acceptodds
Under review as a conference paper at ICLR 2027

Measure the Merge in the Currency You Deploy: Calibrated Path Decisions on a Budget

Abstract

Merging two fine-tuned models by weight averaging can quietly destroy the task: averaging two capable, officially released Qwen2.5 specialists can leave a model that solves nothing, while another pair from the same family survives — and nothing readable from the endpoints, from task labels to weight distance to training logs, says which is which. Part of that cost is exact rather than empirical: the conventional 50/50 average forfeits half the endpoint gap on any asymmetric pair, because net gain over the better endpoint decomposes pointwise as dip depth minus an endpoint-asymmetry wedge. The rest must be measured along the interpolation path, but path evaluations cost GPU time: how many to buy, and what to read them in? How many: a midpoint accept/reject decision is a three-evaluation problem — three points plus one calibrated safety margin, which transfers unchanged to three external populations and is measured wider on the fourth; a denser grid raises coverage on paper but admits no more pairs at the same calibrated risk. For finding a coefficient, density buys real structure at a minimax-optimal k⁻² rate that flattens past about six queries. In what currency: at matched wall-clock budget, rules that read the proxy-loss curve repeatedly committed coefficients below the better endpoint — the worst landing −17.0 points below it — while rules that measured the task metric directly did so far less often; where paths are flat, nothing separates the two currencies. The coefficient, not the recipe, is the operative variable: TIES and DARE collapse at full strength and recover at the measured coefficient where the path allows it. Read the path's depth once, then measure the merge in the currency you deploy.

Primary area:
foundation or frontier models, including LLMs
Keywords:
model merging, interpolation paths, loss landscapes, calibrated risk screening, conformal prediction, query complexity, fine-tuned language models, evaluation protocols

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ctSnOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper seems to suggest a method for model merging. Unfortunately, despite a giving an earnest attempt at understanding this paper, I could not make sense of it.

Critical strengths

None

Critical weaknesses

The authors mentioned that LLMs were used to help with the writing process. Perhaps too much of the paper seems to be written by LLMs? I really could not make heads or tails of this paper.

For example, I already don’t understand the first line of the main technical section:

“For two fine-tuned models with weights θ0,θ1\theta_0,\theta_1, write L(α)=…L(\alpha) = … for the loss along the interpolation, under a specified dataset, metric, and data split. Change any of those choices and the object changes, so we treat a path as an evaluation-conditioned function, not an intrinsic property of two checkpoints”.

All I understood is that you define L to be the loss of the convex combination of two models. What does the additional commentary mean? I am similarly confused by every sentence I read.

I’m sorry I could not provide a more substantive’s review especially regarding the contents of the paper.

Questions for authors

Beyond majorly rewriting the paper, I do not have any suggestions for the authors.

Reviewer_GZVROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

the goal is to determine how to merge two fine tuned checkpoints. consider merging θ1\theta_1 and θ2\theta_2 model weights linearly. we can write this as (1−α)θ1+αθ2(1-\alpha)\theta_1 + \alpha \theta_2. how do we choose the merging coefficient α\alpha? the paper shows that α=0.5\alpha=0.5 is not usually a good default. along which metric do we eval α\alpha? the authors make an argument that you can figure out whether a 50/50 merge is safe by using two endpoint evals and a midpoint eval, and considering a safety margin that they caame to through a conformal prediction argument. you can also specify task specific metrics to choose α\alpha , and its usually worth doing so.

Critical strengths

the authors' main claim, as far as i can tell, is that one should measure a task specific metric (like correctness) and not the loss when assessing weighting in merged model checkpoints. as an example, you could consider the 8th grade math problems dataset (GSM8K) which is a super common NLP task. you could evaluate model checkpoint performance on cross-entropy loss at predicting next tokens during rollouts, OR you could eval the checkpoint on whether it gets the right answer during the rollout. the authors show that these two (unsurprisingly) do not always agree. if you blindly merge according to the loss, you can end up with a merged model that looks better on the loss but actually decreases the performance of the model.

to the authors credit, this is an interesting observation, and likely warrants further exploration. its not obvious to me that you would want to do a linearized parameter merge anyway, but if we assume we want to do that (its cheap, for one), then i think considering how to weight checkpoints is a reasonable research direction.

Critical weaknesses

this paper is almost unreadable. it is clearly unedited text from an LLM. stylistic tells are prominent in the abstract, which does not help me really get a sense of the paper or the contribution. some that particular stick out to me as non-sensical/unscientific "can leave a model that solves nothing", "50/50 average forfeits half the endpoint gap on any asymmetric pair", "the rest must be measured along the interpolation path, but path evaluations cost GPU time: how many to buy, and what to read them in?" this is basically AI slop.

I genuinely struggled to make it through the text in this paper, and for this reason, understanding the setup and contributions were unnecessarily difficult. for me, this is enough to make this paper significantly below the bar for ICLR, where i expect to be able to read papers from any area and understand at least the problem statement and overall contributions from the abstract and introduction.

however, as it is unclear to me what the line is for ai assistance given the new ICLR ai use policy, and to be fair to the authors if they have a good reason for the reliance on LLM text assistance, i made a best effort attempt to understand the contributions as they are.

the authors are able to demonstrate the motivating phenomena, but that is it. they choose small models, and small datasets, and their effect size for their main result is largely the result of a single model checkpoint pair for GSM8K. for PubMedQA, i dont think they are able to demonstrate a meaningful effect at all.

they give some "theory" but its obviously weak for the setting (the conformal certificate thing is clearly not meaningful with n=19).

Questions for authors
  1. why is the theory (in particular Theorem 1 and Propositino 1) you present worth including? can you explain clearly, given the actual parameters of your regime? can you help me digest its relevance to the problem of setting the mixture coeffficent?
  2. the PubMedQA are unconvincing. was there a problem with the setup? if its a small effect, you may need more checkpoints / models?
Reviewer_bvPhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper looks at how much evaluation you need to decide whether two fine-tuned models should be merged, and how to pick their mixing weight. It treats these as two separate problems, checking a fixed 50/50 merge, and searching for a better coefficient. For the first the authors propose a calibrated screen based on loss. For the second, they analyze how evaluating more coefficients gives better coverage of the interpolation path. Across several model families and setups, they compare picking coefficients by proxy loss with picking them by the actual task metric. In a GSM8K experiment with matched budgets, selecting directly by accuracy avoids several mistakes that loss-based selection makes, but a second experiment on PubMedQA shows little difference between the two. The authors suggest running a quick loss scan first to decide when it's worth spending more on task-metric evaluation.

Critical strengths
  1. The paper separates three things: checking a fixed midpoint, choosing a mixing coefficient, and testing the chosen model on held-out data. It also separates beating the average of the two endpoints' losses from beating the stronger endpoint. These distinctions make the evaluation more meaningful. The budget-matched comparison is especially useful because it includes simply picking one of the two original models as a baseline you could actually deploy, instead of assuming that merging has to help.
  2. The experiments keep coefficient selection separate from held-out evaluation, account for the cost of switching coefficients, and check how results change with the evaluation budget. The authors also openly report the dominant outlier, the null result on PubMedQA, and their failure to reproduce a cross-scale merging gain that first looked promising. This openness makes the paper stronger as an empirical study and helps show where its findings stop applying.
Critical weaknesses
  1. Using loss-curve shape to decide when to evaluate on the task metric hasn't been tested prospectively. The supporting correlation rests on eight units and GSM8K and PubMedQA differ in both task and adapters, so curve shape can't explain the gap. Gains over picking the better endpoint are also small and one unit drives 74% of the CE-11 vs. Direct-5 difference.
  2. Theorem 1 only bounds midpoint loss for exchangeable model pairs. It doesn't cover accuracy, other coefficients, or new populations. The recommended 0.05 margin is below the 95% conformal margin (0.0903) and covers only 29 of 32 Qwen2.5 units. The authors should state what risk the procedure controls and when to recalibrate.
Questions for authors
  1. See weaknesses
  2. What threshold separates a "flat" path from one that needs task-metric evaluation? Does the rule predict gains on new model pairs, or only explain the two pools after the fact?
  3. Given the small gains of Direct-3 and Direct-5 over Direct-2, when are the extra evaluations worth it? Can you compare the full procedure, including the loss scan, with endpoint-only selection at the same total budget?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.