acceptodds
Under review as a conference paper at ICLR 2027

Auditing Fact Deletion in Language Models with Separable Memory

Abstract

Language models with separable memory aim to decouple source-specific knowledge from general world knowledge, language understanding and reasoning capabilities. Unlearning then becomes relatively straightforward: simply delete the memory units associated with the targeted knowledge. We show that a naive comparison of model outputs is an unreliable measure of unlearnability, and propose a methodological audit for such systems that measures the effectiveness of memory removal. We apply the audit to Continuous-Query Limited Memory Language Models (Co-LMLM) and Natively Unlearnable Large Language Models (NULLs) on five factual-recall benchmarks. Much of the post-deletion correctness in both systems is also present when separable memory is disabled. This is expected for NULLs by design, but less so for Co-LMLM, which aims to keep facts out of its weights. In NULLs, the backbone recites up to % of audited facts without any source memory, and removing more sources does not remove more facts but damages up to % of nearby ones. For Co-LMLM, up to % of all facts that pass a naive accuracy test as unlearned are still recited by the backbone. A single entry re-inserted into Co-LMLM's memory restores -% of facts the model no longer answered after deletion. We propose the audit as a protocol that future work on separable memory can use to support deletion claims.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM unlearning, memory-augmented language models, knowledge editing

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_xCmVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the problem of auditing fact deletion in language models with separable memory. The authors propose an audit that checks why a fact is still answered or not after deletion. The audit runs each fact under three conditions: with full memory, with the deletion set removed, and with all separable memory turned off. By comparing these conditions, it separates facts kept by memory, facts known by the backbone, and facts that only look deleted because the remaining memory leads to a wrong answer. Experiments on Co-LMLM and NULLs show that most facts answered after deletion are already known by the backbone, and that in Co-LMLM, some of facts that look deleted are still answered by the backbone.

Critical strengths
  1. The paper studies an important problem of how to check whether a fact is deleted, which is ignored by current evaluation.

  2. The method is simple and clear to understand, and can be used to audit separable memory.

  3. The results are convincing and cover two systems and multiple datasets with different deletion rules.

Critical weaknesses
  1. My main concern is about the novelty. The audit is adopted from prior work as the authors mention. Also, turning off retrieval to see what the backbone knows is a common practice in the RAG literature.The new parts are mainly the metrics and deletion rules and system-specific tests. Thus, I feel the contribution is somewhat incremental. Feel free to correct me if I am wrong.

  2. As authors admitted, NULLs is designed to delete content specific to one source, but the audit uses general facts that are shared by many sources and are expected to be kept in the backbone. The results also suggest these facts are not really stored in the source memory. Thus, the audit on NULLs does not really test how well its deletion works, even though the authors mention this in the limitations.

  3. In the paper, each system is tested with only one small model, and there is no comparison with standard unlearning methods on models without separable memory. I suggest applying the same audit to larger models and to these methods, which would show whether separable memory actually deletes facts better.

Questions for authors

See weaknesses.

1 comment

We thank the reviewer for their thoughtful comments.

  1. Novelty, given prior work and RAG practice. As the reviewer notes, turning off retrieval is common in RAG evaluations. In Co-LMLM, this corresponds to our DEL-OFF condition alone and cannot reveal when memory remaining after deletion suppresses an answer the backbone still recites. Comparing DEL-ON and DEL-OFF per fact allows us to quantify this interference. Please also see our response to Reviewer MZmS (point 1) on the contributions beyond Raeesi & Roed (2026).

  2. The NULLs audit uses shared facts. Please see our response to Reviewer MZmS (point 2).

  3. One small model per system and no unlearning baselines. Please see our response to Reviewer J6Fs (point 4).

Reviewer_J6FsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a methodological audit for evaluating the effectiveness of knowledge deletion in language models with separable memory, distinguishing backbone-based fact recitation from memory-dependent responses through three evaluation conditions, and applies the protocol to Co-LMLM and NULLs across five factual-recall benchmarks to reveal limitations in existing deletion evaluations and potential collateral effects of memory removal.

Critical strengths
  1. The paper addresses an important and timely problem concerning the reliability and verifiability of knowledge deletion in language models, which has significant implications for privacy protection and machine unlearning.
  2. The proposed three-condition audit provides a relatively straightforward evaluation protocol to distinguish backbone-retained knowledge from knowledge supported by surviving memory units, exposing limitations of conventional before-and-after comparisons.
  3. The experiments reveal potentially important failure cases in Co-LMLM and NULLs, particularly the persistence of factual knowledge after memory removal and the collateral damage caused by broader deletion operations.
Critical weaknesses
  1. The proposed audit primarily consists of comparing model outputs under three memory configurations. Although useful as an evaluation protocol, the methodological contribution appears relatively incremental, and its fundamental novelty over existing ablation-based and intervention-based evaluations is insufficiently established.
  2. The manuscript does not adequately distinguish between deleting explicit memory entries, suppressing factual recall, and genuinely eliminating the underlying knowledge. These are fundamentally different objectives, and the proposed audit primarily evaluates observable recitation rather than establishing whether the targeted information has been removed from the model.
  3. The audit relies heavily on greedy decoding and answer matching to determine whether a fact is retained. However, failure to generate a fact does not imply that the corresponding knowledge has been removed, while successful recitation does not necessarily establish that the original memory representation remains intact. Consequently, the proposed evaluation may provide an incomplete characterization of unlearnability.
  4. The conclusions may be sensitive to prompt formulations, answer-matching criteria, decoding strategies, and retrieval configurations. The manuscript does not sufficiently establish the robustness of its findings under alternative evaluation settings, which limits the reliability of the reported deletion rates.
  5. The evaluation covers only two separable-memory systems, one model size per system, and a fixed set of deletion rules. The findings therefore provide limited evidence for broader claims about the effectiveness and fundamental limitations of separable-memory architectures.
  6. The authors acknowledge the absence of matched baselines for evaluating how much separable memory actually reduces backbone-based factual recitation. Without comparisons against corresponding models without separable memory, it is difficult to determine whether the observed limitations are intrinsic to these architectures or simply reflect the general tendency of language models to memorize factual knowledge.
  7. The reported degradation of nearby facts after broader deletion does not sufficiently establish the underlying mechanism of interference. More systematic analyses are needed to distinguish semantic overlap, shared memory representations, retrieval interference, and unintended effects of the deletion procedure.
  8. The proposed protocol provides evidence about observable factual recitation under selected evaluation conditions, but it does not establish that information has been irreversibly removed or cannot be recovered through alternative prompts, retrieval mechanisms, or model representations. Therefore, presenting the audit as sufficient support for deletion claims risks overstating what the proposed methodology can verify.
Questions for authors
  1. How do the authors distinguish genuine knowledge removal from the suppression or non-generation of a factual answer? Can the proposed three-condition audit establish that the targeted knowledge has actually been eliminated rather than merely becoming inaccessible under the evaluated prompts and decoding settings?
  2. Have the authors examined the sensitivity of the reported spurious-deletion rates to prompt paraphrasing, alternative decoding strategies, answer-matching criteria, and retrieval configurations? How would these factors affect the conclusions regarding the effectiveness of memory deletion?
  3. Why are only one model size per architecture and two separable-memory systems considered? Can the authors provide matched baselines without separable memory to quantify whether Co-LMLM and NULLs actually reduce backbone knowledge retention compared with conventional language models?
  4. How do the authors justify using observable recitation as evidence for memory deletion, given that latent representations may still retain the targeted information even when it cannot be directly retrieved? Can the proposed audit distinguish irreversible knowledge removal from temporary behavioral suppression?
1 comment

We thank the reviewer for their thoughtful comments.

  1. Incremental over ablation-based evaluations. A before-and-after test only shows whether each fact is still recited after deletion. By evaluating each fact under three memory conditions, the audit distinguishes recitation supported by surviving memory from recitation possible without it, and identifies surviving-memory interference, where the remaining memory suppresses an answer the backbone can still recite (Equations 3–4). A before-and-after test misses this interference and counts these facts as deleted. In Co-LMLM, this occurs for 18.8% and 19.9% of audited facts on T-REx and CounterFact, respectively (Table 1). Please see our response to Reviewer MZmS (point 1), regarding novelty concerns.

  2. Deletion, suppression and knowledge removal are conflated. We agree that these are distinct. The paper measures factual recitation, and we note that backbone recitation is a lower bound on what the backbone retains (Limitations). The representation probe further shows that answer information can be decoded from the query representation even when the backbone does not recite the answer (Appendix E.2). The audit therefore does not certify that knowledge has been removed. Its role is to identify cases where deletion appears successful under a before-and-after test, even though the backbone can still recite the fact.

  3. Sensitivity to decoding, matching and prompts. Interference is measured within each fact, with decoding and prompts held fixed, so only the available memory changes between conditions. For Co-LMLM, a second DEL-OFF implementation changes interference by at most 0.3 percentage points (Appendix C.1); NULLs is more sensitive, as the paper notes. We have not tested paraphrased prompts or sampled decoding.

  4. Matched baselines and model size. Table 1 includes a matched baseline for Co-LMLM (Std-LM-360M): the Co-LMLM backbone recites 25.6% of audited T-REx facts, compared with 46.8% for the baseline, and is below both reference models on every prompt set. The NULLs reference model is not matched, and our evaluation is limited to one checkpoint per system.

Reviewer_zd3eOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces an audit for fact deletion in language models with separable memory (specifically Co-LMLM and NULLs). It evaluates each fact under three conditions: full memory, targeted memory deletion, and memory disabled entirely. Across five entity-centric benchmarks such as T-Rex and PopQA, they found that even after deletion actions, models often retain the knowledge, even in Co-LMLM, which aims to keep facts out of its weights. The authors recommend that papers on separable memory should report to support a deletion claim.

Critical strengths

The paper studies an interesting question and provides a new audit method. The authors conduct an in-depth analysis to understand the models' behavior.

Critical weaknesses

There are several major concerns. First, this paper provides limited technical contribution. The core methodology for knowledge deletion evaluation is based on Raeesi and Roed (2026), and the only novelty seems to be applying it to co-LMLM and NULLs. The paper adds some metrics and comparisons, but overall, I don't think it provides sufficient technical contributions for ICLR. Second, the empirical study is restricted to two special LM architectures, which, in my opinion, are not widely adopted (yet). The analysis and audit methodologies are designed specifically for those models, and it is unclear whether the findings are generally applicable or offer broader insights. The paper also acknowledges that many observations are expected given the model designs, but it doesn't clearly explain what the audit reveals beyond these expected behaviors.
Lastly, the evaluation is based on five atomic factual recall questions via cloze style or short-form open-domain QA questions. Given that the initial motivations focus on real-world privacy scenarios, the authors should include datasets that reflect those scenarios. Also, the small number of audit samples may cause potential concerns about statistical significance. This is not a major concern, but I think the paper's presentation could be improved. The introduction only briefly touches on the method's high-level idea without introducing the core findings. I also found the method and results sections difficult to follow because low-level details are interleaved with the core ideas and takeaways. Introducing the key intuition and findings first, then presenting the supporting details, would make the paper easier to read and its contribution clearer.

Questions for authors
  1. What are the technical novelties beyond Raeesi and Roed (2026)?
  2. Can the authors test this approach on more instances or datasets beyond cloze style or short-form QA about Wikipedia entities?
1 comment

We thank the reviewer for their thoughtful comments.

  1. Limited technical contribution beyond Raeesi & Roed (2026). Please see our response to Reviewer MZmS (point 1).

  2. Two architectures that are not widely adopted, and expected findings. The audit is not specific to Co-LMLM or NULLs. Its three conditions apply to systems where separable memory can be selectively disabled, while the deletion rules and their implementation depend on the architecture. Co-LMLM and NULLs store memory in very different ways, and we demonstrate the audit on both. While backbone recitation is expected for NULLs, our audit also reveals less obvious behavior in Co-LMLM: up to 21.2% of facts counted as deleted by a before-and-after test are still recited by its backbone (Table 1).

  3. Atomic factual-recall benchmarks and sample size. Our current audit measures recitation using answer matching, so we use benchmarks with well-defined answers. We agree that these benchmarks do not directly reflect real-world privacy scenarios, which we have not evaluated. The main Co-LMLM findings are based on substantial cohorts with narrow confidence intervals: on T-REx (n = 7,407), the spurious-deletion rate is 20.5% with a 95% confidence interval of [19.5, 21.4], and on CounterFact (n = 3,479), it is 21.2% [19.8, 22.6] (Table 10). Small cohorts occur for NULLs on three benchmarks, and we interpret those results cautiously.

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

Comparing answers before and after memory deletion cannot reliably distinguish genuine forgetting from residual backbone knowledge or interference from surviving memory. The paper proposes an audit using full, partially deleted, and disabled memory conditions, supplemented by deletion-rule, breadth, probing, and re-insertion analyses. Experiments on factual-recall benchmarks reveal misleading deletion successes, substantial backbone retention, and system-dependent trade-offs between target removal and collateral damage.

Critical strengths
  1. The audit distinguishes suppressed output from backbone knowledge that remains available without separable memory.

  2. Alternative memory-disabled implementations, stricter source checks, and probe controls strengthen the analysis of potential evaluation artifacts.

Critical weaknesses
  1. The central three-condition audit is inherited from prior work. The additional interference measure and analyses are useful, but the contribution remains mainly an application of an existing framework to two systems.

  2. NULLs is designed to retain knowledge shared across sources, so substantial backbone recall is partly expected. Without separating source-unique from shared facts, the audit provides limited evidence about the effectiveness of source deletion.

  3. The strongest re-insertion result adds a memory entry containing the target answer. This exposes a weakness in the geometric deletion rule, but does not distinguish recovery of residual knowledge from answering newly supplied information.

Questions for authors

Can you separate source-unique facts from shared facts in the NULLs cohort and report the audit outcomes for each group?

1 comment

We thank the reviewer for their thoughtful comments.

  1. The three-condition audit is inherited from prior work. Raeesi & Roed (2026) audit a triple-based memory, where the entries supporting a fact can be enumerated. In memory whose units are free-text entries or training sources, the deletion set instead depends on the rule used to identify relevant units. We therefore study surviving-memory interference (with a spurious-deletion rate Φ of up to 21.2% in Co-LMLM, Table 1), the effects of deletion rules (Table 12) and deletion breadth (Figure 2), and single-entry re-insertion against geometric deletion (Table 5). A before-and-after comparison alone cannot reveal whether an apparent deletion is due to surviving-memory interference.

  2. Source-unique vs. shared facts in NULLs. We agree that our evaluation does not explicitly separate source-unique from shared facts, and that such a split would test source deletion more directly than our current cohort. Appendix F suggests that many audited facts are supported mainly by the shared backbone rather than by a particular source-specific sink. Even with a stricter source-verification margin of 1 nat, the backbone still recites 69.7% of audited facts on T-REx and 44.3% on CounterFact (Table 4).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.