acceptodds
Under review as a conference paper at ICLR 2027

Bidirectional Causal Tracing and Intervention: Causal Mechanism-Guided Machine Unlearning

Abstract

Large language model unlearning is commonly evaluated through changes in answer probabilities or output behavior, but such behavioral suppression does not demonstrate that the causal support underlying target knowledge recall has been weakened. Reliable unlearning therefore raises two coupled questions: where is this causal support located, and how should it be weakened? To address the first, we use causal tracing under forward and reverse queries to localize internal components whose restorative effects persist across both retrieval directions. To address the second, we optimize localized adapters to directly reduce these components' causal restoration effects, while directional balancing and retention constraints limit collateral damage to non-target knowledge. We call this framework Bidirectional Causal Tracing and Intervention (BCTI). On 50 target facts with Pythia-6.9B-deduped, BCTI achieves an average 87.8% reduction in target answer scores, while a held-out causal audit records a 98.6% ± 1.5% reduction in causal restoration effect. Applying the same procedure to OLMo-7B yields a 97.8% ± 1.9% causal restoration reduction. These results support weakening and auditing causal recall support as a more reliable basis for assessing forgetting than behavioral suppression alone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine Unlearning, Bidirectional Causal Tracing, Large Language Models, Factual Knowledge Unlearning, Mechanistic Interpretability

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

The paper proposes BCTI, which localizes components causally supporting factual recall via bidirectional causal tracing, then directly minimizes their restoration scores with localized LoRA adapters under retention constraints.

Critical strengths
  1. Distinguishing "the answer is suppressed" from "the mechanism that can restore the answer is weakened" is conceptually important.
  2. The overall method and framework design are well motivated.
Critical weaknesses
  1. In Table 2, the answer-only objective achieves a 98.6% causal-restoration reduction, identical to full BCTI. This suggests that a plain answer-level objective may by itself be sufficient to disrupt the audited causal pathway.
  2. Table 1 lacks the causal-restoration reduction metric for the baselines.
  3. Redundant presentation: Figure 4 largely restates Table 1, and Figure 5 largely restates Table 2.
  4. Several basic definitions are missing. The clean/corrupted/patched runs and the corruption procedure itself are never defined, and some metrics are used before being introduced, which forces the reader to consult prior literature. A more intuitive explanation of why minimizing L_causal leads to forgetting would also help readability. The Table 2 column header "Causal restoration" would be clearer as "Causal-restoration reduction". As written it is easy to misreadr.
Questions for authors

See weaknesses

3 comments

Thank you for recognizing the conceptual importance of distinguishing answer suppression from weakening the mechanisms that can restore the answer, and for your positive assessment of the motivation behind our overall method and framework design. Your detailed observations have helped us sharpen the controlled comparisons, broaden the causal audit of baseline methods, and improve the clarity of the presentation.

1. Identical causal-restoration reduction for answer-only and full BCTI

You correctly observed that the original Table 2 reports the same mean causal-restoration reduction (98.6%) for full BCTI and answer-only. This observation motivated a controlled comparison at similar retention levels. We trained answer-only at the same BCTI-selected update locations and stopped when validation retain KL first reached BCTI's target. At this point, maximum retain-answer drop is 11.7 ± 6.2% for answer-only and 12.5 ± 10.7% for BCTI; target-answer reduction is 64.1 ± 18.4% versus 87.8 ± 7.8%, and causal-restoration reduction is 84.9 ± 16.8% versus 98.6 ± 1.5%. This shared validation retain-KL criterion highlights the contribution of the causal objective. In the causal-only ablation, target-answer scores on held-out prompts decrease by 83.4 ± 9.1% and causal restoration decreases by 97.9 ± 1.9%, demonstrating that the restoration effect provides an effective forgetting signal. The direct- and benign-relearning tests further show slower recovery and lower answer scores for BCTI, with particularly clear differences in causal scores and under benign relearning.

2. Missing causal-restoration reduction for baselines in Table 1

Thank you for suggesting the additional causal-restoration evaluations for baselines. We conducted post-hoc audits of the existing GA+KL and NPO-LoRA checkpoints using their original training results and checkpoint selections. Their causal-restoration reductions are 100.00 ± 0.01% and 69.3 ± 34.1%, respectively, compared with 98.6 ± 1.5% for BCTI. Retention provides essential context: GA+KL has retain KL of 0.655 ± 0.395 and maximum retain-answer drop of 77.0 ± 17.2%; BCTI achieves 0.006 ± 0.004 and 12.5 ± 10.7%, respectively. NPO-LoRA and BCTI each update approximately 2.10M parameters. The expanded table therefore enables a joint comparison of causal-restoration reduction and retention outcomes.

The expanded baseline table will be included in the updated PDF.

3. Redundant presentation in Figures 4 and 5

Thank you for identifying this redundancy. Your suggestion allows us to devote more space to new recovery evidence. We will remove Figures 4 and 5, which overlap with the tabulated comparisons, and use the space for the direct- and benign-relearning trajectories, reporting both answer and causal scores.

4. Definitions, corruption procedure, intuition, and metric naming

Thank you for these precise suggestions. Making the intervention easier to follow should also make the connection between the causal loss and forgetting more accessible. We will introduce the clean, corrupted, and patched runs before the metric is used and describe the corruption operation explicitly. For intuition, in “The capital of France is …”, the clean run uses the original prompt; the corrupted run perturbs the tokens identifying France; and the patched run restores a selected clean activation within the corrupted run. An increase in the score of “Paris” after patching measures the restorative contribution of that state under this intervention. Our causal loss penalizes this patching-induced increase, connecting the operational metric directly to the optimization objective. Finally, following your naming suggestion, we will rename the Table 2 column “Causal-restoration reduction” to specify that the column reports the decrease in the restoration effect.

We hope these controlled and extended evaluations, together with the clearer presentation, address the main issues you raised. Thank you again for helping us strengthen both the evidence and its interpretation. We welcome and look forward to further discussion of any remaining questions. The above discussions and corresponding updates will be reflected in the updated PDF, which we are currently preparing.

The comparison confines answer-only to conditions designed around BCTI, which may prevent it from discovering alternative strategies. Its poorer performance is therefore understandable, especially against explicitly guided BCTI. This supports BCTI’s interpretability and targeted effectiveness in this setting, but not the intrinsic superiority of its causal-mechanistic guidance. In addition, Table 1 still lacks complete results and requires further refinement. I would like to keep my score.

Thank you for your follow-up and for recognizing BCTI’s interpretability and targeted effectiveness. We would like to clarify the purpose of the answer-only ablation and the Table 1 updates already reported.

  1. Your concern about whether the “conditions designed around BCTI” explain this advantage is important; our comprehensive ablations are designed precisely to address it. Answer-only receives the same guidance on update locations, isolating the causal loss’s added benefit and following the earlier request for “similar retention levels” and “the same update locations.” Separately, replacing BCTI-selected update locations with ROME-selected update locations reduces answer reduction to 0.6%. Thus, both components are tested separately, and the shared setup controls for update locations rather than giving BCTI an exclusive advantage.
  2. Regarding Table 1, our previous response already reported supplementary causal-restoration results for GA+KL and NPO-LoRA, alongside their retention metrics. We also explicitly acknowledged that the remaining baseline evaluations were ongoing and would be included in the final version.

We hope this clarifies why the controlled ablation directly addresses the causal objective’s contribution and how we are completing Table 1. Thank you again for helping us strengthen the evidence and its presentation. We welcome further discussion of any remaining concerns.

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

This paper proposes BCTI, a factual unlearning method guided by causal tracing. It uses activation patching to locate layers that support a fact in forward and reverse queries. Low-rank adapters then reduce answer restoration and suppress target answers, subject to retention constraints. Experiments assess forgetting, restoration, and retention across language models. Separate prompts are used for localization, model selection, and evaluation.

Critical strengths
  1. Examining internal recall pathways alongside answer suppression gives a broader view of unlearning.
  2. Testing both query directions can reveal failures hidden by an average forgetting score.
  3. Separating localization, model selection, and evaluation prompts strengthens the experimental design. Retention outliers also reveal damage hidden by average scores.
Critical weaknesses
  1. The benefit of the causal loss over answer-only training remains unclear because the two methods cause different amounts of damage to retained knowledge. The answer-only variant achieves the same restoration reduction and stronger answer suppression. Full BCTI preserves other knowledge better, so neither method clearly dominates. The comparison needs similar retention levels, the same update locations, and the same model-selection budget.
  2. The paper does not show that its causal audit predicts knowledge recovery better than ordinary answer-based evaluation. Training and evaluation use the same type of activation-patching test. Held-out prompts check new wording, but do not test recovery through other procedures. An independent recovery test is needed, such as checking whether the knowledge returns after further training.
  3. The reported restoration reduction is difficult to assess because the evaluation does not clearly specify how invalid cases and reference scores are handled. When corruption barely changes the answer score, the formula assigns zero, but the evaluation describes excluding the case. If unlearning changes which cases are included, the measured reduction may also change. It is also unclear whether the reference scores come from the original or updated model. A paired comparison needs raw scores for the same test cases and the proportion of invalid cases.
  4. The localized intervention may affect closely related facts, but the retention evaluation does not separately test this risk. A module that helps recover the target answer may also support other facts about the same entity. The paper reports a disjoint retain set, but does not separately measure damage to these related facts. A matched test on related and unrelated facts would clarify whether BCTI weakens the target association selectively.
Questions for authors

Please refer to the Weaknesses section.

5 comments

Thank you for recognizing the motivation for studying internal recall pathways, the bidirectional evaluation design, and the separation of localization and evaluation prompts. Your suggestions are particularly helpful in clarifying two closely related questions: whether the causal objective provides a benefit under comparable retention constraints, and whether that benefit persists when the model is trained further.

1. Benefit of the causal loss: similar retention, same update locations, and model-selection budget

You are right to highlight this issue. Unlearning requires balancing multiple forgetting and retention criteria, which makes comparisons across different operating points difficult; we have encountered this challenge in our own experiments. To address your well-founded concern, we added an answer-only control using exactly the same BCTI-selected adaptation locations and stopped training when validation retain KL first reached the BCTI target.

VariantTarget-answer reduction (%) ↑Min-dir. forgetting (%) ↑Causal-restoration reduction (%) ↑Maximum retain-answer drop (%) ↓
Answer-only (retain-matched stopping)64.1 ± 18.456.5 ± 21.584.9 ± 16.811.7 ± 6.2
Causal-only83.4 ± 9.178.1 ± 11.697.9 ± 1.911.7 ± 9.9
Full BCTI87.8 ± 7.883.3 ± 10.498.6 ± 1.512.5 ± 10.7

The stopping criterion matches the validation retain-KL target rather than every retention metric exactly.

The accompanying ablation table will be included in the updated PDF.

2. Independent recovery test and the added information from the causal audit

Thank you for bringing up recovery. This is a particularly useful extension of the comparison above: similar-retention evaluation tests forgetting at the selected checkpoint, whereas an independent recovery test examines whether that effect persists after further training. Following your suggestion, we evaluated BCTI, answer-only, and NPO-LoRA on the same 50 target facts under direct relearning on the forgotten targets and benign relearning on five other attributes of each target entity. We tracked both answer scores and causal restoration scores throughout training. BCTI generally maintains lower answer scores during direct relearning, while its causal restoration scores remain consistently lower. Under benign relearning, the separation is more pronounced in both recovery trajectories. Thus, the similar-retention comparison and the independent recovery tests provide complementary evidence about the value and persistence of the causal objective. We hope this extension, motivated by your suggestion, makes the recovery behavior more informative to assess.

We list several key measurements below; the complete recovery curves will be included in the updated PDF.

Relearning settingMethodAnswer score beforeAnswer score afterCausal score beforeCausal score after
DirectBCTI0.07610.58000.000480.01845
DirectAnswer-only0.21170.64240.006940.03968
DirectNPO-LoRA0.22560.69680.014260.05888
BenignBCTI0.07610.13950.000480.00150
BenignAnswer-only0.21170.39340.006940.02576
BenignNPO-LoRA0.22560.39550.014260.02344

3. Invalid cases, reference scores, and paired comparisons

Thank you for carefully checking the “exclude” wording. The gap threshold is a numerical-stability safeguard: cases with an original-model clean–corrupted reference gap at or below the threshold receive a zero score and remain in the evaluation. Pre- and post-unlearning evaluations use the same original-model reference for each prompt and patching location. Full BCTI and answer-only each have 8,460 paired cases, and all original-model reference gaps exceed the threshold. For the updated BCTI model, 4.26% of gaps trigger the threshold; all such cases remain in Table 2. As an additional check, the mean patched-minus-corrupted probability increase decreases from 0.10264 to 0.00177 for BCTI and from 0.10264 to 0.00168 for answer-only. We will clarify the threshold handling and fixed reference in the paper.

4. Related versus unrelated facts

We share your interest in the retention of facts related to a forget target. Our retention construction prioritizes facts that share the target relation and exhibit causal overlap at selected intervention locations. For instance, when unlearning one capital fact, the constraints protect other capital facts with restorative effects at the same sites. In held-out retention evaluation, BCTI achieves retain KL of 0.006 ± 0.004 and maximum retain-answer drop of 12.5 ± 10.7%; the data construction is also available with the code. Your proposed comparison of other attributes of the same entity against matched unrelated facts offers a more fine-grained view of relatedness. Our current evidence quantifies retention using causal overlap and held-out aggregate metrics. We welcome further discussion of a stratified retention-damage analysis comparing these two groups.

We hope the retention-controlled comparison and independent relearning trajectories address the central concerns about the value and robustness of the causal objective. Thank you again for these constructive suggestions; we would be glad to discuss any further checks that you consider most decisive. The above discussions and corresponding updates will be reflected in the updated PDF, which we are currently preparing.

Thank you for the rebuttal and additional experiments. The retention-matched control and metric clarifications address several of my concerns. However, the relearning comparisons start from different answer scores, so lower final scores alone do not establish stronger resistance to recovery or the causal audit’s added value. Related-fact retention also remains unclear: benign relearning tests recovery of the target fact, rather than preservation of other facts about the same entity. These remaining concerns affect the central contribution, and I therefore maintain my score.

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

This paper proposes Bidirectional Causal Tracing and Intervention (BCTI) for LLM unlearning. It uses forward and reverse queries to locate components that support factual recall. It then trains localized adapters to weaken their causal restoration effects. Retention constraints help preserve other knowledge. Experiments on 50 facts across two models show substantial reductions in target-answer scores and measured causal restoration. But it It remains unclear whether the method scales or resists knowledge recovery.

Critical strengths
  • Considers forward and reverse knowledge retrieval
  • Uses held-out prompts and several useful ablations
  • Updates relatively few parameters through localized adapters
Critical weaknesses
  • Evaluation is limited
  • Reduced causal restoration does not prove complete knowledge removal
  • Robustness to relearning is not considered
Questions for authors

What is the total computational cost? How about resisting recovery method?

2 comments

Thank you for your thoughtful review and for highlighting as strengths our consideration of forward and reverse knowledge retrieval, our use of held-out prompts and useful ablations, and our localized adapters that update relatively few parameters. Your comments on evaluation breadth, the interpretation of causal restoration, and recovery after further training are particularly helpful. Below, we clarify the evidence already presented and describe the additional analyses motivated by this discussion.

1. Evaluation coverage

While we have not covered every setting, we have made a concerted effort to broaden its scope. Our original evaluation covers: (1) two model families, Pythia-6.9B-deduped and OLMo-7B; (2) three factual relations, capital, location, and currency; (3) forward and reverse retrieval using held-out prompts; and (4) three complementary evaluation dimensions, target-answer reduction, causal-restoration reduction, and retention, measured by retain KL and maximum retain-answer drop.

In this rebuttal, we additionally provide: (1) post-hoc causal audits of available baseline checkpoints; (2) a retention-controlled answer-only ablation using the same BCTI-selected adaptation locations; and (3) direct and benign relearning tests.

Across these evaluations, BCTI exhibits strong forgetting with limited retention damage: causal-restoration reduction ranges from 92.04% to 99.95% across Pythia targets and averages 97.8% on OLMo-7B. We hope all the above experiments address your concern about evaluation coverage, and we would be glad to consider further experiments after hearing your additional suggestions.

2. Complete knowledge removal

We partially agree with your concern: reduced causal restoration alone cannot prove complete knowledge removal. Your question points to an interesting broader challenge: how can we test whether apparently forgotten knowledge remains recoverable through other procedures?

Experiments offer one way to investigate this question. As shown in our experiments, BCTI suppresses target-answer scores on ordinary held-out prompts and reduces the ability of the identified internal states to restore those answers under the specified activation-patching audit. The causal score quantifies restoration through the audited recall pathways under the specified interventions. The independent relearning tests below provide additional evidence beyond this audit.

More extensive evaluations may require larger datasets. To our knowledge, a widely accepted comprehensive benchmark for certifying complete knowledge removal is still lacking, and developing one is an important direction for future research.

3. Robustness to relearning and resistance to recovery

Thanks for the question. Our results show that BCTI is robust against the tested relearning procedures. Following your suggestion, we further evaluated the same 50 target facts under (i) direct relearning on the forgotten facts and (ii) benign relearning on five other attributes of each target entity. We compared BCTI with answer-only and NPO-LoRA using answer scores and causal restoration scores across training steps. Under direct relearning, BCTI generally exhibits slower recovery and lower answer scores, with consistently lower causal scores during relearning. The advantage is more pronounced under benign relearning, where recovery in both measures remains substantially lower. These results support BCTI's stronger resistance to recovery under the evaluated settings.

We list several key measurements below; the complete recovery curves will be included in the updated PDF.

Relearning settingMethodAnswer score beforeAnswer score afterCausal score beforeCausal score after
DirectBCTI0.07610.58000.000480.01845
DirectAnswer-only0.21170.64240.006940.03968
DirectNPO-LoRA0.22560.69680.014260.05888
BenignBCTI0.07610.13950.000480.00150
BenignAnswer-only0.21170.39340.006940.02576
BenignNPO-LoRA0.22560.39550.014260.02344

4. Total computational cost

The general purpose of LLM unlearning is to provide a lightweight alternative to retraining, with forgetting effectiveness and knowledge retention remaining central objectives. As you noted among the strengths, localized adapters allow us to update relatively few parameters, reducing the parameter-update workload and offering a training-efficiency advantage over full-parameter updates. To address your concern about BCTI's computational cost, we clarify that its principal additional computation is causal localization.

With N prompts, C corruption strategies, L layers, and S intervention sites per layer, localization requires NC(2 + LS) forward passes. In the five-fact profiling experiment on one RTX 4090 (24 GB), pure localization takes an average of 9.15 minutes per fact, accounting for 48.39% of BCTI's total measured runtime; the remaining stages account for 51.61%. BCTI takes approximately three times as long as NPO-LoRA. BCTI subsequently updates approximately 2.10M parameters, about 0.0306% of the full-parameter baselines' updated parameters. The parameter efficiency therefore reduces update size, while the profiling makes explicit the additional runtime cost of causal localization and the rest of the BCTI pipeline.

We will include the training-dynamics and computational-cost figures in the updated PDF.

We hope the additional baseline audits, retention-controlled ablation, and two relearning protocols help address your concerns and make the contribution and its current limits clearer. We would welcome further discussion of any specific evaluation setting that you consider most informative. The above discussions and corresponding updates will be reflected in the updated PDF, which we are currently preparing.

Comments on the paper

1 comment

Dear AC and Reviewers,

We sincerely thank you for your time, constructive comments, and valuable suggestions. A few days ago, we posted responses to each of your concerns. Now we uploaded the revised PDF here and highlighted the changes for ease of reference.The main revisions are summarized below:

  1. Causal-restoration reduction for the baselines. To address the missing causal-restoration reduction metric for the baselines in Table 1, we added preliminary baseline results that already indicate the overall trend. We are continuing to run the causal-restoration evaluations for the remaining baselines. Although results for all baselines may not be ready within the rebuttal period, we expect to include them in the final version.
  2. Benefit of the causal loss over answer-only training. We added a comparison at similar retention levels and the same update locations: the answer-only variant uses the BCTI-selected locations and stops when validation retain KL first reaches the BCTI target. This comparison helps clarify the benefit of the causal loss while accounting for damage to retained knowledge.
  3. Robustness to relearning and an independent recovery test. To examine whether knowledge returns after further training, we added direct relearning on forgotten facts and benign relearning on other attributes of the same entities. We track answer scores and causal-restoration scores for BCTI, answer-only, and NPO-LoRA to assess their resistance to recovery.
  4. Invalid cases, reference scores, and complete knowledge removal. We clarified how invalid cases and reference scores are handled, including the fixed original-model reference and the inclusion of zero-scored cases, and provided paired comparisons on the same test cases. We also clarified that reduced causal restoration alone does not prove complete knowledge removal.
  5. Basic definitions, redundant presentation, and total computational cost. We introduced the clean/corrupted/patched runs and the corruption procedure before using the metrics, added an intuitive explanation of why minimizing the causal loss leads to forgetting, and renamed the Table 2 column header from “Causal restoration” to “Causal-restoration reduction”. We streamlined the redundant presentation in Figures 4 and 5 to accommodate recovery trajectories and added training-dynamics and total computational-cost analyses.

We hope these revisions and the detailed responses below address your concerns and make the contribution and its current limitations clearer. We look forward to continuing the discussion, answering your questions, and further updating the manuscript.

Sincerely, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.