Resampling a Merge to See Where It Breaks: Calibration of Many-Model Merges

Model merging combines several fine-tuned copies of the same base network into one set of weights by adding up their "task vectors," the weight differences each fine-tune introduced. It lets one checkpoint do the work of many, and recent methods scale it from two models to sixteen or more. Nobody has checked whether the merged model still knows when it is wrong. That matters because merging fails unevenly. Some inputs land where task vectors cancel or fight, and the merged model gets them wrong. A single global confidence correction such as temperature scaling cannot tell those inputs apart from the ones the merge handles well.

This note proposes two things. The first is to measure how confidence calibration and error detection degrade as the number of merged models grows. The second is to test a specific, cheap fix. Most scalable merges already contain randomness or arbitrary choices, such as which delta weights to drop and which sign wins a conflict. If we resample those choices a few times, the resulting sibling merges should disagree mainly on the inputs that interference damaged. The claim is that this disagreement flags interference-induced errors better than temperature scaling, verbalized confidence, or a probe trained on the merged model. It also costs a small constant number of forward passes, independent of how many models were merged.

Core Claim

Many-model merging produces two distinct calibration failures, and they need different remedies. The first is a global shift in logit scale. Averaging shrinks task-vector norms, which is the triangle-inequality effect documented by MAGIC. Rescaling, as in DARE's $1/(1-p)$ factor or MAGIC's norm restoration, pushes the scale back. Temperature scaling should largely repair this first failure. The second failure is a loss of discrimination: the merged model becomes confidently wrong on the specific inputs where task vectors conflict. No monotone rescaling of logits can fix this, because such a rescaling preserves the ranking of confidences.

We hypothesize that the second failure grows with the number of merged models $N$. It should therefore be visible in selective-prediction metrics, such as the area under the ROC curve for separating correct from incorrect answers, even after temperature scaling has driven expected calibration error (ECE, the average gap between stated confidence and accuracy) close to zero. We further hypothesize that disagreement among resampled merges is a targeted signal for it.

Concretely, given base weights $\theta_{pre}$, task vectors $\tau_i$, and merge coefficients $\lambda_i$, each sibling merge is

$$\theta^{(k)} = \theta_{pre} + \sum_{i=1}^{N} \lambda_i \, \frac{m_i^{(k)} \odot \tau_i}{1-p},$$

where $m_i^{(k)}$ is the $k$-th random keep-mask for task $i$ with drop rate $p$, in the style of DARE. For TIES-based merges, a sibling additionally resamples sign election on parameters whose consensus margin is small. The prediction is the average of the $K$ siblings' output distributions, with $K$ around 4 to 8. The spread among siblings serves as the uncertainty score.

A cheaper variant applies masks to the already-merged delta, so only the base and one delta need to be stored. Comparing the per-task and merged-delta variants tests whether knowing which task each weight came from matters.

Relation to Prior Work

The closest paper is Navigating the Alignment-Calibration Trade-off. It interpolates between one pretrained model and its instruction-tuned version and finds a coefficient range that improves both accuracy and ECE. It motivates this work by showing that weight-space operations move calibration. It is strictly a two-model interpolation, however, and it uses merging as the calibration fix. We ask the opposite question: what does merging many different specialists do to calibration, and can the merge's own degrees of freedom expose the damage?

Model Merging by Uncertainty-Based Gradient Matching uses Fisher information as parameter uncertainty to decide how to weight updates. It reports accuracy, not predictive calibration. We use it as a strong merge whose outputs we calibrate, and we test whether its parameter-space uncertainty translates into output-space calibration. HARC and MAGIC use "calibration" to mean matching router probabilities or weight norms, not confidence. MAGIC constrains our design, because its norm restoration may inflate logits, so it is one of the merges we audit. RESM merges honesty-oriented models but measures hallucination rates rather than whether confidence tracks correctness.

On the calibration side, PROBECAL and Large Language Models Must Be Taught to Know What They Don't Know show that hidden states carry correctness information that small trained probes can extract. These supply our strongest competitor rather than our method. The scaling setting comes from FW-Merging, which merges 16 or more models. It also comes from Demystifying Mergeability, whose 20-task vision pool and finding that many-way interference is not captured by pairwise metrics motivate looking at the input level instead.

Research Program

The first stage is measurement. For each merge method, we sweep $N$ over 2, 4, 8, 12, 16, and 20. We report per-task accuracy, ECE, negative log-likelihood, Brier score, and error-detection AUROC, both before and after temperature scaling. Temperature is fit either globally or per task on held-out data. We also label every test input as an "interference error" if the merged model gets it wrong while the relevant specialist gets it right. This gives a ground truth for the failure we care about, separate from errors the specialist would have made anyway.

The second stage builds the resampling estimator for Task Arithmetic, TIES, and DARE merges and tests the per-task and merged-delta variants.

The third stage is a mechanism analysis. We ask whether sibling disagreement concentrates on inputs whose activations load heavily on parameters with sign conflicts or high overlap. We also ask whether RL-trained specialists, which Enough is as good as a feast found degrade less under merging (7 to 8 percent mean decline versus 18 to 22 percent for supervised fine-tuning), also degrade less in discrimination.

Baselines

The naive reference point is the merged model's raw softmax confidence with no correction. Beside it we report each specialist's own calibration, since merging could plausibly inherit specialists that were already overconfident.

The first real baseline is temperature scaling. We run it with a single global temperature and with one temperature per task. Both are fit on the same held-out split that every other method may use, which is 500 labeled examples per task, matching PROBECAL's budget. Temperature scaling is the strongest cheap competitor for ECE. The proposal beats it only if it improves error-detection AUROC or ECE on interference errors, since temperature scaling cannot change AUROC.

The strongest existing method for detecting errors is a PROBECAL-style probe. This is a 3-layer classifier on frozen hidden states of the merged model, trained on the same 500 examples per task. The only reported number in the file is on a different setting: PROBECAL reduced ECE from 0.690 to 0.035 on TabMWP with tool-using models (PROBECAL). No reported result was found on Lacuna for calibration or error detection of merged models at $N > 2$. Beating the probe means higher AUROC on interference errors at matched labeled data. Alternatively, matching it with no labeled data, since resampling needs none, would also be a clear win.

For LLM experiments we add two more baselines. One is base-to-merge SLERP interpolation, following Navigating the Alignment-Calibration Trade-off, which reports a gain of over 4 points in MMLU-Pro accuracy on Gemma-3-27B alongside lower ECE but gives no scalar ECE in the file. The other is verbalized confidence.

The reference ceiling is a full ensemble of the $N$ specialists, which costs $N$ forward passes.

Two ablations isolate the key ingredient. The first replaces task-vector masks with Gaussian weight noise of matched norm, applied to all weights. The second applies MC-dropout on activations of the merged model. Both use the same $K$. If either matches resampling, then the signal is generic stochasticity and not merge structure, and the core claim fails.

For accuracy context, the reported vision anchors are as follows. Task Arithmetic retains 91.2% of specialist performance at 8 tasks (Editing models with task arithmetic). TIES gains 1.8% on ViT (TIES-Merging). FW-Merging gains 15.3% at 16 models.

Why not just…?

Why not just temperature-scale the merged model? If the only damage is a shift in logit scale, this is the right answer. The first stage of the program is designed to find out whether it is. The file offers only indirect evidence that it is not enough: Reducing Supervision Uncertainty Induces Model Miscalibration argues that a single temperature cannot restore instance-level uncertainty. We treat this as a live threat rather than a settled point. If per-task temperature brings interference-error AUROC to within noise of the resampling estimator, we will report merging's calibration cost as a scale problem and drop the method claim.

Why not just train a probe on the merged model's hidden states? Probes are strong and single-pass, and they may well win on in-distribution tasks. Their known weakness is that they go stale when the model changes, and each new merge configuration is a new model that needs fresh labels. Resampling needs no labels. The comparison at matched labeled data, and at zero labels for the probe's retraining after changing $N$, settles which approach is more practical.

Why not just ensemble the specialists? An ensemble costs $N$ forward passes and $N$ stored models, which is the overhead merging exists to avoid (FW-Merging). We include it as a ceiling. The target is to recover most of its AUROC advantage over the raw merge at $K \ll N$.

Why not just ask the model for its confidence? Large Language Models Must Be Taught to Know What They Don't Know shows that verbalized confidence is unreliable in open-ended generation and is sensitive to prompt wording. We include it on the LLM tasks only as a reference.

Proposed Experiments

The main testbed is CLIP ViT-B/32 on the 20-task vision pool from Demystifying Mergeability. This pool includes Cars, DTD, EuroSAT, GTSRB, MNIST, RESISC45, SUN397, and SVHN. Classification makes ECE and AUROC unambiguous, and a 20-model ensemble is affordable as a ceiling.

The language testbed has two parts. One reuses the supervised versus RL specialists across math, code, IFEval, puzzles, and ranking from Enough is as good as a feast. The other is the WizardLM, WizardMath, and Code Alpaca merge on Llama-2-13B from DARE. Confidence is measured as answer-letter probability on MMLU and GPQA and as final-answer sequence probability on GSM8K.

The first decisive check is whether temperature-scaled AUROC falls as $N$ grows. If it stays flat, then the discrimination failure does not exist and stage two is moot.

The second decisive check is whether resampling beats temperature scaling, the probe, and both noise ablations on interference-error AUROC. It must do so at $K \le 8$, with the gap widening as $N$ grows.

The third check is whether the per-task masking variant beats the merged-delta variant. A win would mean that task identity carries the signal.

The fourth check is whether MAGIC-style norm restoration raises raw ECE relative to plain averaging, and whether RL-trained specialists lose less AUROC than supervised ones under the same merge.