acceptodds
Under review as a conference paper at ICLR 2027

COUNTERFACTUAL PLASTICITY: RISK-AWARE SELECTION OF MODEL UPDATES BEFORE EXECUTION

Abstract

An update can teach a language model new information while harming behavior that was previously reliable. When several update procedures are available, the question is not only whether to update, but which update to apply. We introduce Counterfactual Plasticity (CP), a framework that predicts the likely benefit, collateral damage, and damage uncertainty of each candidate before it is run. CP calibrates the resulting selection rule with held-out split conformal risk control, then chooses the highest-benefit admissible candidate or abstains. Controlled experiments show that benefit is easier to forecast than damage, and that simple forecast-based selection can fail when the candidate menu has real benefit and damage trade-offs. In frozen factual-editing evaluations, CP substantially reduces harmful selections in two higher-risk regimes. In an independent Qwen2.5-7B evaluation where the baseline is already benign, calibration selects λ = 0 and leaves the baseline decision unchanged. Together, these results show how a model can compare plausible updates before committing to one, while adding conservatism only when the evidence supports it.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Model Editing, Model Updates, Post-Training Adaptation, Conformal Risk Control, Uncertainty Estimation, Risk-Aware Learning

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

The paper studies how an update candidate for a language model can be selected before running any of the available ones. The presented framework, Counterfactual Plasticity, predicts each candidate’s acquisition gain and collateral damage using only information available before updating. It excludes candidates based on a calibrated damage criterion, and then selects the highest-gain ranking one to realize. They find that gain is easier to predict than damage, and evaluate on a small controlled GPT-2 experiment, locked evaluations on TinyLlama and Gemma, and Qwen2.5-7B.

Critical strengths

S1: The problem is clearly defined, and the importance of the problem is well motivated.

Critical weaknesses

W1 (Engagement with literature): While the paper cites seven works in the related work section, there is no engagement with the literature in any other part of the paper, making it hard to set the contribution into context. The introduction’s motivating claim, that model updates are evaluated mostly retrospectively, is stated without providing evidence or supporting references.

W2 (Presentation): The decision-level calibration (lines 153-161) is hard to follow without prior knowledge of conformal risk control: α\alpha is not defined, and the main inequality is not derived or referenced. It is unclear to me, how choices for parameters like 90 disjoint calibration edits, α\alpha=.10, or the 61 numerical values for lambda are justified. Section 3 mixes the method with experimental settings for E1, splitting it and restructuring the paper would help make the paper clearer. The numbers in Section 4 and Figure 1 do not align for the Original menu.

W3 (Main paper is not self-contained): The forecaster is the component that determines whether the proposed framework is useful. However, the forecaster that predicts gain and damage from pre-update information is not described in the main paper. The guarantee holds only for one fixed training set and pipeline, and it is unclear to me how it applies in other settings.

W4 (Experiments):

  • The controlled GPT-2 experiments use a 10-fold CV-CRC procedure that does not inherit the E1 split-CRC guarantee, so it does not support any previous safety claims.
  • The regimes are not formally defined, making it unclear to me what “harder experiences” or “denser strenghtst” refer to in practice.
  • The primary comparator, B1, shares CP’s six-candidate action space and forecasting machinery and lacks only the calibrated exclusion, so the experiments isolate the effect of calibration, not the value of pre-execution forecasting itself, which is the paper’s motivating question.
  • R_CRC is undefined.
  • There is no reference point for how much any update could improve the model, and baselines from the literature are missing, so the numbers are hard to interpret and it is difficult to assess the achievement of the proposed method.
Questions for authors

Q1: Are the 90 calibration edits representative of the test edits?

Q2: How was α\alpha = 0.10 chosen, and what happens to harm and gain if you use α\alpha = 0.05 or 0.20?

Q3: Calibration requires running every candidate on every one of the 90 calibration edits. How many update runs is that in total, and after how many future edits does forecasting become cheaper than simply running all candidates each time?

Q4: How do the results transfer to other training data or forecasters? Which assumptions does the forecaster have to fulfill?

Q5: Which loss defines gain and damage, and in what units? What does a gain difference of 1 mean in practice?

Q6: What is the gain of the best possible choice for each edit (including abstaining when no candidate is safe), and what percentage of it does CP reach? How does a simple rule that needs no forecaster, such as always using the same fixed setting, compare?

Reviewer_ho2COpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper focuses on LLM model updates, which can be regarded as counterfactuals. It proposes a framework which predicts the potential benefit, potential damage, and damage uncertainty of counterfactual (CF) updates. It proposes a risk-aware update selection procedure to reduce harmful updates.

Critical strengths

The paper actually targets a meaningful problem, particularly for the age of LLMs. It can be regarded as a perspective to explain what good model post-training is (like).

Critical weaknesses
  • The proposed selection proceduer is mostly linear in design. This falls short for paper depth and there is not much insights we can learn from it.
  • The writing is rather sloppy with lots of BX, EX, and SX, etc., in the paper. Even if a couple of them were explained, the original phrases have nothing to do with e.g., B, but it was called B0. This is not a good enough quality of writing.
  • The paper is hard to follow in terms of its content organization, so it is also hard to understand its value. For example,the main text never separates the decision rule from its implementation details.
  • The experiments use a limited number of data points which make us concerned about the trustability. Even worse, no description of datasets are provided at all.

Minor:

  • This is a paper related to CF; however, anything else related to CF from an XAI perspective is hardly discussed. For example, proximity, sparsity, plausibility, etc.
Questions for authors
  • "{p1,…,pM}\{ p_1,\dots,p_M \}" be the protected set. What defines a protected set? Later the paper has "protected set defines the operational meaning of collateral damage rather than an exhaustive notion of perservation". Again, it is still too vague and appears too late.
  • "\texttt{FORCED_ABSTAIN}"---this is not a good writing practise.
  • From "In E1 we enumerate..." to "..and final testing are disjoint", could you explain what information it wants to deliver?
  • Wrt the admissibility rule, were alternatives such as a quantile or distributional risk constraint considered?
  • The locked test uses 90 edits per model and a single calibration split. How trustable are the reported harm-rate reductions across seeds and calibration draws?
  • The paper never defines what good plasticity is. The first paragraph of section 3 is mostly for framework construction. The most close one is "This creates a basic plasticity problem: how should a model be updated when several plausible updates differ in both benefit and collateral effects?", but it is not a formal definition.
  • It gives me a question: if the model updates as a whole have more good than harm, does it mean its plasticity is good enough? If that is good enough, why do we need this framework?
Reviewer_tB96OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work introduces Counterfactual Plasticity, a method that selects the best candidate weight update without applying the updates to the model. To achieve this, the authors introduce acquisition gain and collateral damage metrics, which compare the losses of the original model with the losses of the model after each candidate update. Among the candidates whose predicted damage is below a threshold, the method selects the one with the highest predicted gain. The method is evaluated on factual editing with TinyLlama, Gemma, and Qwen2.5-7B.

Critical strengths
  • The method substantially reduces the rate of harmful updates compared to the uncalibrated baseline in the main factual-editing experiments (from 24.4% to 7.8% on TinyLlama and from 23.3% to 8.9% on Gemma).
  • The finding that acquisition gain is easy to predict while collateral damage is much harder is interesting and potentially useful
Critical weaknesses
  • described problem is not grounded in the existing literature, references are very limited, that is why it’s hard to make it grounded in the literature
  • The claimed contribution of formulating counterfactual update selection is rather minor. The trade-off between learning new information and preserving existing knowledge is well known. The paper does not discuss how its formulation relates to these existing lines of work.
  • obtaining pre-execution fingeprints is not clearly described while it appears to be the crucial part of the work
  • the paper is difficult to follow, most of the notation is not clearly described and authors use many shortcuts which makes the method difficult to understand
  • in Table 1 “Protocol and calibration” overlaps with the table
  • The method is compared only to baselines built from the authors’ own pipeline. There are no comparisons to simple alternatives. Existing editing methods (e.g., ROME, MEMIT) are also not included. This makes it hard to judge whether the added complexity of calibrated selection is necessary
  • μ_G meaning is not stated anywhere. The paper only says it is “the acquisition forecast used for ranking.” It’s never stated whether it’s an ensemble mean or a single predictor’s output.
Questions for authors
  • What versions of Gemma, TinyLlama, and GPT-2 were used in the experiments?
  • Could the authors provide more details on the forecasters (architecture, training hyperparameters, ensemble size) to improve the reproducibility of the method?
  • How does the method compare to other alternatives? A comparison with ROME or MEMIT may show whether the added complexity of calibrated selection gives a real advantage over existing editing methods.
  • How exactly are the pre-execution fingerprints computed?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.