acceptodds
Under review as a conference paper at ICLR 2027

Minimal Functional Modification: Minimal Behavioral Realization of Activation Interventions

Abstract

Activation interventions provide one internal realization of a desired language-model behavior, but need not provide a minimal one. We study minimal behavioral realization: given a frozen reference intervention function and a specified behavioral observation, what is the smallest hidden-state modification required to reproduce the reference behavior across prompting contexts? This problem is context dependent because demonstrations alter both model states and local sensitivities to intervention. We first characterize when directional modification is strictly more economical than optimal scalar contraction. More importantly, we show that pointwise redundancy need not be shareable across contexts: the excess cost of a shared finite-width realization is governed by the spectral structure and input predictability of the pointwise optimal intervention field. We further quantify the cost of preserving richer behavioral observations and establish conditions under which local advantages persist at finite intervention amplitudes. These results motivate Minimal Functional Modification (MFM), a shared nonlinear realization module trained directly to balance reference-output fidelity and hidden-state displacement without explicitly constructing Jacobians. Across two language-model families and three multiple-choice reasoning benchmarks, MFM reduces reference intervention displacement cost by 57.6% on average while preserving four-shot answer distributions. Across changed four-shot demonstration sets, MFM yields contextual-response errors of 0.0085–0.0103 across the six model–task settings. When answer-set probability mass is additionally constrained, MFM still achieves 43.4% displacement-cost savings on the evaluated subset.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, In-context Learning, Representation Geometry, Inference-time Adaptation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UPMkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to reproduce the behavior of an existing activation intervention with a smaller hidden-state modification. It provides theoretical analyses of directional optimization, shared module capacity, and the cost of preserving additional behaviors. Based on these insights, the authors propose MFM, which matches reference outputs while minimizing intervention magnitude. Experiments on two LLMs and three reasoning benchmarks show an average 57.6% reduction in intervention cost with little loss in answer fidelity.

Critical strengths
  • The distinction between pointwise optimal interventions and shared realizability is interesting. The theoretical analysis provides useful insights into how model capacity and context variation affect intervention compression.

  • The experiments show substantial reductions in intervention cost while largely preserving the reference behavior. The comparisons with scalar scaling and coordinate gating are helpful.

Critical weaknesses
  • The methodological novelty is somewhat limited. The main optimization resembles output distillation with norm regularization, while several theoretical results rely on standard ridge regression, projection, and low-rank approximation. Related works on steering geometry and minimum-distortion interventions have explored similar ideas. The paper could better explain what is fundamentally new beyond these existing approaches.

  • The evaluation is quite narrow. The experiments mainly preserve normalized answer distributions on multiple-choice tasks. It is unclear whether the savings would remain when preserving full output distributions or free-form generation. Also, the reduction in hidden-state displacement does not necessarily translate into lower computational cost.

Questions for authors
  • What is the main theoretical difference between MFM and existing geometry-aware or minimum-distortion steering methods? Could you provide a more direct comparison?

  • How much of the reported compression comes from redundancy in the initially trained reference intervention? Would the gains remain as large with a more strongly norm-regularized reference?

  • Would MFM still achieve substantial savings when matching full-vocabulary output distributions rather than only multiple-choice answer distributions?

Reviewer_qbpKOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether a reference activation intervention can be replaced by a smaller hidden-state modification while preserving its observed behavior. It develops a framework for understanding when such savings can be achieved by a shared module and reports an average 57.6% reduction in displacement cost across two models and three benchmarks.

Critical strengths
  1. The distinction between pointwise redundancy and shared realizability is the most interesting part of the paper. Theorem 3.2 gives a useful account of why a shared low-rank module may not be able to capture the savings available at individual inputs.

  2. The experimental results are fairly consistent across models and tasks. I also appreciate that the authors report cases where fidelity fails and examine what happens when the behavioral observation is strengthened.

Critical weaknesses
  1. The savings are measured using a narrow observation: the normalized distribution over answer tokens. Adding answer-set probability mass already reduces the savings from 56.6% to 43.4%. I would like to know whether the result holds when more of the output distribution is preserved.

  2. I am not convinced that Figure 3 isolates the capacity issue discussed in the theory. The failure occurs at width 16, while the reference module has width 64. It is therefore hard to tell whether this reflects the proposed realization tax or simply insufficient capacity.

  3. The choice of displacement metric also matters. The covariance estimate for ARC-C appears to use fewer states than the hidden dimension, so the shrinkage term could have a substantial effect. The Euclidean comparison in Table J.4 is helpful, but it covers only one setting.

Questions for authors
  1. How much of the reported saving remains if the observation includes more of the next-token distribution, rather than only the answer tokens?

  2. Does the width-16 module still fail the fidelity constraints when the displacement penalty is reduced toward zero?

  3. Could the authors report the Euclidean or diagonal-metric results for ARC-C as well? This would help assess how much the savings depend on the covariance metric.

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

The paper basically looks at how to reproduce a frozen activation intervention while making smaller changes to the hidden states. The theory compares changing the direction with simply scaling things down. It also examines how much capacity is needed to use the same realization across different contexts and what happens when more observations need to be preserved. MFM learns a nonlinear module that tries to keep the outputs similar while reducing hidden state displacement. The authors test it on two model families and three reasoning benchmarks, and the results show lower displacement while keeping the answer distributions similar. They also conduct some additional experiments by changing the demonstrations and adding constraints on the answer probability mass.

Critical strengths
  • I think this paper provides useful insight by separating output rank from coefficient predictability. An economical intervention may exist for each context but remain difficult for a shared module to represent. This gives the capacity analysis a clear motivation.
  • The section 4.4 tests preservation of answer probability mass as an additional constraint. This reduces probability mass drift while retaining displacement savings. I appreciate this control because matching renormalized answer distributions alone would miss that behavior.
  • The paper is organized well and easy to read. The explanation of theoretical scope in section 3.5 is helpful. The authors distinguish local guarantees from evidence obtained with nonlinear interventions. This makes it easier to understand what the experiments support without treating their observed agreement as a proof.
Critical weaknesses
  • I noticed that Figure 3 varies output width without separately measuring coefficient prediction error. I think fixing the output subspace and comparing learned coefficients with optimal coefficients would help distinguish prediction difficulty from limited rank. This would connect the experiment more directly to Theorem 3.2.
  • The nonlinear optimality gap in Equation 1 is unresolved by the comparisons in Table 1. Optimizing interventions independently for a small subset of prompts would provide an empirical reference. I think comparing this with MFM would clarify the cost of sharing a module.
  • The paper uses a fixed intervention layer for each model. This leaves sensitivity to intervention depth unclear. I think testing earlier and later layers under matched fidelity constraints would help establish whether the displacement savings depend on the chosen location.
  • The section 4.3 reports fidelity failures on LogiQA when the demonstration context grows, despite small accuracy changes. The cause remains unclear to me. Examining the affected answer distributions would help explain the transfer boundary and suggest when the shared module needs adaptation.
Questions for authors
  • Could you separate coefficient prediction error from output rank limitations in Figure 3? With the output subspace fixed, comparing predicted coefficients against the optimal coefficients would help isolate this source of error. This would strengthen the connection to Theorem 3.2.
  • Could you compare MFM with interventions optimized independently for individual prompts under the same fidelity constraints as Table 1? A small diagnostic subset would help estimate the additional cost of sharing a module. I would interpret this as an empirical reference, without assuming global optimality.
  • What distinguishes the LogiQA prompts that fail the fidelity criteria in section 4.3? Could you relate these failures to changes in the reference answer distribution when demonstrations change? I think this would help explain the transfer boundary and guide when the compressor needs adaptation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.