- Rating
- 2(+1)
- Confidence
- 4
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper identifies cross-entity interference, as distinct from sequence-length forgetting, in recurrent models that maintain a fixed-size shared state over interleaved entity events. It proposes Entity-Conditioned Addressing (ECA), which adds a stable, identifier-derived key to both the write and read addresses of a scan-compatible recurrence. Its principal ECA-Hash variant uses deterministic Rademacher keys without an entity-specific learned table. An idealized superposition analysis predicts expected crosstalk energy proportional to entity load divided by address width, O(M/N), under stated independence assumptions. Evaluation covers the synthetic e-SET entity-state benchmark, an independently constructed security-grammar task, official Mamba-1/2/3, Naju and Gated DeltaNet backbones, and probes on WinTap ACME, CTU-13 and Google cluster traces. The most substantial gains arise when a query must retrieve an entity’s stale state amid interleaved updates; the paper also reports cases where ECA is neutral or harmful.
Critical strengths
- Well-isolated evidence for the address-location mechanism. The comparison holds the identity code and input encoder fixed while varying whether the key is injected into recurrent read/write addresses (Table 2). Several input-only alternatives remain near the no-identity floor at M = 16, whereas ECA-Hash reaches 0.865 on Mamba-2 and 0.977 on Gated DeltaNet. The independent security grammar reproduces the effect across five recurrent backbones (Table 3). Key resampling, write/read mismatch, one-sided keying and query-swap tests (Appendices B and F) provide unusually direct evidence that stable, aligned addressing—not merely the presence of entity metadata—drives retrieval.
- A meaningful characterization of scope and cost, rather than an unqualified accuracy claim. The WinTap ACME entity-state probe shows sizeable gains across five backbones (Table 5), and the study varies entity load independently of horizon and address width (Fig. 2). The constant persistent-state property is explicit; Appendix F also reports parameter, state and throughput measurements. Importantly, the authors document negative results on attack detection, CTU-13 and GCT (Sec. 8–9), helping establish when this inductive bias is appropriate rather than presenting it as a universal security improvement.
Critical weaknesses
- Practical significance beyond a routing-specific diagnostic remains uncertain. The strongest real-data result is a deliberately constructed stale entity-state probe on ACME rather than a demonstrated gain on a native security objective. The paper itself reports no consistent benefit for ACME attack detection or the GCT probe; CTU-13 can even exhibit poorer scenario-held-out recall under keyed configurations (Sec. 8.1–8.3). Thus the evidence supports a conditional memory mechanism more strongly than it supports broad real-world utility. The paper would be more compelling if it quantified how often representative applications actually enter the interference-bound regime.
- The capacity argument does not directly guarantee scaling or efficiency in trained recurrent backbones. The O(M/N) law concerns an idealized random-key superposition with key-independent values and decay (Sec. 4), not the full learned gated systems. Fig. 2 already shows degradation at higher relative load, and the method requires reliable, stable entity identifiers (Sec. 9). Although state size remains constant, Appendix F reports non-fused reference-path throughput reductions for several official implementations; the available fused Naju measurement does not establish equal deployment efficiency across all backbones. This limits the strength of the claimed capacity/performance tradeoff. Figure/presentation concern (secondary): The compact ECA flow diagram (Fig. 1, PDF p. 4) uses small annotations, while the load–accuracy comparisons (Fig. 2 and Appendix Fig. A1) rely on fine dashed traces, small point markers and a crowded legend. Key capacity trends require zooming to inspect, and larger labels and better visual separation of baseline curves would improve interpretability. This visual issue does not independently undermine the reported mechanism.
Questions for authors
- Routing-bound prevalence and native-task impact (assessment-critical). Can the authors measure, on ACME and at least one additional real event stream, the fraction of evaluation queries that actually require retrieving a stale entity-specific value rather than recency/content-based inference? When results are stratified by this property, do ECA gains translate to a naturally occurring downstream objective? A demonstrated benefit would materially strengthen the significance assessment.
- Robustness to identifier instability and implementation budget (assessment-critical). Beyond the key-overlap stress tests already in Appendix F, what happens under realistic identifier churn, aliasing or occasional incorrect IDs on the ACME probe? Can matched-budget throughput or end-to-end latency be reported for a fused ECA implementation in more than the Naju host? These answers would clarify whether the method’s assumptions and computational overhead remain acceptable in its intended deployment regime.
Reviewers: please respond to the author rebuttal
Dear reviewers,
Thank you to those who have already engaged with the authors' responses. The author-reviewer discussion closes on November 18 (AoE).
If you have not yet replied, please read the authors' response to your review and post a reply here before then: whether your concerns have been addressed, what remains open, and whether you are updating your score. If your assessment is unchanged, please say so explicitly. Engaging with the other reviews and the ongoing discussion is also welcome.
Thank you, Area Chair