acceptodds
Under review as a conference paper at ICLR 2027

Intervention Effects Are Execution-Relative

Abstract

Representation-level interventions are commonly compared and tuned through downstream task performance. Yet an intervention-induced state does not directly produce a task output: it is converted by a final execution rule, such as thresholding, decoding, suppression, or candidate scoring. We ask whether conclusions about an intervention family remain stable when this downstream rule changes while the intervened states are held fixed. We call this dependence execution-relative intervention effects and introduce matched-state evaluation, which computes each intervention state once and reuses the same cached state across task-valid executors. Across sound, language, vision, and speech tasks, changing only the executor can alter the measured effect size and, in stronger cases, reverse the ordering of intervention settings or change the selected setting. In sound event detection, fixed-threshold and hysteresis execution give opposite local orderings to the same intervention transition and select different settings under cross-fitting; transferring the fixed-threshold choice to hysteresis incurs a Event-F1 loss on held-out data. Boundary manipulations trace the fixed-threshold drop to sparse crossings of event-formation boundaries, revealing candidate instantiation as one mechanism. A complementary Qwen3-8B study keeps candidate statistics fixed and changes only the scoring rule, showing how candidate scoring reshapes competitive margins and intervention responses while preserving the selected coefficient. These results show that intervention evaluation and selection should account for the executor that turns intervened states into task outputs, and should test whether comparative conclusions remain stable across plausible task-valid execution rules.

Primary area:
interpretability and explainable AI
Keywords:
Activation Steering, Representation-Level Interventions, Downstream Execution, Execution Relativity, Sound Event Detection

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_t7c6OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines how the choice of downstream executor affects the evaluation of an intervention. The authors distinguish three levels of impact: changes in the measured effect size, changes in the ranking of intervention settings, and changes in the setting ultimately selected. Experiments in sound event detection (SED), language modeling, image segmentation, and speech recognition demonstrate that these phenomena can occur.

Critical strengths
  1. The three-level framework provides a useful way to characterize how downstream executors affect the evaluation of interventions.
  2. The paper examines executor-dependent effects across a diverse set of tasks and output types.
Critical weaknesses
  1. Although the experiments cover multiple tasks, they offer limited depth within each one. The tasks use different models, intervention families, and executors. Together, the results show that executor-dependent effects can occur, but they do not establish how often they arise within a given domain or how consistently they appear across intervention methods. The current evidence does not rule out the possibility that the observed effects are specific to the chosen models, interventions, or executor combinations.
  2. The strongest evidence for changes in the selected intervention setting comes from SED. Object detection shows a small local reversal in the ranking of settings, whereas the Qwen3-8B and CTC experiments change the measured intervention response without changing the preferred intervention coefficient. Thus, the experiments support the existence of selection changes but provide limited evidence about their prevalence or practical significance across models and tasks. This limits the support for the paper’s broader conclusions about intervention selection.
Questions for authors

In the Qwen3-8B experiments, did varying the executor produce any local reversals in the ranking of intervention settings or changes in the setting ultimately selected? Can these phenomena be observed across different pretrained models?

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

This paper examines how downstream execution rules affect the measured effects, rankings, and selection of intervention settings. It compares different rules using the same cached intervention states across several tasks.

Critical strengths

The method is partly clear.

Critical weaknesses

Weaknesses:

  1. It is a natural way to isolate the upstream effects by keeping the upstream intervented states fixed while changing only the executor. Why is this presented as a contribution lying on the 2nd item in the Contribution part?

  2. It is a much common conclusion that changing post-processing or decoding rules can affect performance and even reverse the comparison result. I do not see what major conclusion added beyond the expected sensitivity to the evaluation pipeline.

  3. The cross-task experiments use different models, interventions, and execution rules. However, it remains unclear which findings generalize beyond these particular setups, or how the analysis helps people to choose or improve an execution rule. The main recommendation seems to remain “retune for each executor,” which provides limited guidance in standard practice.

Questions for authors

See above

Reviewer_UcgFOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper examines the robustness of representation-based interventions to the decoding and scoring methods through which they are evaluated. The authors asked what conclusions concerning interventions hold once decoding and scoring (which they call the "executor") change, and distinguish between several possible changes: a chance in effect size, in ordering, and in argmax. They examine results for settings across several modalities, and find that indeed changing the executor can change the conclusions regarding the intervention method. They conclude by arguing that accounting for the executor is critical when evaluating intervention methods.

Critical strengths
  • The paper addresses the important topic of robustness in evaluating mechanistic interpretability interventions, a topic that I see as crucial for the community to focus on in order to improve its methods and inferences.
  • The paper tests its conclusions across different modalities.
Critical weaknesses
  • Unfortunately, I do not think that the contribution of this paper is very novel or significant. Despite its jargon, the paper describes basic control procedures in which the decoder or scoring method are changed, which leads to rather trivial changes in the conclusion that are not necessarily related to the fact that this is an intervention based experiment (e.g., the fact that changing a likelihood scoring rule will change the top answer is hardly surprising). The paper also presents basic procedures, e.g., applying the different decoders tested to the same activations (which they term "matched-state evaluation"), as unique contributions. Overall, I must admit that unfortunately I do not see the rather simple tests and conclusions the authors are offering here as particularly informative.
  • The paper is written in an unnecessarily dense and hard-to-read way: it defines terms like executor instead of using accepted terms like decoder, and comes up with new terms like effect-size, order, and selection relativity which simply describe possible conclusions from their experiment. The use of this jargon, as well as the paper's style more generally (which reminds a lot of LLM writing), obfuscates the rather simple points the authors are trying to make.
Questions for authors

None.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.