acceptodds
Withdrawn from ICLR 2027

SlideTrace: Governed Memory for Cross-Project Slide Personalization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-project slide personalization requires an agent to carry reusable user preferences across projects while preserving their scope and evolution over time. Raw feedback histories leave these properties implicit, making it difficult to determine which preferences remain applicable as interactions accumulate. We introduce SlideTrace, a governed memory framework that represents reusable feedback as atomic, scoped preference claims and maintains them across project working memory and canonical long-term memory. Semantic relation proposals and deterministic validation govern how claims are updated, related, and consolidated across projects. We evaluate SlideTrace against four memory systems and a no-memory reference under a shared generator. After 20 sequential training projects, frozen memory states are evaluated on 15 held-out tasks. SlideTrace achieves the highest overall Score under both GPT-5.4 and Kimi-K3, exceeding the strongest baseline by 0.38 and 0.51 points, with the clearest gains in requirement realization. Ablations further support relation-governed canonicalization and structured memory packages. These findings show that cross-project personalization benefits from governing reusable preference state rather than merely retaining interaction history. Code and data are available in our anonymous repository at https://anonymous.4open.science/r/slidetrace-anonymous-artifact-1315.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
preference memory, slide generation, agent memory, iterative generation, personalization, presentation generation, evidence-governed memory

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_igr9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents SlideTrace, a memory framework for cross-project slide personalization. It represents user feedback as structured preference claims with explicit scope, relations, lifecycle states, and supporting evidence. The framework maintains project-level working memory and consolidates selected preferences into long-term memory for reuse in subsequent projects. Experiments with synthetic personas show improved requirement realization and overall scores compared with several memory baselines.

Critical strengths

The paper addresses a meaningful problem: retaining reusable user preferences across projects while preventing temporary or outdated feedback from influencing future generation. Its explicit separation of preference content, applicability, and lifecycle state provides a reasonable approach to this problem.

Critical weaknesses
  1. The technical contribution beyond existing personalized memory systems is limited. MemSlides already maintains persistent user profiles, project-level working memory, preference routing, conflict handling, and cross-project consolidation. The main difference in SlideTrace therefore appears to be its more explicit representation and governance of preference claims, rather than the overall memory architecture. While this is a reasonable extension, the paper does not sufficiently establish which additional capabilities it enables or why existing consolidation mechanisms are inadequate. The Append-Only ablation in Table 3 provides some supporting evidence, but it does not directly evaluate the correctness or necessity of individual relation-governed updates.

  2. The evaluation does not directly validate the claimed ability to manage evolving and conflicting preferences. A central motivation is that preferences may be refined, contradicted, restricted to particular scopes, or replaced over time. However, the evaluation mainly uses synthetic personas with a fixed preference applicability map and does not separately test these forms of preference evolution. The paper also reports little difference when removing the evidence gate or using the No-Scope variant. Moreover, Section 4.3 describes No-Scope as a generation-time variant constructed from the original frozen store, leaving unclear whether it actually removes scope governance during memory construction and retrieval. Consequently, the experiments do not establish how reliably the proposed system handles the preference-management cases that motivate its design.

  3. Table 2 shows that only 23 of 180 held-out requirement slots are matched to claims retrieved and exposed during generation. Although 22 of these 23 requirements are satisfied according to both evaluators, the comparison with non-exposed requirements is observational and subject to selection effects, as the authors acknowledge. The overall score improvements are encouraging, but the analysis does not clearly establish how much of the gain results from explicitly transferred preferences. The small five-task ablation also limits confidence in the relative contributions of the proposed components.

  4. The baseline memory states and computational costs require further clarification. In Appendix E, Figure 13 reports 88 active long-term claims for SlideTrace but only 6 active entries for MemSlides, with 83 additional entries marked as superseded. Although these lifecycle labels may have different meanings across systems, the substantial difference raises questions about the effective memory available to each method. Similar total store sizes alone do not establish comparable usable memory. Furthermore, the efficiency analysis in Figure 2 measures memory tokens injected into generation but does not account for the full cost of semantic analysis, verification, and memory updates. These details are important for interpreting both the performance improvements and the claimed efficiency benefits.

Questions for authors
  1. What specific preference-update capabilities does SlideTrace provide beyond the profile reconciliation and consolidation mechanisms in MemSlides? Can the authors provide concrete examples of relation-governed updates that existing methods fail to handle?

  2. How accurately does SlideTrace identify contradictions, exceptions, temporal updates, and preference scopes? Does the benchmark contain explicit cases that test these operations? In the No-Scope ablation, are scope assignment and retrieval filtering also disabled?

  3. Can the authors quantify the contribution of explicitly transferred preferences through a matched removal experiment on held-out tasks? How sensitive are the transfer-trace results to the heuristic claim-matching procedure in Appendix B?

  4. Why does the adapted MemSlides baseline retain only 6 active entries after training? How do these states affect retrieval, and what are the total memory-management costs across the compared systems?

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

SlideTrace is a memory framework for cross-project slide personalization. It represents user feedback as scoped preference claims and manages their updates through semantic relations, lifecycle rules, and execution evidence. Experiments with a shared generator and frozen memory show improvements in requirement realization over the compared memory systems.

Critical strengths
  1. The paper addresses a practical problem and makes a useful distinction between project-specific feedback and reusable preferences. Tracking applicability and updates gives the framework a clear motivation.

  2. The shared-generator and frozen-memory protocol provides a reasonable basis for comparing memory systems. The ablations also support the value of memory consolidation and structured input.

Critical weaknesses
  1. My main concern is the gap between the governance mechanisms emphasized in the paper and what the experiments directly validate. Removing the evidence gate or scope constraints has little effect, and there is no targeted evaluation of conflicting, withdrawn, or conditional preferences. I think the results support the overall pipeline more clearly than these specific governance claims.

  2. I am less convinced that the evaluation establishes user-specific personalization. Some preferences resemble general requirements for professional reports, and the limited synthetic-user setting makes it difficult to separate learning an individual’s preferences from improving task completeness. The lack of repeated runs and human evaluation further limits confidence in generalization.

Questions for authors
  1. Could the authors evaluate controlled cases involving preference conflicts, withdrawal, and scope changes? Such tests would help clarify when the governance mechanisms matter and whether they prevent incorrect memory updates.

  2. Could the authors compare users with incompatible preferences on the same task, ideally with human assessment? I would find this helpful for distinguishing personalization from general improvements in report quality.

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

The paper studies cross-project personalization for slide generation. Reusable user preferences must persist across projects without mixing with project-specific edits. The authors argue raw feedback histories and retrieval alone are insufficient. They leave scope, applicability, and lifecycle implicit. SlideTrace instead represents preferences as atomic, scoped claims with direction, priority, scope, conditions, and lifecycle state. It maintains two timescales: project working memory and canonical long-term memory. Semantic proposals generate claims and relations. Deterministic validators govern scope, lifecycle transitions, and canonical updates. Consolidation into long-term memory requires execution evidence and project-level confirmation. Evaluation uses a synthetic benchmark with two personas, 20 sequential training projects, and 15 held-out tasks. Under a shared frozen generator, SlideTrace is compared to four memory systems and a no-memory baseline. It reports the highest overall Score under both GPT-5.4 and Kimi-K3, exceeding the strongest baseline by 0.38 and 0.51 points. Gains are clearest in requirement realization. Ablations examine relation-governed canonicalization and structured memory packages.

Critical strengths

S1. Well-motivated problem formulation. Framing cross-project personalization as governed preference-state management is a sensible conceptual contribution. Atomic, scoped claims with explicit lifecycle address a real failure mode: conflating project-local edits with reusable preferences. This framing may be useful beyond slides.

S2. Controlled evaluation protocol. Sequential training, frozen memory, then held-out generation prevents test-order contamination. A shared generator and fresh sessions isolate memory effects from generator differences. The protocol design itself is a useful reference for evaluating agent memory.

S3. Transparent ablation reporting. The paper includes generation-time and write-time ablations. Null results, such as the evidence gate showing no consistent gain, are reported openly. This honesty helps readers calibrate which components actually matter.

Critical weaknesses

W1. The improvement over baselines is marginal. The headline gains are 0.38 and 0.51 points on a 10-point scale, over the strongest baseline. With only 15 held-out tasks and no significance testing, it is unclear whether these differences are meaningful or within noise. The practical significance for end users is not established.

W2. The additional benefit of memory itself is not demonstrated. The paper does not show that governed memory improves efficiency, cost, or convergence. There is no analysis of revision rounds to acceptance, token usage, latency, or dollar cost, and no learning curves over the 20 training projects. The governance pipeline adds LLM calls for proposal, relation classification, and validation, yet this overhead is never quantified. Without a cost-benefit analysis, the added complexity is hard to justify over simpler baselines or no memory at all.

W3. Synthetic and narrow validation. Two synthetic personas and one fixed curriculum are used. No human users, no real drift, no order variation. LLM-as-judge is used without human correlation. Generalization to real feedback is unproven.

Questions for authors

Q1. What is the extra cost of SlideTrace? Please report LLM calls, tokens, and latency per project vs Append-Only and MemSlides.

Q2. Does memory reduce the number of generation/revision rounds needed to reach acceptable quality? Please report rounds-to-acceptance or score-vs-project learning curves for SlideTrace and baselines.

Q3. Are the 0.38/0.51-point gains statistically significant? Please report confidence intervals or paired tests across the 15 held-out tasks. What does a ~0.4-point Score gain mean for a user? Is there any human study showing users prefer SlideTrace outputs?

Q4. Given the evidence-gate ablation shows no consistent gain, under what conditions does governed canonicalization outperform simple append-only memory?

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

Related papers

Discussion (0)

Sign in to comment.