PaperlayerView exact revision on arXiv

Source-linked AI summary

Predicting and Repairing Merge Collapse in Large Language Models

Jungseob Lee, Seungyoon Lee, Sugyeong Eo, Hyeonseok Moon, Jaehyung Seo, Heuiseok Lim

arXiv:2610.03199v1cs.LGcs.CL

TL;DR

Merging specialists can cause collapse, while existing operators do not warn of destructive outcomes before evaluation. The paper uses task-vector interference to predict risk and calibrate PRISM, which repairs destructive merges by thresholding the averaged task vectors. Across the reported configurations, the score flags destructive merges while PRISM keeps them near the base model.

  • Problem

    Some merges of specialists fine-tuned from a shared base collapse far below the base, and common operators do not warn of destructive outcomes before evaluation.

    Source: §1 · pp. 1–2

  • Method

    The paper measures interference as task-vector variance, uses it to score pre-merge disturbance, and sets PRISM’s soft-threshold for the averaged task vectors.

    Source: §1 · p. 2

  • Results

    The score flagged five of seven destructive merges and none of fifteen harmless merges, and twelve of fourteen further merge outcomes were predicted correctly.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Applying PRISM above the score threshold while retaining plain averaging below it kept destructive merges near the base and preserved plain averaging for harmless merges.

    Source: §1 · p. 2

    The screen and evaluations target capability benchmarks and do not measure safety behavior, so merged models need separate safety and alignment evaluation before deployment.

    Source: ETHICS STATEMENT · p. 9

Abstract

from arXiv · show

Large language models fine-tuned from a shared base can be merged by averaging their task vectors, but some merges collapse far below the base model, and common merge operators give no warning before evaluation. We show that one statistic of the specialists' task vectors both predicts this collapse and calibrates its repair. The power that averaging removes equals the variance of the task vectors across specialists, our measure of interference. Under a working noise model, the disturbance that a merge injects grows with the merge coefficient and with interference, yielding a pre-merge score. In our experiments on twenty-two merge configurations from four model families, only destructive merges exceed a threshold on this score. We find that statistics of sign conflict between specialists, a common target of existing merge operators, are anti-predictive. We then predicted the outcomes of fourteen merges before evaluating them, and twelve predictions were correct, including the destructive outcome of a specialist pair pushed past the threshold by continued pretraining. To address this collapse, we introduce PRISM, an operator that averages the task vectors first and then soft-thresholds each layer at a level set by the layer's interference. Without data or tuning, PRISM keeps all five destructive merges above the threshold within evaluation noise of the base model, where plain averaging falls at least 14.4 points below it or collapses entirely. We apply PRISM only above the threshold and keep the plain average for merges below it, which include all fifteen harmless ones. Code is available at https://github.com/js-lee-AI/PRISM.

1 INTRODUCTION

Averaging specialists can erase task-vector updates, and existing approaches do not establish whether a merge will be destructive before evaluation. The paper uses task-vector interference to screen merge risk and set PRISM’s repair, then tests the score and repair across merge configurations.

1 INTRODUCTION

  • Averaging specialists fine-tuned for different tasks cancels some of their differently directed weight changes, while existing merge approaches leave destructiveness unknown until evaluation.Destructive merges fall substantially below the base model; harmless ones stay near or above it.Source: §1 · pp. 1–2
  • Interference is the variance of task vectors across specialists, equal to the mean-squared power removed by averaging; under a working noise model, the pre-merge score grows with the merge coefficient and square root of interference.Interference includes cancellation and dispersion, including one-sided drift.Source: §1 · p. 2
  • PRISM averages task vectors before soft-thresholding each layer at a level set by interference, using one threshold scale without data or tuning.A cancellation-share gate returns the plain average when cancellation is small, preserving one-sided drift.Source: §1 · p. 2
  • The score flagged five of seven destructive merges and none of fifteen harmless merges; twelve of fourteen further outcomes were predicted correctly.Sign-disagreement statistics were anti-predictive.Source: §1 · p. 2
  • Continued pretraining pushed a Japanese-web and biomedical specialist pair past the threshold, where plain averaging became destructive as predicted and PRISM stayed near the base.At a fixed learning rate, the score grew roughly with the square root of training tokens.Source: §1 · p. 2
  • PRISM keeps destructive merges near the base, while plain averaging falls far below it; the paper applies PRISM above the score threshold and averages plainly below it.Below-threshold merges include all fifteen harmless ones.Source: §1 · p. 2

2 RELATED WORK

Prior model-merging work includes prune-then-merge operators, data-dependent and training-free approaches, and spectral or subspace methods. This paper adds pre-merge collapse screening and a repair that thresholds the averaged task vector using a weight-derived interference scale.

2 RELATED WORK

  • Prune-then-merge work includes magnitude trimming with sign election in TIES-Merging, random dropping and rescaling in DARE, and magnitude-dependent dropping in DELLA.These examples illustrate operators that prune task vectors before aggregation.Source: §2 · p. 3, Figure 1 · p. 3
  • Other approaches use auxiliary data, training-free merging, task-specific masks, spectral flattening, task singular vectors, or subspace alignment.The related work also describes input-dependent combination of shared and exclusive knowledge.Source: §2 · p. 3
  • The paper distinguishes its contribution as screening collapse risk before merging and soft-thresholding the averaged task vector in linear time, using weights alone rather than pruning each vector or computing singular-value decompositions.The threshold is grounded in an exact variance identity and a noise model.Source: §2 · p. 3

3 INTERFERENCE IN TASK-VECTOR MERGING

Interference is the variance of scaled task vectors across specialists, equal to the power averaging removes. Under a working noise model, it also sets a pre-merge estimate of the disturbance introduced by merging.

3.1 INTERFERENCE IS THE CROSS-MODEL VARIANCE

  • Task-vector interference is the population variance across specialists, equal to the reduction in mean-squared power caused by averaging.Identical scaled vectors have zero interference even if their shared drift is large; unequal magnitudes and one-sided drift can produce positive variance.Source: §3.1 · pp. 3–4

3.2 WHAT A MERGE INJECTS

  • Under an independent, shared-signal noise model, merge disturbance grows with the coefficient and interference, yielding a pre-merge screening scale.Using measured interference as a proxy for disturbance variance is a modeling choice; estimated disturbance power grows as c^2 at fixed interference.Source: §3.2 · p. 4

3.3 CANCELLATION AND DISPERSION

  • Interference comprises cancellation from opposing signs and dispersion, which includes potentially useful one-sided specialist drift.The cancellation ratio ρ distinguishes conflict-dominated from dispersion-dominated merges and ranges from 0.003 to 0.720 across four model families.Source: §3.3 · p. 4

4 REPAIRING A DANGEROUS MERGE WITH PRISM

PRISM averages task vectors before applying an interference-calibrated soft threshold, with a gate that leaves dispersion-dominated means unchanged. Its reported merge procedure uses norm scaling and applies PRISM to MLP projections while retaining base attention weights.

4 REPAIRING A DANGEROUS MERGE WITH PRISM

  • PRISM averages task vectors first, then soft-thresholds the aggregate using each layer’s interference scale.A coordinate survives only if its aggregate magnitude exceeds the threshold, and surviving magnitudes are reduced by that threshold.Source: §4 · p. 4, §4.1 · pp. 4–5

4.1 INTERFERENCE-CALIBRATED SOFT THRESHOLDING

  • PRISM thresholds under substantial cancellation but sets the threshold to zero for dispersion-dominated merges, leaving their mean unchanged.The default gate is τg = 0.1; the reported ratios imply any threshold strictly between 0.093 and 0.229 makes the same decisions on these settings.Source: §4.1 · p. 5

4.3 MERGE PROCEDURE AND SCOPE

  • The default procedure norm-scales task vectors, applies PRISM to MLP projections, and leaves attention weights at their base values.Extending PRISM to attention merges 2.3× more keys and stays within 0.1pp of the six-task mean obtained with MLP projections alone.Source: §4.3 · p. 5

5 EXPERIMENTS

The experiments evaluate model-family and specialist variants, then test PRISM against merging baselines, coefficient and threshold choices, and a pre-merge interference screen. PRISM preserves capability on destructive merges, while the screen distinguishes many destructive merges from harmless ones and supports prospective predictions.

5.1 SETUP

  • The evaluation covers Qwen2.5 and Llama-3.1 7–8B models, additional Mistral-7B and DeepSeek-7B breadth tests, and six benchmarks plus WikiText-2 perplexity.The main setting pairs Qwen2.5 Math and Coder specialists; the default coefficient is c = 1/K.Source: §5.1 · p. 5

5.2 MERGING UNDER SEVERE CONFLICT

  • PRISM leads Task Arithmetic, TIES, and DELLA by at least 6.7pp on the six-task mean, and leads every baseline on four generation metrics by at least 25.0pp on HumanEval.On Qwen2.5-1.5B, PRISM leads the strongest baseline by 3.5pp while Task Arithmetic collapses outright.Source: §5.2 · p. 6

5.3 ROBUSTNESS TO THE SCALING COEFFICIENT

  • Across the coefficient sweep, PRISM holds a 64.8–68.0% mean, while four baselines at c = 0.7 fall 19 to 33 points below the base.At c = 1.0, Task Arithmetic, TIES, and DELLA score zero on generation GSM8K; LEWIS scores 17.4.Source: §5.3 · p. 6

5.4 IS IT JUST CONSERVATIVE SHRINKAGE?

  • The norm-matched control avoids collapse and stays near the base, but PRISM concentrates its update on 0.1% of coordinates versus 85.1% for the control.PRISM scores 65.9 on HumanEval against 62.2 for the control and 62.8 for the base.Source: §5.4 · pp. 6–7

5.5 DATA-DEPENDENT BASELINES NEED DATA TO REACH PARITY

  • AdaMerging ties PRISM at 67.9 versus 67.8 after unlabeled-data entropy minimization and a coefficient sweep; data-free Iso-C reaches 66.6 when tuned away from its collapsing default.Greedy Soup returns the base model, Fisher-weighted averaging scores 47.5, and WUDI-Merging diverges in every tested configuration.Source: §5.5 · p. 7

5.6 ABLATIONS AND MECHANISM

  • At the same interference scale, soft thresholding beats hard truncation by 4.4pp on the six-task mean; truncation costs 19.2pp on GSM8K and up to 28.6pp on HumanEval.Across κ=0.25 to 3, the six-task mean varies by at most 0.5pp; magnitude pruning lowers it by 3.7pp at 7B and 6.9pp at 1.5B, with the gap closing at K=3.Source: Figure 3 · p. 7, §5.6 · p. 8

5.7 THE SAME SCALE SUPPORTS PRE-MERGE SCREENING

  • The score exceeds 1.9×10−3 for five of seven destructive merges and none of fifteen harmless merges; mean incoherence and ρ are anti-predictive.Interference, total power, and signal power each reach AUC 0.98 or higher.Source: §5.7 · p. 9
  • Twelve of fourteen pre-evaluation predictions were correct; for the constructed destructive endpoint, plain averaging fell 20.3pp below the base while PRISM stayed within 0.3pp.The recipe applies naive averaging below the score threshold and gated shrinkage above it; all fifteen harmless configurations remain below the threshold.Source: §5.7 · p. 9

6 CONCLUSION

The paper presents task-vector variance as both a predictor of merge collapse and a calibration scale for its repair. Its reported experiments support pre-merge screening and data-free PRISM repair within the tested configurations.

6 CONCLUSION

  • Across twenty-two merges from four model families, the interference score exceeded its threshold only for destructive merges, and twelve of fourteen pre-evaluation predictions were correct.PRISM kept above-threshold destructive merges near the base without data or tuning.Source: §6 · p. 9
  • PRISM soft-thresholds the averaged task vectors, using their variance across specialists to calibrate the repair.Source: §6 · p. 9
  • Because the score grows with training, model developers can track it to keep their specialists mergeable.Source: §6 · p. 9

ETHICS STATEMENT

The ethics statement notes that the work uses public checkpoints and benchmarks without human subjects or personal data, while warning that capability-focused evaluation does not measure safety behavior. The accompanying argument connects task-vector variance to the disturbance introduced by scaled averaging.

ETHICS STATEMENT

  • The work uses publicly released checkpoints and benchmarks, involves no human subjects, and collects no personal data; its capability evaluations do not measure safety behavior.Merged models need separate safety and alignment evaluation before deployment.Source: ETHICS STATEMENT · p. 9

A FULL PRISM ALGORITHM

  • The PRISM algorithm applies per-projection safeguards, including a 0.1% keep-rate floor and small numerical clamps for threshold variance and denominators.The procedure is summarized in Section 4 and illustrated in Figure 1.Source: A FULL PRISM ALGORITHM · p. 14

B DENOISE-ONCE-THEN-AMPLIFY: FULL ARGUMENT

  • Equation 1 defines I = P − S as empirical variance across scaled task vectors; under the working model, averaging reduces disturbance variance by K while applying coefficient c scales squared disturbance by c^2.The measured layer interference I_l supplies the plug-in screening scale.Source: B DENOISE-ONCE-THEN-AMPLIFY: FULL… · p. 14
  • In the coefficient sweep, baselines deteriorate at c≥0.7, with the worst reaching perplexity 961 at c = 0.7, while PRISM retains low perplexity across the tested sweep.PRISM thresholds the average before applying the coefficient.Source: B DENOISE-ONCE-THEN-AMPLIFY: FULL… · p. 14

C MECHANISM ANALYSIS

PRISM’s threshold varies with layer interference, while a cancellation gate bypasses shrinkage when dispersion dominates. The ablations examine how calibration granularity, threshold scale, specialist checkpoints, and model scale affect its behavior.

C MECHANISM ANALYSIS

  • At κ=0.5, layer retention and coherence correlate at r=0.91, while the ρ-gate returns the plain average when dispersion dominates.For fixed layer power and dimension, CR →1 implies σl →0 up to the numerical stabilizer; the advantage over Task Arithmetic vanishes on the single-specialist merge.Source: C MECHANISM ANALYSIS · pp. 14–16

D.1 CALIBRATION GRANULARITY

  • At c = 0.7, global and layer-wise threshold estimates score 67.8 and 67.9; at c = 1.0, they score 65.0 and 64.8.The six-task means are within evaluation uncertainty across the compared granularities; layer-wise calibration is retained to vary the threshold across depth.Source: §D.1 · p. 16, Table 5 · p. 16

D.2 THRESHOLD SCALE AND SCOPE

  • With the floor disabled at κ=1, the six-task mean is 68.6 at perplexity 8.9, within evaluation uncertainty of the floored operator.The 0.1% floor binds in every layer for κ≥1; on HumanEval, HumanEval+, MBPP and MBPP+, it adds 0.3 to 2.5pp over the no-floor scores.Source: §D.2 · p. 16, Table 6 · p. 16

D.3 INTERFERENCE DISTRIBUTION

  • Across 28 MLP layers, the standardized aggregate interference has excess kurtosis 0.44 and near-zero skewness, with mild heavy-tail deviations beyond |z|>3.Layer coherence correlates negatively with mean interference (r=−0.41); the six-task mean stays between 67.3 and 67.8 across the κ sweep.Source: §D.3 · pp. 16–17

D.4 SPECIALIST CHECKPOINTS AND EVALUATION UNCERTAINTY

  • PRISM leads the listed merges with both instruction-tuned and base specialist pairs; Task Arithmetic scores 6.5 points higher with instruction-tuned specialists.Individual benchmark standard errors range from 1.7 to 2.2pp, while propagated error is at most 0.8pp for a six-task mean.Source: §D.4 · p. 17, Table 7 · p. 17

D.5 PRISM-DENOISE AGAINST PRISM-PRUNE

  • With two specialists, PRISM-Denoise leads PRISM-Prune by 3.7pp at 7B and 6.9pp at 1.5B; the variants tie at K =3.At K =1, Denoise equals Task Arithmetic because λl =0, while Prune trails by 3.2pp.Source: §D.5 · p. 17, Table 8 · p. 17

D.6 MODEL SCALE

  • At 3B and 14B, the below-threshold Instruct+Coder-Instruct merges keep their plain averages, which gain 0.8 and 1.0pp over the base.Source: §D.6 · p. 17, Table 9 · p. 18

D.7 LEWIS VARIANTS AND THE SCORING ABLATION

The scoring and baseline ablations probe when interference-based thresholding is useful and how it compares with alternative merge rules. They also examine low-rank simulations and distinguish conflict from task-exclusive updates.

D.7 LEWIS VARIANTS AND THE SCORING ABLATION

  • On the Math+Coder merge, activation-based LEWIS trails weight-sparsity LEWIS by 8.8pp; interference-based thresholding adds 1.3pp over activation-norm scaling.Denoising with the activation-based scale scores 10.4pp above activation-based pruning.Source: §D.7 · p. 18, Table 10 · p. 18, Table 11 · p. 18

D.8 DATA-DEPENDENT BASELINES

  • AdaMerging reaches parity with PRISM only at its tuned coefficient and scores 51.6 at shared c=0.5; Greedy Soup returns the base, and Fisher-weighted averaging scores 47.5.The compared data-dependent baselines use unlabeled C4 or WikiText-2 texts.Source: §D.8 · p. 18, Table 12 · p. 19

D.9 BASELINE FAIRNESS PROTOCOL

  • The baseline protocol fixes c=1/K for the main task-vector comparisons; PRISM uses κ=1.0 without setting-specific tuning.The GSM8K and MATH-500 generation evaluations fix TIES trim at k=0.2 and use weight-sparsity LEWIS.Source: §D.9 · p. 18

D.10 WHY BASELINES FAIL ON QWEN K =2

  • At c=0.5, DARE collapses at both tested drop rates; default Iso-C and all four evaluated WUDI-Merging configurations also collapse.Iso-C at c = 0.1 reaches 66.6, 1.2pp below untuned PRISM. Community SLERP collapses, while Model Stock loses 14.8pp at perplexity 35.Source: §D.10 · pp. 18–19

D.11 SCALING-COEFFICIENT SWEEP

  • At c=0.7, Task Arithmetic, TIES, DELLA and LEWIS have perplexities 961, 100, 116 and 105; at c=1.0, they have 4.6×105, 1.1×104, 1.6×104 and 5.9×103.The unmerged base has a harness mean of 68.1 and scores 83.8, 38.2, 56.7 and 63.8 on GSM8K, MATH-500, HumanEval and MBPP.Source: §D.11 · p. 19, Table 14 · p. 20

D.12 LOW-RANK SPECIALISTS

  • Across ranks 8 to 128, PRISM stays within 0.6pp of its full-weight six-task mean of 67.8 and outperforms Task Arithmetic on all three measures.Source: §D.12 · p. 19, Table 15 · p. 20

E COHERENCE RATIO AND SIGN AGREEMENT

  • For K =2, layer coherence and sign agreement correlate at r =0.998, but sign agreement alone does not determine interference magnitude.The coherence-derived scale relates interference to power as σ2_l = I_l = P_l − S_l ≃ P_l(1 − CR_l), neglecting the stabilizer.Source: E COHERENCE RATIO AND SIGN AGREEMENT · p. 20

F CONFLICT AND EXCLUSIVITY

  • The exclusivity score distinguishes one-specialist concentration from similar magnitudes across specialists; thresholded elements are predominantly conflict elements, while many exclusive elements retain consensus.Using cancellation alone for calibration scores 60.8 at c=1.0, compared with 64.8 for PRISM.Source: F CONFLICT AND EXCLUSIVITY · pp. 20–21

G IMPLEMENTATION DETAILS

The default matched comparison merges MLP projections and leaves other listed components at their base values. In coding evaluation with instruction-tuned specialists, PRISM-Denoise matches the base on MBPP and leads the no-calibration-data merges on both benchmarks.

G IMPLEMENTATION DETAILS

  • The default matched comparison merges gate_proj, up_proj and down_proj while leaving attention, embeddings, layer norms and the output head at base values.MLP-only merging stays within 0.1pp of the attention-inclusive variant’s six-task mean with 84 merged keys rather than 196.Source: G IMPLEMENTATION DETAILS · p. 21

H CODING GENERATION

I CODING EVALUATION WITH EVALPLUS

Under the EvalPlus recipe, several merge baselines produce degenerate generations. The evaluation uses a prompt, code-extraction procedure, and test suites that differ from the lm-eval-harness recipe.

I CODING EVALUATION WITH EVALPLUS

J BREADTH EXPERIMENTS: DETAILS

Across the evaluated configurations, the interference score separated destructive from harmless merges and supported advance predictions. Broader tests included public specialist pairs, alternative model settings, continued pretraining, and increasing expert counts.

J BREADTH EXPERIMENTS: DETAILS

  • Among 22 evaluated configurations, seven were destructive and 15 harmless; the 1.9×10−3 threshold was placed in the gap between harmless and destructive scores for the nine unstarred configurations measured first, and 12 of 14 pre-evaluation predictions were correct.Source: J BREADTH EXPERIMENTS: DETAILS · p. 22
  • All 28 screened public specialist pairs scored below half the threshold, ranging from 0.002×10−3 to 0.853×10−3.The screen covered Qwen2.5-7B, Llama-3.1-8B, and Mistral-7B pairs across varied specialist domains.Source: J BREADTH EXPERIMENTS: DETAILS · p. 22
  • At one billion tokens and learning rate 10−4, the continued-pretraining merge was harmless at a score of 0.699×10−3; six-task means were 55.4 for Task Arithmetic, 56.2 for PRISM, and 56.1 for the base.The Japanese and biomedical specialists were trained from Qwen/Qwen2.5-1.5B and merged as a two-specialist pair.Source: J BREADTH EXPERIMENTS: DETAILS · p. 22
  • Magnitude statistics reached AUC 0.98 or higher, and total and signal power separated all 22 outcomes, while both direction statistics ranked destructive merges below many harmless ones.The statistics use parameter-count-weighted layer means; Table 18 compares rival pre-merge statistics across the evaluated configurations.Source: J BREADTH EXPERIMENTS: DETAILS · p. 22, Table 18 · p. 23
  • On the six-specialist Qwen2.5-7B merge, the base, Task Arithmetic, and PRISM scored 68.1, 68.0, and 68.2, respectively, with an interference score of 0.005×10−3.The specialists had conflicting objectives.Source: J BREADTH EXPERIMENTS: DETAILS · p. 22
  • For the Llama-3.1-8B merge of OpenMath2, Hermes-3, and Llama-Guard, the data-free ρ-gate detected a dispersion-dominated regime at ρ=0.021 and returned Task Arithmetic.The merge used three specialists with unscaled task vectors.Source: J BREADTH EXPERIMENTS: DETAILS · p. 22, Table 19 · p. 23
  • As matched-recipe self-SFT expert count grew from 2 to 8, ρ rose monotonically from 0.34 to 0.70, while every method remained within a point of the base.Within this curve, neither increasing K nor increasing ρ indicated harm.Source: J BREADTH EXPERIMENTS: DETAILS · p. 23, Table 20 · p. 23

Related papers

Loading 2610.03199v1…