acceptodds
Withdrawn from ICLR 2027

Not All Drift Is a Mechanism: Tracing Prompt Adaptation in LLMs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

12 views

Abstract

Prompt-based adaptation offers several ways to specialize a frozen language model, ranging from natural-language prompt optimization to learned continuous prompts. These methods are usually compared through downstream performance, although similar outputs need not arise from similar internal computations. We study how discrete prompt optimization, prompt tuning, and prefix tuning alter the processing of dense and mixture-of-experts language models. Our analysis shows that the location of the adaptation predicts how task information is carried to the output: optimized text primarily changes response construction, input-level prompts write their influence into content representations, and layer-wise prefixes maintain direct control through attention. In MoE models, these differences further induce distinct expert-routing patterns. At the same time, common representation-level measurements respond strongly to generic context changes and do not reliably distinguish useful adaptations from matched placebos. We argue that understanding prompt adaptation requires separating representation drift from the part of that drift that is aligned with, and causally used by, the model's task computation. This perspective provides practical guidance for selecting, evaluating, and interpreting prompt adaptation methods.

Primary area:
interpretability and explainable AI
Keywords:
prompt adaptation, prompt optimization, mechanistic interpretability, causal interventions

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_MEGtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors examine three prompt adaptation methods on a text classification task, assigning civility labels (e.g., “insult” or “threat”) to online comments (Appendix D). The prompt adaptation methods are GEPA, prompt tuning, and prefix tuning, which differ by “insertion site,” e.g., GEPA optimizes hard prompts and prompt tuning optimizes soft prompts. The paper delineates several observations that distinguish the methods’ impacts on select models on the task, such as differences in the assignment of experts in MoE models (Section 4).

Critical strengths

The two research questions the authors pose (on pages 3 and 7) are interesting. I believe these questions are worth exploring, and I would be happy to read an ICLR paper investigating them.

Critical weaknesses
  1. In my view, the paper's severe readability issues warrant a careful revision, without which the work is difficult to consume, evaluate, or reproduce. For instance, several terms and/or variables are non-standard, undefined, or ambiguous, and some details vary throughout the paper for reasons unclear. Additionally, there are confusing references to an earlier workshop draft (Appendix D, Appendix E.4, Appendix E.7) that include unexplained differences in settings.

  2. It seems to me that the first of the two research questions (Section 3) remains unanswered, and the experimental design does not seem adequate to address such inquiry. While, indeed, the three methods differ by insertion site, the results do not appear to causally attribute downstream effects to this distinction.

  3. The results of Section 3 are performed on only one small model, namely Gemma-2 2B. This is not a deal-breaker, but it limits the generalizability of the findings.

Questions for authors
  1. Can the authors clarify the details of the dataset being used? I believe CivilComments has ~2M rows, with ~100k in the test set, but Appendix E.1 describes the test set as 2000 rows. Please let me know if I am misunderstanding.

  2. What is the authors’ intention with Section 3.2? It adds padding to the seed prompt to (1) rule out that the inclusion of generic context improves task performance and (2) rule out that task performance increases monotonically with distance from the seed prompt (Figure 7a), but I am not sure why a reader would expect either of these possibilities.

Reviewer_zk7hOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how discrete prompt optimization, prompt tuning, and prefix tuning interfere the processing of dense and MoE LLMs.

Critical strengths

N/A

Critical weaknesses

This paper is AI slop. Every sentence introduces new things that have absolutely no connection. Return the paper for rewriting. I will fight to have the paper rejected.

Questions for authors

N/A

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

This paper aims to study the underlying mechanisms of three different prompt optimization methods. Both dense models and MoE models are tested.

Critical strengths
  • This paper studies different models with many different analyses.
Critical weaknesses
  • The paper is very difficult to understand. There are many terms introduced without definition. For example, what is residual movement? What is shift energy?
  • Many results are listed in short, compact sentences without interpreting what the results mean or connecting the results to a hypothesis.
Questions for authors

Can the authors define the new terms that are introduced in the paper? What is a generic context drift? What are task-aligned changes?
What practical guidance does this paper provide to the field?

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

Related papers

Discussion (0)

Sign in to comment.