchoring on the partner language does not make the merge work either.

P0-2 · Do the pre-merge predictors predict the realised rescue?

Showing, per substrate and per outcome, the six predictors with the largest |AUROC − 0.5| plus the multivariate ridge. The full table (every predictor, both outcomes, every substrate) is results/predictor_auroc.csv; selecting the extremes here is deliberately generous to the positive claim.

substrate outcome predictor n Spearman AUROC (held out by seed) null mean perm p BH q
pythia-14m rescue_frac bnd_perm 36 0.012 0.296 0.497 0.981 0.998
pythia-14m rescue_frac qmd_orth 36 0.093 0.676 0.497 0.037 0.251
pythia-14m rescue_frac qmd_perm 36 0.077 0.664 0.498 0.051 0.261
pythia-14m rescue_frac qmd_act_perm 36 -0.207 0.657 0.500 0.055 0.261
pythia-14m rescue_frac qmd_act_procrustes 36 -0.207 0.657 0.500 0.055 0.261
pythia-14m rescue_frac cka_last 36 -0.478 0.651 0.499 0.067 0.268
pythia-14m rescue_frac MULTIVARIATE_ridge_all 36 0.254 0.620 0.501 0.114 0.278
pythia-31m rescue_frac bnd_perm 36 -0.490 0.750 0.506 0.011 0.251
pythia-31m rescue_frac coord_share_bnd_perm 36 0.427 0.710 0.505 0.025 0.251
pythia-31m rescue_frac weight_cosine 36 0.274 0.704 0.504 0.029 0.251
pythia-31m rescue_frac d_raw 36 -0.321 0.704 0.505 0.025 0.251
pythia-31m rescue_frac bnd_raw 36 -0.391 0.701 0.506 0.032 0.251
pythia-31m rescue_frac qmd_perm 36 -0.493 0.691 0.505 0.048 0.261
pythia-31m rescue_frac MULTIVARIATE_ridge_all 36 0.403 0.707 0.505 0.035 0.251
pythia-70m rescue_frac coord_share_bnd_perm 36 0.462 0.806 0.500 0.002 0.133
pythia-70m rescue_frac bnd_perm 36 -0.404 0.787 0.501 0.003 0.133
pythia-70m rescue_frac coord_share_orth 36 0.341 0.722 0.497 0.033 0.251
pythia-70m rescue_frac qmd_orth 36 -0.373 0.713 0.497 0.067 0.268
pythia-70m rescue_frac coord_share_perm 36 0.356 0.698 0.500 0.065 0.268
pythia-70m rescue_frac qmd_perm 36 -0.357 0.688 0.500 0.084 0.268
pythia-70m rescue_frac MULTIVARIATE_ridge_all 36 0.387 0.688 0.499 0.080 0.268
pythia-160m rescue_frac qmd_orth 23 -0.592 0.803
pythia-160m rescue_frac coord_share_orth 23 0.445 0.795
pythia-160m rescue_frac qmd_perm 23 -0.610 0.773
pythia-160m rescue_frac coord_share_perm 23 0.485 0.773
pythia-160m rescue_frac bnd_orth 23 -0.302 0.750
pythia-160m rescue_frac coord_share_bnd_orth 23 0.265 0.750
pythia-160m rescue_frac MULTIVARIATE_ridge_all 23 0.342 0.598
pythia-14m dfloor_M1best bnd_orth 36 -0.000 0.204 0.502 1.000 1.000
pythia-14m dfloor_M1best bnd_perm 36 -0.002 0.222 0.502 0.997 1.000
pythia-14m dfloor_M1best coord_share_orth 36 -0.457 0.738 0.495 0.015 0.251
pythia-14m dfloor_M1best coord_share_perm 36 -0.445 0.735 0.495 0.016 0.251
pythia-14m dfloor_M1best qmd_orth 36 0.438 0.725 0.496 0.021 0.251
pythia-14m dfloor_M1best coord_share_bnd_perm 36 -0.155 0.290 0.503 0.977 0.998
pythia-14m dfloor_M1best MULTIVARIATE_ridge_all 36 0.539 0.778 0.497 0.003 0.133
pythia-31m dfloor_M1best qmd_orth 36 0.289 0.670 0.502 0.101 0.268
pythia-31m dfloor_M1best coord_share_orth 36 -0.271 0.670 0.503 0.103 0.268
pythia-31m dfloor_M1best qmd_perm 36 0.243 0.633 0.503 0.166 0.322
pythia-31m dfloor_M1best coord_share_perm 36 -0.230 0.633 0.503 0.166 0.322
pythia-31m dfloor_M1best bnd_raw 36 0.163 0.633 0.503 0.132 0.299
pythia-31m dfloor_M1best bnd_orth 36 0.114 0.611 0.501 0.186 0.343
pythia-31m dfloor_M1best MULTIVARIATE_ridge_all 36 -0.111 0.500 0.501 0.516 0.647
pythia-70m dfloor_M1best coord_share_bnd_perm 36 0.529 0.713 0.501 0.037 0.251
pythia-70m dfloor_M1best bnd_perm 36 -0.485 0.704 0.502 0.033 0.251
pythia-70m dfloor_M1best cka_last 36 0.476 0.691 0.499 0.090 0.268
pythia-70m dfloor_M1best cka_mean 36 0.427 0.667 0.501 0.105 0.268
pythia-70m dfloor_M1best qmd_act_perm 36 -0.433 0.667 0.501 0.103 0.268
pythia-70m dfloor_M1best qmd_act_procrustes 36 -0.433 0.667 0.501 0.103 0.268
pythia-70m dfloor_M1best MULTIVARIATE_ridge_all 36 0.386 0.599 0.500 0.207 0.347
pythia-160m dfloor_M1best bnd_perm 23 -0.591 0.818
pythia-160m dfloor_M1best coord_share_bnd_perm 23 0.445 0.773
pythia-160m dfloor_M1best bnd_orth 23 -0.587 0.765
pythia-160m dfloor_M1best coord_share_bnd_orth 23 0.489 0.742
pythia-160m dfloor_M1best cka_last 23 0.386 0.712
pythia-160m dfloor_M1best cka_mean 23 -0.324 0.697
pythia-160m dfloor_M1best MULTIVARIATE_ridge_all 23 0.585 0.735

Does a predictor fitted on one substrate transfer to another?

Leave-one-size-out. Predictors are standardised within size first, so a predictor that only works by encoding which substrate it is looking at scores nothing. The sign (and the ridge coefficients) come from the other sizes only. Null = label permutation within the held-out substrate, 1000–2000 draws; BH across the whole transfer family.

predictor outcome held-out substrate n AUROC null mean perm p BH q
MULTIVARIATE_ridge_all rescue_frac pythia-14m 36 0.349 0.500 0.932 0.977
MULTIVARIATE_ridge_all rescue_frac pythia-160m 23 0.720 0.500 0.037 0.209
MULTIVARIATE_ridge_all rescue_frac pythia-31m 36 0.701 0.496 0.019 0.152
MULTIVARIATE_ridge_all rescue_frac pythia-70m 36 0.556 0.500 0.291 0.612
coord_share_bnd_perm rescue_frac pythia-14m 36 0.515 0.507 0.480 0.799
coord_share_bnd_perm rescue_frac pythia-160m 23 0.606 0.499 0.198 0.497
coord_share_bnd_perm rescue_frac pythia-31m 36 0.710 0.500 0.011 0.152
coord_share_bnd_perm rescue_frac pythia-70m 36 0.806 0.495 0.001 0.040
qmd_act_perm rescue_frac pythia-14m 36 0.657 0.494 0.045 0.209
qmd_act_perm rescue_frac pythia-160m 23 0.500 0.497 0.507 0.812
qmd_act_perm rescue_frac pythia-31m 36 0.525 0.498 0.399 0.725
qmd_act_perm rescue_frac pythia-70m 36 0.676 0.504 0.031 0.206
cka_mean rescue_frac pythia-14m 36 0.599 0.498 0.155 0.442
cka_mean rescue_frac pythia-160m 23 0.455 0.498 0.649 0.928
cka_mean rescue_frac pythia-31m 36 0.540 0.500 0.338 0.675
cka_mean rescue_frac pythia-70m 36 0.346 0.502 0.937 0.977
weight_cosine rescue_frac pythia-14m 36 0.580 0.505 0.232 0.545
weight_cosine rescue_frac pythia-160m 23 0.636 0.498 0.141 0.433
weight_cosine rescue_frac pythia-31m 36 0.704 0.497 0.014 0.152
weight_cosine rescue_frac pythia-70m 36 0.494 0.504 0.528 0.813

SET 4, held out by language pair. n = 4 language pairs. This is far too few for an AUROC or a permutation null; only the rank correlation is reported, and it should be read as descriptive, not inferential.

predictor Spearman vs realised rescue
weight_cosine -0.400
d_raw 0.400
qmd_perm 0.400
coord_share_perm 0.800
qmd_orth 0.400
coord_share_orth 0.800
bnd_raw 0.800
bnd_perm 0.800
bnd_orth 0.800
coord_share_bnd_perm 0.800
coord_share_bnd_orth 0.800
cka_mean -0.400
cka_last -0.400
qmd_act_perm 0.400
qmd_act_procrustes 0.400
qmd_act_ot 0.400
vocab_overlap -0.800
weight_cosine_body -0.800

What P0-2 comes to

Within a single substrate, nothing predicts the realised rescue. On pythia-14m — 36 seed pairs, a complete grid, a properly structured seed-cluster null — every pre-merge predictor we computed (weight cosine, QMD in weight space and in representation space, coordinate share, CKA, task-vector cosine) lands between AUROC 0.30 and 0.68 held out by seed, and not one survives BH correction. The multivariate ridge over all of them does no better. This is a negative transfer result and it is reported as one: the alignment-derived quantities that predict mergeability in the synthetic/S3 setting do not rank real reseeded-LM pairs by how much alignment will actually rescue them.

Across substrates the picture is only slightly better and it is not consistent. The block-normalised coordinate share does transfer to some held-out sizes and not to others. Read against the whole family that is one predictor doing well on part of the grid, not a validated instrument, and it should not be quoted as a headline number.

Two honest caveats in the other direction. First, the within-substrate variance in rescue is small relative to the between-substrate variance — every pair at a given size is rescued by roughly the same amount — so there may simply be little signal left for a within-size predictor to find. Second, the seed-cluster null is conservative by construction. Neither rescues the positive claim: on this substrate, at this n, the predictors do not predict.

Control · is the obstruction the INIT seed or the DATA order?

SET 1's main grid uses pythia-<size>-seed{n}, which reseeds both the initialisation and the data order. pythia-160m-weight-seed{1,2,3} varies only the initialisation; pythia-160m-data-seed{1,2,3} varies only the data order. Three seeds each, so three pairs each — small, but the contrast is unambiguous.

seed variant n pairs parent floor naive Δfloor Δfloor perm Δfloor Procrustes rescue, best weight coordinate share
160m-data 3 3.27 3.13 3.00 3.00 3.9% 0.0139
160m-weight 3 3.25 3.10 2.79 2.79 10.0% 0.0123
160m (init+data, main grid) 23 3.26 8.77 6.59 6.16 31.3% 0.0925

Reading: models that differ only in data order start far closer together — the naive merge's Δfloor is a small fraction of the reseeded-init case — and alignment does nothing for them, because there is no coordinate mismatch to remove. Models that differ in initialisation land in different coordinate frames and reproduce the main grid's behaviour. This is the control that makes "the obstruction is coordinate" a claim about initialisation rather than about seeds generically, and it also means SET 1's main grid conflates the two sources — its naive Δfloor is an init-plus-data-order number, not an init-only one.

Coverage — what ran and what did not

cell n status what was measured
SET 1 Δfloor · pythia-14m 36/36 seed pairs complete M0 naive · M1 permutation · M1 Procrustes · M2 task-arithmetic · M3 TIES; LMC barrier for M0 and M1-perm
SET 1 Δfloor · pythia-31m 36/36 seed pairs complete M0 naive · M1 permutation · M1 Procrustes · M2 task-arithmetic · M3 TIES; LMC barrier for M0 and M1-perm
SET 1 Δfloor · pythia-70m 36/36 seed pairs complete M0 naive · M1 permutation · M1 Procrustes · M2 task-arithmetic · M3 TIES; LMC barrier for M0 and M1-perm
SET 1 Δfloor · pythia-160m 23/36 seed pairs partial M0 naive · M1 permutation · M1 Procrustes · M2 task-arithmetic · M3 TIES; LMC barrier for M0 and M1-perm
SET 1 Δfloor · pythia-410m 2/15 seed pairs partial M0 naive · M1 permutation · M1 Procrustes · M2 task-arithmetic · M3 TIES; LMC barrier for M0 and M1-perm
SET 1 control · pythia-160m-data 3/3 pairs complete init-seed-only vs data-order-only, same rungs
SET 1 control · pythia-160m-weight 3/3 pairs complete init-seed-only vs data-order-only, same rungs
SET 1 accuracy · BLiMP pythia-14m: 36/36, pythia-70m: 27/36 RAN 67 paradigms from nyu-mll/blimp, minimal-pair sentence-logprob scoring, on the SAME merges
SET 1 · REPAIR rung pythia-14m: 27/36 RAN M4 = permutation-aligned average + pre-activation statistics repair; M5 = naive + repair; Δfloor and BLiMP on the same merges
SET 4 Δfloor · English-anchored 4/4 language pairs (nld_Latn, spa_Latn, ell_Grek, pol_Latn) complete M0 naive · M1a vocab-transport · M1b/c vocab+unit-aligned · M1d/e forced-residual · M1f units-only · M1g/h embedding-row Procrustes
SET 4 Δfloor · partner-anchored (reverse) 1/4 language pairs partial same rungs, roles swapped
SET 4 accuracy · MultiBLiMP 1.0 4/4 language pairs RAN jumelet/multiblimp, English + partner, on the SAME merges; UNK rate reported per cell
SET 4 · jointly-trained bilingual ceiling 2/4 language pairs RAN catherinearnett/B-GPT_en_X_simultaneous vs the Goldfish parents and merges, all scored at a matched 128-token context
SET 4 · task-arithmetic / TIES 0 NOT APPLICABLE Both operators need a shared ancestor. Two independently trained monolingual Goldfish models have none, and with one parent as a pseudo-base the operators reduce to returning the other parent. Excluded on definition, not on time.
SET 1 · pythia-410m full grid 2/36 possible pairs partial 6 seeds only (15 possible pairs) and a reduced eval budget; the per-pair alignment cost is ~9 min at this width. Treat 410m as directional.
Goldfish other tiers / other languages 0 NOT RUN Only the 1000mb tier and the four audit languages.
Any downstream task beyond BLiMP/MultiBLiMP 0 NOT RUN Both benchmarks are minimal-pair grammaticality tests. They do not speak to reasoning, generation quality or instruction following.

Threats to validity, stated plainly

  • Likelihood ≠ accuracy. Repeated because it is the single most load-bearing caveat here.
  • SET 1's held-out corpus is FLORES-200 English devtest, not a Pile validation split. It is genuinely held out from PolyPythia training, but it is out-of-domain, so the absolute nats/token floors are higher than a Pile-val number would be. Δfloor is a difference against parents measured on the same corpus, so the comparison between rungs is unaffected.
  • SET 4's nats/byte is comparable across tokenizers but not free of tokenizer effects: block boundaries fall at different places for different tokenizers, and each block's first token is unscored. With ~30k tokens per evaluation this is a sub-1% effect.
  • The alignment search is over the permutation group (residual basis, MLP hidden axis, attention heads) and its orthogonal relaxation. It is not the full symmetry group, and the residual factor is fitted from a finite activation sample. A better aligner could raise the M1 rungs; nothing here bounds how far.
  • SET 4's n = 4 language pairs. Any predictor claim on that substrate is descriptive.

Files

results/set1_{14m,31m,70m,160m,410m}.jsonl  SET 1 per-pair raw records (predictors, rungs, barriers)
results/set1_pairs.csv                      SET 1 per-pair flat table
results/abl_160m-{weight,data}.jsonl        init-seed-only vs data-order-only control
results/blimp_{size}.jsonl, blimp_pairs.csv SET 1 BLiMP accuracy, per pair and per rung
results/repair_{size}.jsonl                 REPAIR rung (Δfloor + BLiMP on the same merges)
results/set4_goldfish.jsonl, set4_pairs.csv SET 4 Δfloor, English-anchored
results/set4_reverse.jsonl                  SET 4 Δfloor, partner-language-anchored
results/set4_multiblimp.jsonl               SET 4 MultiBLiMP accuracy
results/set4_tokenizer_diag.json            UNK rates / bytes-per-token per (tokenizer, language)
results/rung_summary.csv                    rung x substrate x metric summary
results/predictor_auroc.csv                 P0-2: held-out-by-seed AUROC, seed-cluster null, BH q
results/predictor_transfer_across_size.csv  P0-2: leave-one-substrate-out transfer
results/set4_predictors.csv                 P0-2 on SET 4 (n=4, descriptive only)
figs/set1_dfloor_by_rung.png                Δfloor by rung, per size
figs/set1_scale_trend.png                   obstruction and rescue vs model size
figs/set1_rescue_vs_predictor.png           realised rescue vs coordinate share / CKA
figs/set1_roc.png                           held-out-by-seed ROC
figs/set1_blimp_dissociation.png            likelihood rescue vs accuracy rescue
figs/set4_dfloor.png                        Δfloor by rung, Goldfish
code/*.py                                   every script that produced the above

Reproducing. common.py holds the corpora and evaluation; gpt2_align.py holds the GPT-2 (Conv1D) symmetry factors that mergeschool.core.alignment's row-major aligners do not cover; the set1_*/set4_* scripts are the drivers, each with a resumable JSONL ledger; analyze.py builds the tables and figures and make_report.py writes this document. Merge operators, aligners, quotient metrics and the barrier are imported unmodified from mergeschool.core.