acceptodds
Under review as a conference paper at ICLR 2027

Audience-Bound Persistent Memory: Authorization Across the Memory Lifecycle

Abstract

A personal language agent that acts for its owner across private and shared conversations can learn a fact from one audience and later place it in the context it assembles for another. We study authorization before context across the whole memory lifecycle. Each memory item carries the audience present when it was recorded, and derived items are partitioned by audience, receive the intersection of their sources’ audiences, or are suppressed. An audience is widened only by an explicit, object-specific grant, and an item enters a model attempt only when every current viewer already belongs to one of its authorized audiences; ambiguous viewers fail closed to public-only. Under explicit identity, provenance and complete-mediation assumptions, this admission is sound and policy-complete on the exact assembled context, enforced by exclusion rather than by model behavior. We realize it in two independently persisted reference architectures, a flat store and a relationship graph, and separately, with descriptive validation, in a native agent-memory runtime. In a prospectively frozen confirmation over 10,000 independently generated multi-party histories and 480,000 paired retrieval cells, all eight prespecified entitled-recall non-inferiority checks against unscoped, siloed, bounded post-filter and principal-postings retrieval pass at the −0.05 Recall@5 margin. The architectures match the policy-equivalent baselines exactly in every history, so authorization before ranking costs no recall, and they recall more entitled evidence than unscoped retrieval, with a Holm-confirmed advantage that grows with distractors. Neither produced a wrong-principal substitution, against 7 of 9,984 unscoped answers; at this base rate the frozen test cannot confirm the reduction, so the prespecified joint decision is not established in either view.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
agent memory, access control, information flow control, contextual integrity, LLM agents, retrieval-augmented generation, provenance, privacy

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_xpEqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes lifecycle authorization for a personal agent’s persistent memory. Records inherit their origin audience; derived records are partitioned, restricted by audience intersection, explicitly declassified, or suppressed. Object-specific grants remain separate, and admission is checked against a snapshot before each physical model attempt. The policy is implemented in two independently persisted reference architectures and descriptively integrated into a native agent runtime.

The conditional policy appears sound, and exact assembled contexts are an appropriate endpoint for the stated confidentiality property. The 10,000-history confirmation provides credible implementation and retrieval evidence within its controlled generator. However, the strongest results mainly exercise a familiar admission predicate, while the distinguishing lifecycle contribution receives substantially thinner validation. My recommendation is marginally below acceptance because the paper has not yet established enough additional capability beyond existing authorization and information-flow mechanisms.

Critical strengths

Precise and implementable policy semantics.The contract clearly distinguishes origin audiences, separate grants, transitive source currentness, and per-attempt snapshots. In particular, separate grants cannot be unioned to authorize a larger audience, and source grants do not automatically authorize derived records.

An appropriate structural security endpoint.Measuring forbidden items in the exact assembled context avoids relying on the model to suppress information after reading it. The paper correctly separates this property from answer-level leakage and broader contextual integrity.

Meaningful implementation validation.B-FLAT and B-GRAPH use independent schemas and query paths. Differential, mutation, and transformation tests are useful evidence that the policy is implemented across multiple access pat

Critical weaknesses
  1. The research increment over established information-flow control needs stronger demonstration

Audience containment, conservative propagation of derived-data permissions, and explicit declassification are established information-flow concepts. The paper acknowledges this foundation, but its discussion of agent IFC emphasizes control and effect flows more than the confidentiality mechanisms already present in that literature.

For example, FIDES explicitly describes permitted-reader sets, intersection for derived data, email recipients as a source of confidentiality labels, and conservative propagation through LLM histories and outputs. These overlap substantially with the paper’s basic labeling mechanisms. FIDES, Sections 4.1–4.3.

The defensible increment is therefore the persistent-memory lifecycle integration: exact grants, source currentness, durable projections, and admission for each physical attempt. This is potentially valuable engineering, but the paper needs to show more directly what that integration enables or prevents beyond a conventional audience-aware store. The conditional proofs largely establish faithful enforcement of the specified policy; they do not by themselves demonstrate a substantial new capability.

  1. The most distinctive lifecycle mechanisms receive weaker evidence than the inherited admission rule

The large confirmation establishes context exclusion, recall parity, and resistance to unauthorized distractors. Direct evidence for derived-memory handling consists mainly of casewise transformation tests, eight intersection-versus-suppression cases, and descriptive native scenarios.

Importantly, Appendix E.2 explicitly states that grant revocation, item deletion/revocation, and member removal are not evaluated. The full conformance counts reported in Section 6.1 also come from an earlier revision, while the evaluated revision has a smaller selected check set.

The native results do not fully close this gap. They span heterogeneous earlier revisions, and approved widening and complete intersection each produce five incorrect abstentions. Incomplete-provenance scenarios record non-emitting requests rather than demonstrated runtime refusals (Appendix E.6).

These observations do not refute the conditional guarantee. They limit the evidence for the paper’s central extension across the memory lifecycle. A focused replay on one frozen revision should follow derived artifacts through grants, source revocation or membership changes, persistence/restart, and the next model attempt. Inputs, source closures, resulting labels, and assembled contexts would make this contribution directly assessable.

  1. The recall study establishes a narrow operating regime rather than a general absence of recall cost

Both architectures match principal postings and bounded post-filtering exactly in every history and at every reported (k). This is positive evidence for policy conformance and the stated non-inferiority objective.

However, the post-filter’s 20-item window is never exhausted by the eight-distractor manipulation. The main superiority result therefore compares authorized retrieval with unscoped retrieval under deliberately introduced unauthorized competitors; it does not distinguish the proposed lifecycle implementation from the strong authorization-aware alternatives.

The independently authored corpora do not resolve this limitation. Their audience overlay contains no competing cross-audience evidence, so all arms—including unscoped retrieval—produce identical retrieval results (Appendix E.4).

Consequently, the unqualified statement that authorization before ranking “costs no recall” should be restricted to the evaluated lexical retrieval regime. One additional regime with a binding candidate window, greater audience overlap, or nontrivial derived-memory competition would be more informative than further repetitions of the current generator.

  1. The semantic outcome claims lack sufficiently independent validation

The primary context-exposure and retrieval measurements are deterministic and do not depend on an LLM judge. The answer-level evidence is less secure.

All 15,229 semantic-review subjects receive AI judgments through several routes, none is independently validated, and the approved consistency audit was not executed (Appendix E.5). Native correctness labels were assigned by the executing AI session. With only seven detected substitutions in the unscoped arm, modest classification error could materially affect the observed counts.

The paper appropriately does not claim that H3a/b succeeded. Nevertheless, “zero substitutions” should consistently be described as zero detected substitutions under the current labeling procedure. Blinded verification of the seven positive cases and their paired architecture answers, together with an independently reviewed negative sample, would strengthen the semantic interpretation without changing the prespecified statistical decision.

Questions for authors

What lifecycle behavior is demonstrated beyond an audience-aware store with conventional label propagation and object-specific grants? Can this be isolated against the previous single-rule implementation?

Can the final evaluated revision be replayed through complete derivation–grant–revocation–restart trajectories, including the four retained unfavorable native scenarios?

How do recall and candidate utilization behave when the post-filter window becomes binding or derived items compete with authorized raw evidence?

Can the rare substitution labels and a stratified sample of negative judgments receive blinded independent verification?

1 comment
AuthorsOpenReview ↗

Response to Reviewer xpEq

Thank you for the detailed review.

Increment and isolation. The contribution is persistent-memory lifecycle integration; §2 explicitly credits FIDES's reader sets and intersection. On the fixed panel (Table 7), the complete mechanism exposes 0/27 opportunities per reference store and 0/8 natively. Removing refresh exposes 19/27 and 6/8 respectively; retrieval-only checking exposes 5/8 natively. Derived labels retain 18/18 challenge and 16/16 ordinary operations per reference store, versus 0/18 and 0/16 under suppression. These are controlled omissions within our implementation, not a head-to-head run of the earlier system. Separate evidence comes from Shen et al. (2026): default Graphiti retrieval returns revoked policies (App. D).

One-version trajectory and native scenarios. One post-submission candidate passes 12 controls using persisted SQLite authorization and native admission into a local receiver. Source invalidation blocks a granted derivative after restart; an unaffected item remains usable, and incomplete source records are refused (App. F.4). Both stores and the retrieval, context-assembly, grant-authority, oracle and provenance modules match the submitted artifact byte-for-byte (supplement, CODE-COMPARISON.json). All four native scenarios are replayed from retained candidates. The current query-aware tie-break selects admissible evidence for approved widening and complete intersection. Overlapping audiences remains a top-1 miss; consolidation still misses some supporting evidence.

Recall regime. Parity is now scoped to complete candidate coverage (§6.2). The added LoCoMo regime, with constructed audiences, makes the window bind: a 20-slot post-filter underfills 145/1,981 queries. Recall@5 over 1,531 eligible questions is 56.03% with that window versus 56.39% authorizing first (Table 22).

Human verification. Yes. A non-author reporting no prior exposure assessed 141 cases with method names and prior verdicts hidden: seven flags, 14 paired answers and 120 stratified random negatives. An author with prior exposure assessed 57 overlapping cases. Each originally marked two flags as substitutions, sharing one positive. Neither marked a definite substitution in the paired answers. The non-author's negatives contain 0 Yes, 117 No and 3 Not sure. Original and corrected records remain separate (App. F.8, Table 24).

Presentation. The prose is simplified; the formal model, proofs, hypotheses and main result tables are retained. The general response lists what changed.

Reviewer_4iQKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses cross-audience memory leakage in personal language agents by binding memories to their origin audiences and enforcing authorization before model invocation. Derived memories are partitioned, intersected, or suppressed, while audience widening requires explicit, object-specific grants. Experiments on 10,000 synthetic multi-party histories report zero forbidden context inclusions and a 0.298 improvement in entitled Recall@5 over unscoped retrieval. However, the reduction in wrong-principal substitutions is not statistically significant.

Critical strengths
  1. The paper addresses an important privacy problem in persistent agent memory by enforcing authorization on the assembled model context rather than relying on model behavior or output filtering.
  2. The framework covers write-time binding, derived-memory authorization, explicit grants, and per-attempt admission, with conditional soundness and policy-completeness guarantees under stated assumptions.
  3. The evaluation includes two independently persisted reference architectures, 10,000 paired histories, and extensive conformance tests. The prespecified statistical protocol and complete outcome accounting strengthen the evaluation.
Critical weaknesses
  1. The core admission rule follows [1], while related authorization and provenance principles are established in [2,3]. Although the lifecycle extensions are clearly described, their additional theoretical value beyond these principles remains insufficiently established.
  2. The main evaluation relies on synthetic histories generated from a single controlled event law with deterministic language templates. The LongMemEval, LoCoMo, and EverMemBench overlays contain no competing cross-audience evidence, resulting in identical retrieval performance across all six arms. Moreover, Appendix E.4 states that semantic grading has not been performed on the 7,647 external answer outputs. For 140 questions, different generation models are used across arms, making their paired answer comparisons model-confounded. These experiments provide limited evidence of generalization to realistic multi-party conversations.
  3. The practical advantage over simpler authorization-aware retrieval strategies remains unclear. Table 6 reports identical Recall@5 for B-FLAT, B-GRAPH, bounded post-filtering, and principal postings. Section 6.2 states that the distractors never exhaust the post-filter's 20-item window. Moreover, Table 14 reports positive-retrieval median latencies of 575 ms for B-FLAT and 815 ms for B-GRAPH under warm-cache conditions, while Table 15 shows only 134 objects in the corresponding workload. More challenging distractor settings and scalability experiments with larger memory stores are needed.
  4. The effectiveness of the proposed lifecycle mechanisms is not sufficiently evaluated. The intersection-versus-suppression study includes only eight controlled cases. Appendix E.6 reports 15 incorrect abstentions across three native scenarios involving approved owner-private widening, complete intersection, and overlapping audiences. Appendix E.2 also states that grant revocation, deletion, and member removal are not evaluated as complete policy behaviors. Additional end-to-end experiments are needed to establish the availability and effectiveness of the proposed lifecycle mechanisms.
  5. The reduction in wrong-principal substitutions is not established by the prespecified statistical tests. Section 6.3 reports zero substitutions for B-FLAT and B-GRAPH versus seven for unscoped retrieval, but H3a/b do not reject the null hypotheses. Furthermore, Appendix B.3 and E.5 describe multiple amendments to the AI judging procedure. Served model versions and per-judgment metadata are unavailable for some judging routes, and the approved consistency audit was not completed. These limitations weaken the answer-level evidence.

[1] Liu, Sibo. "Authorization Before Context: A Model-Neutral Audience Boundary Against Cross-Audience Memory Leakage in Agentic Systems." arXiv 2026. [2] Namboothiri, Rohith. "Authorization-First Retrieval: Enforcing Least Privilege in Multi-Agent RAG Systems." Workshop on TrustNLP 2026. [3] Rezazadeh, Alireza, et al. "Collaborative memory: Multi-user memory sharing in llm agents with dynamic access control." arXiv 2025.

Questions for authors

Please refer to the weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 4iQK

Thank you for the detailed review.

Novelty and theory. We extend [1]'s admission rule across stored writes, derivation, grants and later requests. [2] checks document access before learned retrieval; [3] uses access graphs and source-tagged memory. Our audiences come from authenticated conversation participants, and checks cover each request's final input (§2, Table 12). P1–P4 extend the earlier properties to grants and source validity; P5 covers derivation. §3.5 separates four properties rather than treating authorization as one filter.

External data. Merged LoCoMo contains 5,882 records from ten human-edited, model-generated conversations with constructed audiences. Over 1,531 eligible questions, Recall@5 is 56.39% authorizing first versus 52.08% unscoped. Post-filtering reaches 52.08%, 55.28% and 56.03% at windows 5, 10 and 20; window 20 underfills 145/1,981 queries (App. F.6, Table 22). These retrieval metrics do not depend on answer-generation models (App. E.4).

Simpler strategies. Principal postings enforces the same policy through another index; equal recall shows no measured recall loss from authorization. With fixed scores and stable ties, post-filtering matches the allowed top-k only when its window reaches the k-th allowed record (§3.5). The window sweep measures this difference directly.

Latency. Table 15 reports similar whole-operation retrieval medians for audience labels and postings: 575 versus 585 ms for B-FLAT and 815 versus 812 ms for B-GRAPH, on a 134-object workload with warm caches. The larger-corpus study measures retrieval quality.

Withdrawal and native results. Grant removal, revocation and invalidation are now tested; membership changes affect later writes. The complete mechanism exposes 0/27 prohibited opportunities per reference store and 0/8 natively; stale state exposes 19/27 per reference store, and retrieval-only checks expose 5/8 natively (Table 7). Derived labels retain 18/18 challenge and 16/16 ordinary operations per reference store, versus 0/18 and 0/16 under suppression. For the 15 incorrect native abstentions, the needed evidence is admissible in all three scenarios. In the retained-candidate replay, the current query-aware tie-break selects it in two; overlapping audiences remains a top-1 miss (App. F.4).

Judging. H1, H2 and H3c use retrieved records and the fixed evidence ledger, not AI answer judgments. The completed AI audit and human assessment address answer labels (Apps. F.3, F.8); the general response states the unchanged H3a/b outcome.

Thank you for the detailed rebuttal and substantial revisions. I appreciate the authors' efforts to address my concerns through additional experiments and clarifications. In particular, the merged LoCoMo evaluation, lifecycle ablations, end-to-end request checks, and human assessment provide stronger evidence for the proposed framework. These additions have addressed several of my concerns regarding the evaluation and implementation.

However, I still have some reservations about the novelty, practical benefits, and generalizability of the approach. The main contribution remains largely an integration of established authorization and provenance mechanisms. While the new experiments demonstrate the importance of lifecycle enforcement and candidate completeness, the benefits over simpler authorization-aware baselines remain relatively modest. In Table 22, authorize-first achieves 56.39% Recall@5 compared with 56.03% for bounded post-filtering, while principal postings achieves identical performance. The external evaluation also relies on constructed audiences rather than naturally evolving multi-party permissions.

I appreciate the additional lifecycle tests and human verification, which have strengthened the evaluation. However, the improvement in wrong-principal substitutions remains statistically inconclusive.

Overall, the rebuttal has strengthened the paper and clarified several important issues, but my main concerns regarding novelty, practical advantages, and generalizability remain. I will therefore maintain my original rating of 2.

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

This paper addresses cross-audience information leakage in persistent agent memory by enforcing authorization throughout the memory lifecycle. It introduces audience binding at write time, provenance-aware rules for derived memories, object-specific grants, and authorization checks before memory enters the model context. The framework is implemented in two independent reference architectures and a native agent-memory runtime.

Experiments on 10,000 synthetic multi-party histories report zero unauthorized memory inclusions, while maintaining Recall@5 parity with policy-equivalent baselines and substantially outperforming unscoped retrieval. The reported reduction in wrong-principal substitutions remains statistically unconfirmed.

Critical strengths

Important problem and well-defined security boundary. The paper addresses a practical privacy risk in persistent multi-user agents. Evaluating unauthorized exposure at the assembled model context, rather than only at the generated output, is a meaningful and well-motivated design choice.

Coherent lifecycle framework. The proposed mechanisms cover memory creation, derivation, sharing, and retrieval under a unified authorization contract. The conditional guarantees are clearly stated, and the independent flat-store and graph implementations strengthen the systems contribution.

Careful experimental design and transparent reporting. The large-scale paired evaluation, policy-equivalent baselines, prespecified statistical tests, and explicit reporting of inconclusive and unfavorable results provide useful evidence without overstating the observed outcomes.

Critical weaknesses

Unclear propagation of dynamic authorization changes. The framework permits audience narrowing, but it is unclear how previously derived memories remain consistent when a source's audience is subsequently restricted. Does source-specific narrowing trigger transitive relabeling or invalidation, including for descendants with existing grants? An explicit lifecycle invariant and corresponding tests would strengthen the correctness claim.

Insufficiently specified provenance for mixed-audience derivations. The partitioning mechanism assumes that derived outputs can be assigned appropriate source audiences. However, when an LLM transformation consumes inputs from multiple audiences, its outputs may depend on information beyond their assigned source partition. Are transformations isolated by audience, or does each output inherit the complete input provenance? Clarifying this execution boundary and testing potential cross-audience contamination would be valuable.

Ambiguity in grant-issuer authority. The paper requires grants to be authenticated and object-specific, but does not clearly specify who is authorized to widen an item's audience. Authentication alone does not establish declassification authority, particularly when the original information was contributed by another participant. The framework would benefit from an explicit grant-issuer authorization rule and tests for unauthorized issuers.

Limited evidence for the full lifecycle claims and retrieval generality. The strongest empirical results concern authorization under deterministic lexical retrieval, where authorized candidate enumeration largely explains recall preservation. The bounded post-filter baseline is not stressed beyond its candidate window, and the more extensive conformance results partly come from earlier implementation revisions. Current-revision end-to-end lifecycle tests and clearer qualification of the "no recall cost" claim would better support the paper's broader conclusions.

Questions for authors
  1. Dynamic authorization propagation: When a source memory's audience is narrowed after derived memories have been created, how are the descendants' permissions updated? Does the implementation enforce transitive relabeling or invalidation, including for descendants with existing grants?

  2. Provenance in mixed-audience derivation: When an LLM transformation consumes memories from multiple audiences, how does the system ensure that partitioned outputs do not incorporate information from unauthorized sources? Are transformations isolated by audience, or does each output inherit the complete input provenance?

  3. Grant authority: Who is permitted to authorize the widening of a memory's audience? Does the framework distinguish an authenticated grant issuer from one actually authorized to declassify the underlying information, particularly when that information originated from another participant?

  4. Scope of empirical validation: Which lifecycle guarantees have been verified end-to-end on the evaluated implementation revision, rather than only on earlier versions? Additionally, should the reported "no recall cost" conclusion be interpreted as specific to the exhaustive, fixed-ranking retrieval setting used in the experiments?

1 comment
AuthorsOpenReview ↗

Response to Reviewer Jrtx

Thank you for the detailed review.

Q1, narrowing. Recorded audiences stay unchanged; membership changes govern later writes. Historical withdrawal uses grant removal or item/source invalidation. Invalidation blocks every descendant, even one holding its own grant, from the next request (§3.3, App. E.2). The lifecycle test confirms this after restart, while an unaffected item remains usable (App. F.4).

Q2, mixed-audience derivation. The reference stores support both options: an audience-separated output uses sources with one identical audience; a combined output records every input and permits only readers allowed to access them all (§3.2). Intersection labeling retains 8/8 legitimate derived operations in the matched test. The native publisher accepts only same-audience outputs. Request checks refuse altered text, undeclared inputs and incomplete source records (App. F.4). A separate post-submission successor binds outputs to their generating requests and rejects added or omitted sources. It passes 16 regression tests, seven producer probes and two invalidation/restart cases. Table 21 retains the original diagnostic beside the successor results.

Q3, grant authority. The owner may approve owner-authored sources or sources whose policy names the owner as authority. Other material requires its contributor's authenticated consent for that item and destination; mixed items require authority over every source. Authentication alone grants no sharing authority (§3.2). Eight authority checks pass, including rejection of owner-only approval where contributor consent is required, and of missing-consent, wrong-destination and wrong-principal variants (App. F.4).

Q4, revision and recall. The confirmation and 30 selected checks ran on the evaluated revision. The post-submission candidate passes 94 differential cases, detects 71 applicable mutants and passes 22 transformations. Both stores and the retrieval, context-assembly, grant-authority, oracle and provenance modules match the submitted artifact byte-for-byte (supplement, CODE-COMPARISON.json).

Yes, recall parity requires complete candidate coverage; §6.5 tests binding windows. Exact cosine on all 10,000 histories reproduces exclusion and baseline parity, with Recall@5 0.995 versus 0.892 unscoped. There, forbidden text cannot change the selection, given corpus-independent fixed vectors, exhaustive authorized ranking before truncation and stable ties (App. F.1).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response: summary of the revision

We thank the reviewers for their careful and constructive feedback. The revision shows how access restrictions survive summarization, sharing, source invalidation and restart. It separates four properties requiring different evidence:

  1. Context exclusion. No forbidden record entered 200,000 reference-store contexts. On merged LoCoMo with constructed audiences, unscoped retrieval returns cross-conversation records in 1,291/1,981 queries; every access-checked method returns none (App. F.6).
  2. Candidate completeness. The confirmation needed at most 13 post-filter slots, so its 20-slot window never bound. On merged LoCoMo, that window underfills 145/1,981 queries. Recall@5 over 1,531 eligible questions is 56.39% authorizing first, 56.03% with the window and 52.08% unscoped (App. F.6, Table 22).
  3. Selection independence. It holds for the tested exact-cosine configuration under App. F.1's conditions. An alternative restricting lexical statistics and fusion to allowed text changes top-5 order in 33,207/100,000 cells and lowers LoCoMo recall by 1.43 percentage points (§6.5).
  4. Lifecycle validity. A post-submission code version passes 12/12 request controls: after source invalidation and restart, an item is blocked despite its grant (App. F.4). On the fixed ablation panel, the complete mechanism exposes 0/27 prohibited opportunities per reference store and 0/8 natively; stale state exposes 19/27 and 6/8, and retrieval-only checks expose 5/8 natively (Table 7).

Answer-level checks. All 10,000 unscoped answer requests carried forbidden records; no reference-store request did (App. F.7). Original scoring recorded 0 detected substitutions per reference store versus 7 unscoped; H3a/b and the joint criterion remain not established. The AI consistency audit is complete (App. F.3). Neither human assessor, one a non-author, marked a definite substitution in the 14 paired reference-store answers or the negatives they assessed. Table 24 reports the flagged-case assessments, original and corrected records, and unsure responses separately.

What changed. The paper is rewritten for readability; the admission rule, hypotheses, tests, original results and proofs are retained. §3.1 and App. E.2 now state the implemented withdrawal rule. Added material appears in §3.5, §5.3, §§6.4–6.5 and App. F. Derivation and admission exchange §§3.2–3.3; submitted Tables 6, 14 and 15 become Tables 5, 15 and 16. "Competing records" means distractors, "wrong-person" means wrong-principal, and "reference stores" means the two architectures.

Area_Chair_deVQOpenReview ↗

Reminder: Please Read and Discuss the Author Rebutta

Dear Reviewers,

Thank you again for your valuable reviews and discussions. As the author-reviewer discussion period will end on November 18 (AoE), I would kindly ask you to take another look at the authors’ rebuttal if you have not yet done so.

Please make use of this final opportunity to discuss any remaining questions or concerns with the authors, particularly if your current assessment is still borderline or uncertain. I hope the discussion can help clarify the key issues and lead to a fair and clear final assessment.

Thank you very much for your time and contribution.

AC, ICLR 2027

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.