acceptodds
Under review as a conference paper at ICLR 2027

Bridging the Representation–Behavior Gap in Knowledge Editing with Calibrated Activation Steering

Abstract

Knowledge editing aims to update specific knowledge in large language models while preserving unrelated behavior. Recent distribution-based approaches edit knowledge by steering hidden representations toward distributions associated with the target knowledge. However, it remains unclear whether better representation alignment necessarily leads to the desired output behavior. Through systematic layer-wise analysis, we identify a **representation–behavior gap** with two distinct failure modes: **behavioral under-realization**, where well-aligned hidden representations fail to sufficiently express the target knowledge in output behavior, and **behavioral overshooting**, where stronger interventions continue to improve edit success but push the output behavior beyond that induced by explicit conditioning on the target knowledge. To address these failure modes, we propose **B**ehavior-aware **E**diting via **C**alibrated **A**ctivation**S**teering (**BECAS**). BECAS uses the model's output distribution when explicitly conditioned on the target knowledge as a behavioral reference and adjusts the strength of activation steering accordingly. Experiments on knowledge editing benchmarks show that BECAS improves generalization and locality while maintaining high editing efficacy. These results demonstrate that output behavior provides a useful reference for controlling representation-based knowledge editing.

Primary area:
interpretability and explainable AI
Keywords:
Knowledge Editing, Activation Steer

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dHUDOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This article studies knowledge editing methods based on activation distribution (such as SAKE) and finds that there is a gap between representation space alignment and output behavior: the layer with the best representation alignment cannot make the model output reflect the target knowledge (behavior under implementation), while increasing intervention intensity can improve the success rate of editing, but it will cause the output to deviate from the output distribution of the model when "new knowledge is explicitly given in context" (behavior overshoot) and damage locality. The author proposes BECAS based on this: using the transmission direction of SAKE, taking the output distribution under multiple contextual conditions as a reliable behavior reference, and jointly searching for intervention layers and strengths for each editor, so as to align the target comprehensive behavior, achieve sufficient editing effect, and maintain locality. In the CounterFact and zsRE experiments on GPT2-XL, BECAS improved generalization and locality while maintaining an editing success rate of approximately 93%, and significantly reduced behavioral differences.

Critical strengths
  • The diagnostic analysis layer by layer and intensity by intensity clearly demonstrates the two phenomena of "behavior under implementation" and "behavior overshoot", providing an intuitive and useful perspective for the selection of intervention intensity in activated knowledge editing.
  • The method is simple and easy to implement, without the need to train new transmission maps, and has significant improvements in locality and behavior alignment compared to SAKE.
Critical weaknesses
  • The experiment was only conducted on an older small model, GPT2-XL, and the candidate layer was also selected for it. It was not validated on modern models such as Llama and Qwen, nor did it consider more practical scenarios such as batch or sequence editing. Therefore, there is a great doubt about whether the conclusion can be generalized, which is the main reason why I tend to reject it.
  • There is a certain degree of cyclicality in the evaluation. The main advantages of BECAS are reflected in two indicators: behavioral JS and overshoot, which are its direct optimization goals; At the same time, each editor uses neighborhood queries for calibration, which are close to the distribution of queries evaluating locality. Therefore, the significant improvement in these indicators cannot fully demonstrate that the editing quality has truly improved.
  • The baseline comparison is not complete enough. The paper explicitly lists ICE as a comparative method in the experimental setup, but its results are not presented in the results table; The behavioral reference used by BECAS comes from the output distribution under contextual conditions. Without this comparison, it is difficult to determine whether the benefits come from activating the intervention itself or borrowing the behavioral reference. In addition, recent methods such as AlphaEdit mentioned in related work have not been included in the comparison.
Questions for authors
  • How effective is BECAS on larger and newer models such as Llama-8B or Qwen series? Does the selection of candidate layers require re diagnosis for each model?
  • Why is there no ICE in the results table? It uses the same behavioral reference as BECAS, and this comparison is crucial for understanding the sources of revenue.
  • How are neighborhood queries used for calibration and evaluation of locality divided? Are they from the same neighborhood set?
  • What is the additional computational cost for each editor to perform layer and intensity searches?
3 comments
AuthorsOpenReview ↗

Response to Reviewer dHUD

Thank you for your thoughtful review and constructive feedback.

W1: Cross-model and more practical editing scenarios

Cross model: Refer to Q1. Appendix C.2 of the submitted paper reports single-fact editing results on LLaMA-3-8B-Instruct.

More practical editing scenarios: Batch and sequential editing are both important practical settings. We are currently evaluating BECAS in multi-edit settings(ongoing) and will report the latest results before the discussion deadline if time permits.


W2: behavioral JS and overshoot improvement cannot fully demonstrate that the editing quality has truly improved.

Behavioral JS and overshoot alone are insufficient to establish editing quality, given their relationship to the calibration objectives. The "Gen." and "Loc." reported in Table 1 (can also refer to Q2) of the main paper are evaluation results on held-out paraphrase and neighborhood queries, respectively. These queries were excluded from calibration and from the per-edit selection of intervention layer and strength. Across both datasets, BECAS improves generalization to unseen paraphrases and locality on unseen neighborhood queries relative to SAKE and ICE. These gains on held-out task metrics provide evidence of editing effectiveness beyond the behavioral alignment metrics.

As described in Appendix A.3, calibration uses the original rewrite query, one sampled paraphrase, and one sampled neighborhood query per edit. The remaining paraphrase and neighborhood queries are reserved for evaluation. The calibration and evaluation neighborhood queries come from the same edit-specific pool, so the locality results measure preservation on unseen queries within that neighborhood. We will state this split alongside Table 1 and refer to Appendix A.3.


W3: More baseline and where benefits come from

More baseline: Refer to Q2. We now report ICE on the same CounterFact/ZsRE edits, using identical held-out evaluation prompts and metric definitions.

Benefits come from activating the intervention itself or borrowing the behavioral reference. ICE and BECAS both use model outputs conditioned on the new fact, but they use those outputs differently. ICE uses a context-conditioned teacher for a weight update based on the rewrite prompt. BECAS uses a reference built from multiple contextual queries to select an activation intervention. The comparison therefore reflects differences in both the intervention mechanism and the information available during editing.

Table 3 in main paper includes a direct ablation, "BECAS w/o LbehL_{beh}". Following the evaluation protocol in Appendix A.3, both variants are evaluated on held-out queries for the same edits. The comparison isolates the effect of including the behavioral reference in calibration while retaining the other calibration objectives. The results demonstrate that, within the BECAS framework, the behavioral reference aligment contributes to the performance(Gen, Loc) of BECAS.


AuthorsOpenReview ↗

Response to Reviewer dHUD

Q1: Cross-model

Appendix C.2 of the submitted paper reports single-fact editing results on LLaMA-3-8B-Instruct. On this model, BECAS improves both "Gen." and "Loc." relative to SAKE. We will move this comparison into the main text and specify the candidate layers used for LLaMA-3-8B-Instruct. Performance on the Qwen family remains to be evaluated.

When transferring BECAS to a new model, we recommend running a layer-wise diagnostic once to identify its candidate layers, since layer indices and responses to activation interventions can differ across architectures. The resulting candidate set is reused across edits on that model. For each edit, BECAS selects an intervention layer from that set and tunes the intervention strength.


Q2: More baseline

We thank the reviewer for pointing out this omission: while we included ICE in our experimental setup, we failed to report the corresponding results. We now report ICE alongside SAKE and BECAS on the same CounterFact/ZsRE edits on GPT2-XL, using identical held-out evaluation prompts and metric definitions.

Method on CounterFactEFF.↑Gen.↑Loc.↑B.JS↓OS↓
ICE99.8771.6256.9026.2115.37
SAKE96.2087.5484.6727.7943.46
BECAS93.3793.1390.8415.927.67
Method on ZsREEFF.↑Gen.↑Loc.↑B.JS↓OS↓
ICE98.9078.4161.8726.7118.61
SAKE95.4786.2787.3128.0842.13
BECAS92.7292.0190.3210.8211.24

ICE and BECAS both use model outputs conditioned on the new fact, but they use those outputs differently. ICE uses a context-conditioned teacher for a weight update based on the rewrite prompt. BECAS uses a reference built from multiple contextual queries to select an activation intervention. The comparison therefore reflects differences in both the intervention mechanism and the information available during editing.

We agree that AlphaEdit would broaden the comparison with recent editing methods. We do not yet have AlphaEdit results under the matched evaluation protocol, so the revised table reports the directly comparable ICE results and the behavioral-loss ablation. A matched comparison with AlphaEdit remains open.


Q3: Neighborhood queries for calibration and evaluation

Yes. As described in Appendix A.3, for each edit we sample one query from its CounterFact neighborhood pool for locality calibration and reserve the remaining neighborhood queries for evaluation. In the revised 1,000-edit run, we evaluate four remaining neighborhood queries per edit and verify that none overlaps with the calibration query after text normalization.


Q4: Additional computational cost

We benchmarked 100 paired single-fact edits on GPT2-XL using one NVIDIA RTX A6000, FP32, and six CPU threads. Each method ran independently. Timings exclude model loading, warm-up, and final quality evaluation.

MethodMean time per edit (s)P95(s)Peak GPU memory (GiB)
SAKE @L42,α=1\alpha=14.7145.4765.98 GiB
L=42, continuous α\alpha5.6126.0135.98 GiB
ICE4.1884.2977.05 GiB
BECAS11.09811.6735.99 GiB

Compared with fixed-layer SAKE, BECAS adds 6.384 seconds per edit (95% paired confidence interval [6.243, 6.530]). BECAS performs the search once per edit and reuses the selected intervention for subsequent queries. Another time-consuming aspect is that BECAS also needs to construct transport mappings for the two candidate layers separately.


Thank you again.

Thank you for the detailed rebuttal and the additional experiments. In light of this, I am raising my score from 1 to 2.

Reviewer_UG1YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Behavior-aware Editing via Calibrated Activation Steering (BECAS) to address the mismatch between hidden-representation alignment and final output behavior in representation-based knowledge editing. The paper refers to this mismatch as the representation–behavior gap and divides it into two cases. The first is behavioral under-realization, in which the hidden representations are already well aligned with the target distribution, but the target knowledge is not sufficiently reflected in the model’s outputs. The second is behavioral overshooting, in which stronger interventions continue to improve editing success but cause the model’s output behavior to go beyond the behavioral range observed when the target knowledge is explicitly provided. BECAS reuses the representation transport direction introduced by SAKE and treats the output distribution produced when the model is explicitly given the target knowledge as a behavioral reference. It then selects an appropriate intervention layer and intervention strength for each edit. The method is calibrated against three objectives: behavioral alignment, sufficient editing efficacy, and locality preservation.

Critical strengths

1.The motivation is intuitive, and the paper is generally clearly presented. The paper points out that aligning a model with the target distribution in representation space does not necessarily guarantee appropriate output behavior. Analyzing knowledge editing in both representation space and output space offers a reasonable and potentially valuable perspective.

  1. Using the knowledge-conditioned output distribution as a behavioral reference is well motivated. Compared with merely increasing the probability of the target answer, the output distribution produced when the target knowledge is explicitly provided contains richer behavioral information. The paper also reliability-weights the behavioral reference using multiple context templates, context agreement, and target support, resulting in a relatively comprehensive method design.
Critical weaknesses

1.The paper does not sufficiently explain what fundamentally new scientific problem the representation–behavior gap constitutes relative to previously identified issues in knowledge editing. Behavioral under-realization is closely related to existing concepts such as superficial editing, under-generalization, insufficient portability, and incomplete knowledge internalization. Behavioral overshooting, meanwhile, closely resembles the long-standing problems of over-editing and the efficacy–locality trade-off. It therefore remains unclear whether the paper identifies a genuinely new failure mode in knowledge editing or primarily reframes existing problems using new representation- and behavior-distance measures.

  1. The methodological novelty is also relatively limited. SAKE already performs knowledge editing through representation-distribution transport, while ICE already uses the context-conditioned output distribution produced when the target knowledge is explicitly provided as a richer editing signal. BECAS mainly combines SAKE’s representation transport with an ICE-style behavioral reference and conducts per-instance searches over intervention layers and strengths. It therefore appears more like a useful but incremental calibration extension than an entirely new knowledge-editing framework.

  2. The experimental setup and results supporting the paper’s central claims are insufficiently clear, and the contribution of the behavioral reference itself is not adequately isolated. First, the paper directly compares raw Gaussian W2W_2 distances across different Transformer layers. However, hidden states at different layers may have different scales and geometric structures, making it unclear whether these distances are directly comparable. Moreover, the remaining computational pathways differ between early and late layers. Consequently, the observation that different edits have different optimal intervention layers may not, by itself, be sufficient evidence for an independent representation–behavior gap.

Questions for authors
  1. Could the authors further clarify what is fundamentally novel about the proposed representation–behavior gap relative to previously studied issues such as superficial editing, insufficient generalization, and the classical efficacy–locality trade-off or over-editing problem?

  2. Could the authors explain why raw Gaussian W2W_2 distances can be directly compared across different Transformer layers? When controlling for the intervention layer and steering strength, could the authors report normalized representation discrepancies and the relative improvement in representation alignment?

  3. Could the authors explain the apparent inconsistency between the results in Tables 2 and 3? The three gap categories in Table 2 collectively contain 1,000 editing instances. However, after weighting the locality results across these groups, BECAS appears to reduce overall locality. In contrast, Table 3 reports a substantial improvement in locality on what is described as the same 1,000-instance CounterFact subset.

3 comments
AuthorsOpenReview ↗

Response to Reviewer UG1Y

Thank you for your thoughtful review and constructive feedback.

W1: Scientific contribution of the representation–behavior gap

As detailed in Q1, we agree that the phenomena we study overlap with previously identified editing failures. We explain how this diagnostic relates to superficial editing, insufficient generalization, and the efficacy–locality trade-off.

The paired analysis in Q2 report normalized representation discrepancies and the relative improvement in representation alignment while controlling for the intervention layer and the steering strength. That results support the representation–behavior gap can be observed with the intervention layer and downstream computation held fixed.


W2: Novelty

Our method, BECAS, does not introduce a new representation transport mechanism, it is a calibration of representation-based knowledge editing, inheriting SAKE’s distributional transport and builds on the use of knowledge-conditioned output distributions explored in ICE. The methodological contribution is how BECAS uses that distribution to select a steering intervention for each edit. We will clarify this scope of the methodological contribution in the revision.

As shown in Q2, the fixed-layer comparison shows why this selection criterion matters. Table 3 directly evaluates the contribution of including the behavioral loss through the "BECAS w/o LbehL_{beh}" ablation discussed in W3.


W3: Contribution of the behavioral reference and evidence for the representation–behavior gap

Contribution of the behavioral reference. Table 3 in main paper includes a direct ablation, BECAS w/o LbehL_{beh}. Following the evaluation protocol in Appendix A.3, both variants are evaluated on held-out queries for the same edits. The comparison isolates the effect of including the behavioral loss in calibration while retaining the other calibration objectives. The results demonstrate that, within the BECAS framework, the behavioral reference alignment contributes to the performance(Gen, Loc) of BECAS.

Evidence for the representation–behavior gap. Refer to Q2.

AuthorsOpenReview ↗

Response to Reviewer UG1Y

Q1. Novelty about the proposed representation–behavior gap

We agree that behavioral under-realization and behavioral overshooting overlap with previously studied editing failures. Our contribution is to measure changes in hidden representations and output behavior under the same intervention. This lets us test a specific question: When an intervention brings hidden representations closer to the target distribution, does the output distribution also move closer to the distribution observed when the new fact is explicitly provided? We refer to the latter distribution as the knowledge-conditioned behavioral reference.

Superficial editing and insufficient generalization. These concepts concern whether the model uses the new fact and whether the edit generalizes to prompts such as paraphrases. Our analysis adds a measurement of how far the hidden representations move toward the constructed target distribution under the same intervention. This joint measurement can also complement generalization analyses by tracking representation changes alongside output outcomes on the prompts being evaluated.

Over-editing and the efficacy–locality trade-off. The classical efficacy–locality trade-off concerns whether stronger edit efficacy comes at the expense of preserving behavior on non-target queries. We additionally examine what happens on in-scope prompts: can the output distribution move further away from the knowledge-conditioned behavioral reference even as the target answer becomes more likely to be generated?


Q2. Cross-layer comparability of Gaussian W2W_2 and within-layer normalization

Hidden states can differ in scale and covariance structure across Transformer layers, so raw distances alone do not justify a cross-layer ranking of representation alignment.

Normalized representation discrepancies. For each edit and layer, we measure the pre- and post-intervention (W_2) distances to the same target distribution. We report the normalized residual distance (R=W2,after/W2,beforeR = W_{2,after}/W_{2,before}) and the relative reduction in W2W_2 distance (1−R1-R). This ratio cancels a common scale factor within a layer and expresses the change relative to each edit’s starting distance. At L8, a large relative reduction in W2W_2 coexists with very low Eff, illustrating behavioral under-realization.

Fixed layerSteering strength α\alphaMedian RR ↓Median 1−R1-R ↑Eff. ↑B. JS ↓
L81.009.07%90.93%0.83%0.2922
L361.002.08%97.92%90.88%0.2212
L421.001.69%98.31%94.88%0.2292

Relative improvement in representation alignment. To test whether representation alignment and behavioral alignment can diverge within the same layer, we compare α=0.75\alpha=0.75 with α=1.00\alpha=1.00 using the same edits and in-scope prompts at both strengths. The trends in Eff and B. JS are consistent with Figure 1(b) of main paper. The new W2,afterW_2,after ↓ but B. JS ↑ column counts edits whose representations move closer to the target distribution while their output distributions move further from the knowledge-conditioned behavioral reference, providing per-edit evidence of behavioral overshooting.

Fixed layerPaired editsMedian$1-R$ ↑Eff. ↑B. JS ↓Edits with W2,afterW_2,after ↓W2,afterW_2,after ↓ but B. JS ↑ [95% CI]
L3690075.01% → 97.92%66.50% → 90.88%0.1636 → 0.2212900/900752/900 (83.56%) [80.90%, 85.75%]
L4290075.17% → 98.31%79.39% → 94.88%0.1559 → 0.2292900/900835/900 (92.78%) [90.79%, 94.20%]

These results establish that the representation–behavior gap can be observed with the intervention layer and downstream computation held fixed. Behavioral JS divergence provides information about the effect of a candidate intervention even when representation alignment and aggregate edit efficacy improve. We will use the within-layer normalization and paired results to clarify this motivation in §4.1.


Q3. Explanation: Locality results in Tables 2 and 3

Thank you for checking the weighted locality results in Table 2. The original Table 2 inadvertently reported group-level scores from a pilot run of 100 edits alongside group sizes for the full set of 1,000 edits.

The corrected results below use the same 1,000 CounterFact edits as Table 3 in main paper(fixed point means L=42L = 42 and α=1.0\alpha = 1.0). Following Q2, we use the within-layer normalized representation distance to compute the gap score (GeG_e) and reassign edits to the low-, medium-, and high-gap groups.

GapMethoodeditsGen. ↑Loc. ↑B.JS ↓Δ\DeltaB.JS ↑OS ↓
Lowfixed point33389.7882.7421.33-30.37
BECAS89.1285.9315.755.586.34
Mediumfixed point33391.2785.0726.5-41.81
BECAS95.0889.6417.269.2410.72
Highfixed point33490.8285.8932.71-54.9
BECAS94.9891.7717.7414.9716.63

Thank you again.

Given that the phenomenon studied in this paper overlaps with previously identified failures in knowledge editing and lacks clear practical significance, I will maintain my original score.

Reviewer_YxroOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper first analyzes the relationship between representation alignment and output behavior in knowledge editing, identifying two failure modes: behavioral under-realization and behavioral overshooting. Based on these observations, it proposes BECAS, which constructs a behavioral reference from multiple knowledge-conditioning contexts and uses it to calibrate the intervention layer and steering strength while balancing editing efficacy and locality.

Critical strengths
  • The related work section is fairly comprehensive and covers different approaches to knowledge editing, making it easier for readers who are less familiar with the field to follow the paper.

  • The paper uses clear notation and mathematical formulations, which make the proposed method relatively easy to understand.

  • I appreciate that the authors first conduct an analysis to identify the two failure modes, behavioral under-realization and behavioral overshooting, and then use these observations to motivate their proposed method. Using the knowledge-conditioned output distribution as a reference for calibrating intervention strength is intuitive and provides a coherent connection between the analysis and the proposed solution.

Critical weaknesses
  • The method largely builds on SAKE, combining its existing distributional representation transport with a knowledge-conditioned behavioral reference, an idea already explored in ICE. The main contribution is then calibrating the intervention layer and steering strength using this reference while considering efficacy and locality. BECAS does not introduce a new representation transport mechanism but rather calibrates an existing one. Overall, I find the methodological novelty somewhat limited.
  • In Section 4.1, the authors compare absolute representation distances across different transformer layers. However, these layers may have substantially different hidden-state distributions, variances, and geometries. Therefore, a smaller Wasserstein distance at layer 8 compared to layer 42 does not necessarily indicate better knowledge-related representation alignment. Furthermore, interventions at earlier layers must propagate through many subsequent transformer layers, so even comparable levels of representation alignment at different layers may lead to very different output behaviors. I am therefore not fully convinced that this layer-wise comparison provides sufficient evidence for the claimed representation-behavior gap.
  • I am not fully convinced that the behavior of a pretrained model conditioned on a new fact necessarily represents the desired behavior of a successfully edited model. For example, suppose the model has learned that the sky is blue, and we want to edit this knowledge to make it believe the sky is red. Simply providing the sentence "The sky is red" in the context may not fully override its existing knowledge. The conditioned model might still assign 35% probability to "blue" and 65% to "red". whereas a successfully edited model might assign 90% to "red". In this case, BECAS could penalize the latter for deviating from the behavioral reference, despite it expressing the target knowledge more strongly. Although the authors consider agreement across conditioning contexts and target support, these criteria do not necessarily guarantee that the reference represents the desired post-edit behavior. I think this assumption needs further justification, particularly since it is central to the proposed method.
Questions for authors

Please refer to the weaknesses mentioned above. I will also take the other reviewers' feedback into consideration. If my concerns are at least partially addressed and the other reviewers are generally leaning toward acceptance, I would be open to raising my score.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Yxro

Thank you for your thoughtful review and constructive feedback.

W1: Methodological novelty

We agree that BECAS does not introduce a new representation transport mechanism, it inherits SAKE’s distributional transport and builds on the use of knowledge-conditioned output distributions explored in ICE. As discussed in our Related Work, ICE uses this distribution as a self-induced editing target. We use it as a behavioral reference to diagnose and calibrate interventions in representation space.

Our analysis uses this reference to identify two failure modes: behavioral under-realization and behavioral overshooting. These observations motivate BECAS to calibrate steering strength for each edit using the behavioral reference together with efficacy and locality criteria. The methodological contribution is how BECAS uses that distribution to select a steering intervention for each edit. We have also added relevant statements in the "Related Work" section of the paper.


W2: Cross-layer Wasserstein distances and the representation–behavior gap

We agree that only the smaller absolute Wasserstein distance at L8 does not establish better knowledge-related representation alignment than at L42. Hidden states can differ in scale, variance, and geometry across layers. Interventions at different layers also propagate through different downstream computations. We clarify what the cross-layer measurements show in §4.1 and add a paired analysis at fixed intervention layers.

Cross-layer measurements. For each edit and layer, we measure the pre- and post-intervention (W_2) distances to the same target distribution. We report the normalized residual distance (R=W2,after/W2,beforeR = W_{2,after}/W_{2,before} and the relative reduction in (W_2) distance (1-R). All layers use full SAKE transport (α=1.00\alpha=1.00). The pre- and post-intervention distances are means across edits.

LayerMean W2,beforeW_{2,before} ↓Mean W2,afterW_{2,after} ↓median 1−R1-R ↑
L814.951.3690.93%
L1220.981.8091.38%
L1841.372.4194.18%
L2487.023.2196.28%
L30161.704.5697.11%
L36307.856.2797.93%
L42539.948.7198.30%

L8 has a smaller mean W2,afterW_{2,after} than L42, but its median reduction(1−R1-R) is also smaller. The relative reduction quantifies how much the intervention closes the initial distance to the target distribution within each layer. Differences in hidden-state geometry and downstream computation still limit cross-layer interpretation.

Paired analysis at fixed layers. At L36 and L42, we compare the same CounterFact edits at α=0.75\alpha=0.75 and α=1.0\alpha=1.0, holding the intervention layer and its downstream computation path fixed within each comparison. For each edit, we test whether the absolute hidden-state (W2,afterW_{2,after}) at α=1.0\alpha=1.0 is lower than at α=0.75\alpha=0.75 and whether the JS divergence(B. JS) between the edited output distribution and the knowledge-conditioned reference increases.

Fixed layerPaired editsW2,afterW_{2,after}(1.0)<W2,afterW_{2,after}(0.75)W2,afterW_{2,after} ↓ but B. JS ↑95% Wilson CI
L36900900/900752/900 (83.56%)80.90%–85.75%
L42900900/900835/900 (92.78%)90.79%–94.20%

These paired results show that, even at a fixed intervention layer, the same edit can move closer to the target distribution in representation space while moving farther from the knowledge-conditioned reference in output space. This result will be used **as direct evidence that the two forms of alignment can diverge **(i.e., representation-behavior gap) and we will present the early-layer results in §4.1 as descriptive observations.



W3: Validity of the knowledge-conditioned behavioral reference

We agree that a pretrained model’s output distribution conditioned on a new fact need not represent the desired behavior of a successfully edited model. BECAS uses this distribution as a soft calibration reference alongside efficacy and locality.

In ablation table 3 of the main paper, adding LbehL_{beh} improves Gen and Loc , with a small decrease in Eff. Behavior JS and overshoot also decrease. These aggregate results show the empirical benefit of the behavioral term for generalization and locality.

To examine the reviewer’s concern at the edit level, we counted how many of the 952 in 1,000 edits with a reference that passed both checks retained at least 10%, 20%, or 30% weighted probability on the original answer. At each cumulative threshold, we compared "full BECAS" and "BECAS w/o LbehL_{beh}" loss on the same held-out paraphrase per edit.

Original-answer supportEdits (% of 952)BECAS w/o LbehL_{beh}, correct / incorrectFull BECAS, correct / incorrectIncorrect → correctCorrect → incorrect
≥10%34 (3.57%)22 / 1230 / 491
≥20%10 (1.05%)5 / 55 / 500
≥30%2 (0.21%)0 / 21 / 110

In the correct-to-incorrect case, a reference that passed both checks retained ≥10% support for the original answer, and adding LbehL_{beh} led BECAS to select weaker steering. We will describe the limitation: a reference can pass both checks while retaining support for the original answer.

Changes in the revised manuscript. In the Abstract, Introduction, and §3.2, we will describe the conditioned distribution as a soft behavioral reference for diagnosis and calibration. In §5.1, we will explain how a reference can pass both checks while retaining substantial support for the original answer. In §5.2, we will clarify that the efficacy term is a soft penalty that becomes zero at the threshold. We will add the paired threshold analysis and failure case to the appendix, with a pointer in §6.4.


Thank you again.

I thank the authors for their rebuttal and responses. I still find the methodological novelty somewhat limited. Therefore, I will maintain my original rating of weak rejection.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.