acceptodds
Withdrawn from ICLR 2027

Does Prompt Sensitivity Improve Cross-Family Degradation Prediction?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

14 views

Abstract

Models whose answers barely change when instructions are reworded might also hold their accuracy when the data change. Does prompt sensitivity predict how much accuracy a new model family will lose, beyond source accuracy? We test two source-side features, accuracy dispersion and label-free disagreement across ten wordings, on fifteen open-weight models, three classification tasks and seven natural shifts. Both lower held-out error under cross-validation over five development families. Under one frozen fit, disagreement raises error by 6% in a first held-out cohort and lowers it by 35% and 30% in a second and a third, of three and two families; dispersion has the same signs. No cohort reaches the registered level, and no test pools them. Disagreement's falls in the second and third cohorts exceed a within-source permutation reference, but in development the invalid-response rate alone predicts degradation nearly as well, and the first cohort's rise reverses when one training family is dropped. The evidence does not establish that the gain transfers to a new family, nor that sensitivity carries no information about shift.

Primary area:
foundation or frontier models, including LLMs
Keywords:
prompt sensitivity, distribution shift, accuracy-on-the-line, degradation prediction, held-out model families

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_MmZHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies whether changes in answers across prompts help predict a model’s loss of accuracy on different data. The authors test fifteen language models on three classification tasks and seven dataset shifts. They measure changes in accuracy and disagreement between answers across ten wordings. They test whether adding these measures improves predictions for model families excluded from training. The measures help in two test groups and hurt in another, leaving their usefulness for new model families unclear.

Critical strengths
  1. The paper asks a useful practical question. Predicting how much accuracy a new model will lose could reduce the need to test every model on every dataset.

  2. The study covers several models and tasks. It evaluates fifteen models, three classification tasks, and seven dataset shifts.

Critical weaknesses
  1. The writing makes the main findings harder to assess. Long sentences and repeated labels interrupt the explanation. The reader must work through many qualifications to understand what each experiment establishes and why it matters.

  2. The paper adds two simple measures to a standard regression model. Its value therefore rests mainly on the empirical findings, which do not yet give a clear explanation or a reliable prediction rule.

  3. The paper gives limited guidance on when sensitivity will help. It reports improvements in two groups and worse predictions in another. The checks identify factors that change the results, but they do not give a clear way to predict which outcome to expect for a new family.

  4. The predictor is trained on only five model families. Removing the Qwen 3 family changes both sensitivity measures from hurting to helping in the first held-out group. This makes that result difficult to treat as stable.

  5. Different kinds of prompt changes are combined into one measure. The wordings change label order, whitespace, wording, and other features, sometimes together. The paper does not isolate which changes provide useful information. This limits what we learn about the source of the measured sensitivity.

Questions for authors
  1. Does either sensitivity measure improve predictions when the baseline already includes both source accuracy and the invalid response rate? This would help establish what sensitivity adds beyond response formatting.

  2. How much do the results change when the evaluation items are resampled? The current checks vary the prompt wordings, but do not show how much the findings depend on the particular 180 items selected.

  3. How does adding prompt sensitivity compare with adding model size to the baseline, using the same held-out groups? This would help establish its value over a simpler model property.

  4. Under the same number of model calls, is measuring sensitivity more useful than evaluating more source examples to estimate accuracy? This comparison would help establish whether repeated prompting is a useful use of the evaluation budget.

Reviewer_wNNXOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates whether prompt sensitivity can be used to predict accuracy degradation in unseen model families relative to the source accuracy. The paper compares ridge regressions using source accuracy and task/shift indicators with augmented regressions that additionally use either accuracy dispersion or prediction disagreement across ten instruction variants. Using fifteen open-weight models, three classification tasks, and seven natural shifts, the authors test these features across cross-validated development folds and three held-out cohorts. Consequently, the evidence does not establish that performance gains from prompt sensitivity reliably transfer to new model families.

Critical strengths
  1. The research question that whether prompt changes can help predict the model's accuracy degradation on different datasets is interesting and worth analyzing.
  2. The paper provides empirical analysis on a wide range of open-source models and datasets, which is informative, although it was likely done using AI research tools.
Critical weaknesses
  1. The paper lacks novelty and empirical contribution. The paper aims to answer the original research question by conducting analytical experiments across many open-source models and different datasets. However, it does not provide a clear conclusion on the usefulness of sensitivity for predicting degradation across different model families. Since no clear novel contribution is identified, the paper also does not propose any future work based on the current observations that could potentially make a greater contribution.
  2. The presentation of this paper is of poor quality. The expressions, such as the claims, arguments, and observations, are difficult to follow. The writing style makes it difficult for readers to understand. For example, the paper lacks clear definitions or explanations of several terms such as "families", "cohorts", "reading", and "resolution" before repeatedly using them. The discussion of experiments also moves between measured effects and protocol-specific interpretation rules without consistently distinguishing them.
Questions for authors
  1. What conclusion about transfer across model families can the primary design establish, given that every cohort is too small to attain the chosen significance threshold? Could the authors clarify what substantive conclusion is supported beyond the descriptive results on the particular models evaluated?
  2. Could the authors disclose a more specific usage of AI assistance across the research process, including the preparation of the supplementary materials? The paper's AI usage statement only describes the authors' contribution to idea generation, which does not meet this conference's standards. Please clarify whether AI tools or autonomous research agents contributed to idea generation, cohort protocols, and interpretation rules. Please also explain how the authors verified the code, planned the methodological implementations, and provided analysis and conclusions based on the empirical observations.
  3. What scientific insights or practical guidance can be derived from the current observations, and how could they inform future research on robustness evaluation or degradation prediction? Section 7 suggests larger cohorts and more development families, which is fine. But could the authors provide a specific follow-up hypothesis or methodological solution motivated by the present findings?
Reviewer_AiWFOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies whether the prompt sensitivity of LLMs helps predict their drop in accuracy from in-domain to distribution-shifted data, compared to a baseline that takes into account only in-domain performance, and pre-defined task and distribution-shift labels. The authors suggest two measures for prompt sensitivity (accuracy dispersion across prompt templates, and disagreement between wordings) and use them to build performance drop predictors for distribution-shifted data in cross-validation (train predictor on 4 model families, test on a 5th) and held-out settings (train predictor on 5 development set model families, test on 3 held-out sets with 2-3 families). The authors find that adding prompt sensitivity helps in cross-validation and two of the held-out sets, while it hurt in one held-out set, though the experiments were designed in such a way that due to the small size of held-out sets, statistical significance could not be established.

Critical strengths
  1. The paper is well motivated in that we'd like to be able to cheaply and reliably predict performance under distribution shifts, and that prompt sensitivity is a variable that could plausibly improve such prediction for LLMs.
Critical weaknesses
  1. The paper's main results are inconclusive, partially due to flawed experimental design. As such, it is unclear what to take away from the paper. Since each held-out set only covers 2-3 model families, statistical significance could by design not be established for the held-out set comparisons. On top of that, the results are mixed across the 3 held-out sets (prompt sensitivity helps in two, but hurts in one). Furthermore, the results hinge on LLMs used as judges without human validation for key methodology such as verification of prompt equivalence, which adds further uncertainty.

  2. The paper aims to make relatively general claims about performance under distribution shifts of language models, but the experimental setup as-is cannot reliably support those. The paper covers 15 LLMs from approx. 2024-2025 that are nowadays in the SLM (small language model) range, and only classification tasks. The paper also relies on pre-defined distribution shift labels - adding a new shift type would require us to score all the training models on it and re-train the predictor, which is not necessarily more practical than simply evaluating a new model directly on the shifted data.

Questions for authors
  1. What motivated your particular selection of models/model families?
  2. Have you tried to extend your setup to generative tasks?
  3. How did you ensure the reliability of LLM judges, and that of LLM agents within the experimental loop?

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

Related papers

Discussion (0)

Sign in to comment.