Merge collapse is predictable from the weights alone: PRISM screens merges in CPU-minutes and repairs the destructive ones

A new preprint shows that model-merge collapse can be predicted from the specialists' weights alone, before any evaluation runs: the statistic is simply the variance of the task vectors across specialists — the power that averaging removes. The same statistic calibrates PRISM, a data-free repair operator that keeps all five flagged destructive merges within 0.5 points of the base model, where plain averaging falls at least 14.4 points below it or collapses entirely.

What shipped

Jungseob Lee and five co-authors posted "Predicting and Repairing Merge Collapse in Large Language Models" to arXiv on October 2, 2026 (23 pages, 5 figures, 20 tables), with MIT-licensed code released as the prism-merge package. The CLI's screen subcommand scores a merge from weights alone; merge applies the repair. The core depends only on numpy; the package installs as a standalone library (no mergekit integration yet).

The mechanism, from the paper and the code:

• At every weight coordinate, the total power of the scaled task vectors splits into signal (the squared averaged task vector) and interference. As the paper puts it, "the mean-squared power that averaging removes from the task vectors equals their variance across specialists."
• The pre-merge score is c·√(Ī/K): the merge coefficient times the square root of the parameter-weighted mean interference over merged layers, divided by the number of specialists. The shipped code hard-codes a frozen screening threshold of 1.9206e-3.
• On 22 merge configurations across four model families (Qwen2.5, Llama-3.1, Mistral, DeepSeek), with up to eight specialists in a single merge, seven merges are destructive at the default coefficient. The threshold flags five of the seven and none of the fifteen harmless merges.
• The team then froze the score and threshold and predicted the outcomes of fourteen further merges before evaluating them: twelve were correct, including the destructive outcome of a specialist pair pushed past the threshold by continued pretraining.
• Sign-conflict statistics are anti-predictive: incoherence and the cancellation ratio ρ get AUCs of 0.08 and 0.25 (1 = perfect) versus 0.98 for the interference score. "Statistics of sign disagreement, the conflict that sign election resolves, are anti-predictive."

PRISM itself is deliberately simple: average the norm-scaled task vectors first, then soft-threshold each layer at λ_l = σ_l·√(2·log n_l) — the universal threshold of wavelet shrinkage (Donoho & Johnstone) — with σ_l² set to the layer's mean interference. No data, no tuning, and the repair runs in linear time; the statistics take about two CPU-minutes on a 7B model.

The headline numbers, all from the paper's own runs:

• On the five flagged destructive merges, PRISM keeps the six-task mean within 0.5 points of the base in every case. The continued-pretraining pair (Japanese web text + biomedical abstracts, Qwen2.5-1.5B) goes 56.1 → 35.8 → 55.9; R1-Distill+Coder goes 68.1 → 49.7 → 67.7; Math-PRM+Coder goes 68.1 → 53.7 → 67.6. On the last two, the README notes naive averaging falls 18.4 and 14.4 points below the base.
• On Qwen2.5-7B Math+Coder, PRISM's six-task mean of 67.8 beats Task Arithmetic (49.5), TIES (61.1), DELLA (60.3), and LEWIS (56.1, which uses calibration data) by at least 6.7 points, and leads every baseline on all four generation metrics (GSM8K, MATH, HumanEval, MBPP).

Why it matters

Screening before you spend GPU hours. The score needs only the weights and minutes of CPU time, so a merge workflow can screen candidates before any evaluation. The prospective test — 12 of 14 predictions correct, made before evaluation with a frozen threshold — is the right shape of evidence for a pre-merge gate.

The contrarian result is the interesting one. TIES-style operators are built around sign conflict: they elect a sign per coordinate and drop entries that disagree. But the paper finds sign-conflict statistics are anti-predictive, because the destructive merges the score misses are dominated by one-sided drift (one specialist moves a weight the others leave near zero) rather than cancellation. The two destructive merges the score misses — Qwen2.5-7B Math+Finance and Coder+Coder-Instruct — have the lowest cancellation ratios among the destructive merges (ρ = 0.036 and 0.013), and PRISM's gate (ρ < 0.1) declines to repair them. If you are building a merge operator, the conflict you are resolving is not the failure mode you should be optimizing against.

Repair without data or tuning. One threshold scale serves all settings without data or tuning — a meaningful difference from calibration-data approaches like LEWIS — and the operator is cheap enough to run on every merge, not just the ones you care about.

Caveats

• Everything above is self-reported. The preprint went up on October 2, 2026, the repo has zero stars or forks, and we found no independent reproduction of the evaluation numbers. The per-task generation columns and the baseline runs are "outside this release," so the headline comparisons are not fully reproducible from the repo alone.
• The threshold is a hard-coded constant (1.9206e-3, "frozen" in the source). The paper's noise-model derivation is the part to scrutinize; the prospective 12/14 test mitigates, but does not eliminate, the risk that the threshold was tuned to the retrospective 22 merges.
• The score misses 2 of 7 destructive merges, and PRISM's gate deliberately leaves them as plain averaging (37.0 and 39.3 vs. a 68.1 base). "Predicting and repairing merge collapse" covers five of seven destructive cases.
• Interference is not the only separator. Total power and signal power separate the 22 merges perfectly (AUC 1.00) versus 0.98 for interference, whose raw score ranges overlap (min destructive / max harmless = 0.83×). In our reading, the paper does not fully explain why variance beats total power.
• Scope is narrow by design: only the MLP projections (gate/up/down) are merged per decoder layer; attention, embeddings, norms, and the output head stay at base values. Specialists trained for the paper, including the continued-pretraining pair central to the headline result, are not released.
• The six-task mean is an lm-evaluation-harness average over gsm8k, arc_challenge, mmlu, truthfulqa_mc2, hellaswag, and winogrande (500 examples per task, seed 42).

Sources

• arXiv abstract — Predicting and Repairing Merge Collapse in Large Language Models
• Full text (arXiv HTML)
• GitHub repository — js-lee-AI/PRISM
• README
• prism/core.py (score, threshold, soft-thresholding, gate)
• data/breadth.csv (22-merge table)
• data/titration.csv (continued-pretraining runs)
• experiments/table18_rival_statistics.py (rival-statistic AUCs)
• experiments/evaluate.py (six-task benchmark definition)
• pyproject.toml (package metadata)