acceptodds
Under review as a conference paper at ICLR 2027

MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision

Abstract

Personalized presentation generation requires more than conditioning on a current prompt or template: an agent must preserve stable user preferences across tasks, retain newly introduced preferences during multi-turn revision, and carry out local edits without disturbing the rest of the deck. We present MemSlides, which organizes preference and execution memory by lifetime and role. User profile memory stores intent-conditioned preferences and personalizes the initial deck; working memory keeps session constraints active across later edits; and tool memory stores reusable execution experience for local revision. Before generation, intent-matched profiles are reconciled with the current request, with explicit instructions taking precedence over stored preferences. Stable preference signals are consolidated after each session for reuse across tasks. During revision, execution contracts define edit scope and targets, local snapshots guide targeted patches, and coverage checks verify completion before finalization. Across three model families, MemSlides outperforms DeepPresenter and SlideTailor on most persona-alignment dimensions while maintaining competitive general presentation quality. Human evaluation further supports improved preference alignment, while localized-revision experiments show that MemSlides more reliably completes requested changes without disturbing non-target content.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Personalization, Slides Generation, Agent Memory, Multi-turn Interaction, Local Revision, Preference, Self-improving

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_576WOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MemSlides, a hierarchical memory framework for personalized slide generation and multi-turn local revision. It separates persistent user profiles, session-level working memory, and reusable tool experience. A Plan–Act–Guard editing pipeline uses explicit edit scopes, local snapshots, and coverage checks to complete revisions while preserving non-target content. The evaluation includes persona alignment, general presentation quality, human preference ratings, and controlled comparisons of memory organization and localized editing.

Critical strengths
  1. The paper addresses persistent personalization and reliable local editing, both of which are important for interactive presentation authoring.
  2. Distinguishing user preferences, session constraints, and tool experience provides a sensible organization for managing information across tasks and revision turns.
  3. Explicit scope control, snapshot-bound patches, and coverage checks provide a practical mechanism for reducing unintended changes.
  4. The paper combines model-based judgments, human ratings, and controlled memory comparisons, providing evidence from several perspectives.
Critical weaknesses
  1. The profiles include predefined persona information, so it is unclear how much improvement comes from learning user preferences. A comparison without this extra information would help.
  2. Some results are inconsistent. Table 9’s Overall scores do not always match the stated averaging rule, and the GLM-5 quality claim in Appendix A.5 conflicts with Table 2. These should be clarified and corrected.
Questions for authors

How does performance change when profiles contain only interaction-derived preferences, without seeded completion?

How many distinct decks were included in the human study, which decks did each participant evaluate, and was condition order randomized?

2 comments

Acknowledgment

We thank the reviewer for recognizing the importance of persistent personalization and reliable local editing, our memory organization and revision mechanism, and the evaluation from several perspectives.

Weaknesses

W1 & Q1 Comparison using only interaction-derived preferences

The profiles include predefined persona information, so it is unclear how much improvement comes from learning user preferences. A comparison without this extra information would help.

How does performance change when profiles contain only interaction-derived preferences, without seeded completion?

Thank you for suggesting this comparison. Without predefined persona information, interaction-derived preferences yield a 4.83-point improvement in preference following, from 3.00 to 7.83 on a 0–8 scale.

We evaluate 6 cases, 1 per simulated user, under 3 conditions: No profile, only preferences retained from 3 prior interaction rounds, and the same interaction-derived profile processed with seeded completion. Each case uses the same new task and material across conditions; the current request does not repeat the preferences. We have added this comparison to Appendix A.18 (Table 21).

Table 1. Preference following across profile conditions. Preference Score is the mean total across 4 interaction-derived preference ratings (0–2 each; 0–8 total), using the median of 3 blind GPT-5 judgments per preference. Full Match reports how many of the 6 cases satisfy all 4 preferences.

ConditionPreference Score ↑Full Match ↑
No profile3.000/6
Only interaction-derived preferences7.835/6
Interaction-derived preferences + seeded completion6.833/6

No profile vs. only interaction-derived preferences: The 4.83-point gain and increase in Full Match from 0/6 to 5/6 show that profiles containing preferences expressed in prior interactions effectively guide new tasks. The improvement does not rely on extra predefined persona information.

Only interaction-derived preferences vs. interaction-derived preferences + seeded completion: Preference Score is lower with seeded completion (6.83 vs. 7.83). This may reflect changes in how the interaction-derived preferences are prioritized when combined with additional persona information. The rubric evaluates only the 4 interaction-derived preferences, so additional persona characteristics receive no separate credit. Both profile conditions outperform No profile (3.00), further supporting the role of interaction-derived preferences in guiding new tasks.

W2 Scoring definition in Table 9 and wording correction in Appendix A.5

Some results are inconsistent. Table 9’s Overall scores do not always match the stated averaging rule, and the GLM-5 quality claim in Appendix A.5 conflicts with Table 2. These should be clarified and corrected.

Thank you for pointing out the unclear caption in Table 9 and the incorrect description in Appendix A.5.

  1. Table 9: An LLM assigns Overall scores by considering Content, Structure, and Visual jointly, and we report the mean of 3 such judgments. Overall is not the arithmetic mean of the Content, Structure, and Visual scores. The Average row averages each column across 10 personas before rounding. We have corrected the caption and scoring description; the reported values are unchanged.
  2. Appendix A.5: We incorrectly stated that MemSlides achieved the highest Avg. on GLM-5. Table 2 reports 3.74 for MemSlides, below DeepPresenter’s 3.86. We have corrected the statement to indicate the highest Avg. only on GPT-5 and revised the related discussion.

Questions

Q2 Human-study decks, assignment, and condition order

How many distinct decks were included in the human study, which decks did each participant evaluate, and was condition order randomized?

The study included 9 distinct pre-generated decks. For each of 3 preference profiles, MemSlides, Mem0, and No profile each generated 1 deck.

Each of the 25 participants evaluated the 3 decks for the profile matching their preference-questionnaire responses. They rated each deck according to their own preferences on Direct, Content, Organization, and Visual using a 1–5 satisfaction scale, yielding 75 deck ratings.

Yes, condition order was automatically randomized by the questionnaire platform. Decks were presented under anonymous labels.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your valuable feedback and thoughtful evaluation of our work.

In our rebuttal, we added a comparison showing that interaction-derived preferences alone improve Preference Score from 3.00 to 7.83 without predefined persona information. We also clarified the Overall scoring procedure in Table 9, corrected the GLM-5 quality statement in Appendix A.5, and explained the human-study deck assignment and randomized presentation order.

We would like to follow up to ask whether our rebuttal and additional experiments have adequately addressed your concerns. If any questions remain, we would be happy to provide further clarification.

If you feel that your concerns have been resolved, we would greatly appreciate it if you could consider raising your score accordingly.

Thank you for your time and consideration.

Best regards,
The Authors

Reviewer_9rtuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents MemSlides, a personalized presentation-generation agent that organizes memory into user profile memory, working memory, and tool memory. Before generation, intent-matched profile items are reconciled with the current request, with explicit instructions taking precedence over stored preferences. During multi-turn revision, an execution contract defines edit scope and targets, local snapshots guide targeted patches, and coverage checks verify completion. The validation includes persona-alignment judgments on three model families against DeepPresenter and SlideTailor, a general-quality check, a 25-rater human preference study, a five-persona comparison with Mem0 and no memory, a nine-pair tool-memory ablation, and a 24-task localized revision suite.

Critical strengths
  • The system design is coherent and operationally specific. Separating user profile memory, working memory, and tool memory, with explicit reconciliation and precedence rules, is a sensible response to the different lifetimes of personalization signals.
  • The Plan-Act-Guard revision protocol, with execution contracts, snapshot hash binding, and a coverage gate, is a concrete mechanism for preserving already aligned content.
  • The paper evaluates three model families, two strong slide-generation baselines, a human preference study, a five-persona memory comparison, a matched-pair tool-memory ablation, and a 24-task revision suite. The persona-alignment gains on GLM-5 and Gemini 3.1 Pro are large and consistent, and the human ratings favor MemSlides on all four items.
Critical weaknesses
  • Table 1 reports averages over three personas, but the personas are not identified, and the GPT-5 results do not appear consistent with the ten-persona results in Table 9. It is therefore unclear how the reported scores were obtained or whether the two evaluations use the same profiles. The authors should clarify the persona selection and explain why the larger evaluation is only conducted with GPT-5.

  • The tool-memory ablation is based on only nine matched pairs. Although Strict Verify and Core Tool Time Ratio show improvements (p = 0.0195), neither remains significant after correcting for multiple comparisons. The other two metrics are also not statistically significant, and the overall time ratio is sensitive to an extreme case. These results provide limited evidence for the claimed benefits of tool memory.

  • Table 9 states that Overall is the average of Content, Structure, and Visual, but the reported scores do not follow this definition. For example, the Marketing Manager has an Overall score of 5.95, while the average of the three dimensions is 8.50. The average row also appears inconsistent with the stated calculation. The authors should clarify the scoring method and correct the reported values.

Questions for authors

See Weakness

7 comments

Acknowledgment

We thank the reviewer for recognizing the coherent system design, the concrete revision mechanism, and the positive persona-alignment and human-rating results.

Weaknesses

W1 Clarifying the results in Tables 1 and 9

Table 1 reports averages over three personas, but the personas are not identified, and the GPT-5 results do not appear consistent with the ten-persona results in Table 9. It is therefore unclear how the reported scores were obtained or whether the two evaluations use the same profiles. The authors should clarify the persona selection and explain why the larger evaluation is only conducted with GPT-5.

Thank you for pointing out the unclear description. We clarify the three concerns below:

  1. The three personas in Table 1: Financial manager, Legislator, and Software developer.
  2. Score differences and calculation: Table 9 reports a separately conducted extension to 10 profiles, rather than a per-profile breakdown of Table 1, so the scores need not match. Table 1 uses 3 profiles from Table 9, with identical profile content. Each score in Table 1 averages 3 blind judgments per profile and then averages across 3 profiles. Table 9 reports per-profile scores, with its Average row averaging each column across 10 profiles.
  3. Persona selection and why only GPT-5: Table 1 uses a 3-persona subset to compare slide-generation systems across GPT-5, GLM-5, and Gemini 3.1 Pro, testing personalization across different model families. Table 9 tests whether personalization remains effective across 10 profiles, so we use a single model, GPT-5, to focus the comparison on profile variation.

W2 Tool-memory effect sizes and statistical robustness

The tool-memory ablation is based on only nine matched pairs. Although Strict Verify and Core Tool Time Ratio show improvements (p = 0.0195), neither remains significant after correcting for multiple comparisons. The other two metrics are also not statistically significant, and the overall time ratio is sensitive to an extreme case. These results provide limited evidence for the claimed benefits of tool memory.

Thank you for your attention to statistical robustness. We expanded the tool-memory ablation to 18 matched pairs (6 per model) using the same configuration as Table 4 and observed substantial gains in editing reliability and efficiency. Closed-Loop Completion increases from 0.815 ± 0.170 to 0.963 ± 0.108, and Strict Verify from 0.310 ± 0.109 to 0.534 ± 0.102 (mean ± SD across pairs). Mean Time to First Correct Edit decreases from 567.8 to 239.0 s (57.9% reduction), and the geometric mean Core Tool Time Ratio is 0.327× (67.3% reduction).

The expanded tool-memory ablation shows statistically significant improvements on all 4 metrics after correcting for multiple comparisons (all corrected p < 0.05; Table 1), supporting gains in workflow completion, verification, and editing efficiency.

Table 1. Paired tool-memory results across 18 pairs from 3 models. Exact one-sided sign tests exclude ties and unavailable pairs.

MetricWithout tool memoryWith tool memoryWins / losses / ties / NARaw pCorrected p
Closed-Loop Completion ↑0.8150.9639 / 1 / 8 / 00.0107420.042969
Strict Verify ↑0.3100.53416 / 2 / 0 / 00.0006560.002625
Time to First Correct Edit (s) ↓567.8239.015 / 3 / 0 / 00.0037690.015076
Core Tool Time Ratio ↓1.000×0.327×16 / 2 / 0 / 00.0006560.002625

Values are means across pairs; Core Tool Time Ratio is the geometric mean of paired time ratios, with time without tool memory normalized to 1.000×. Corrected p = min(1, 4 × Raw p) uses Bonferroni correction across the 4 metrics, calculated from unrounded p-values.

The Core Tool Time Ratio is not sensitive to a single extreme case. It remains 0.303–0.347× after leaving out any 1 pair and 0.217–0.505× after leaving out any 1 model, with all ratios below 1.

Together, the expanded paired evaluation and sensitivity analysis provide consistent evidence of improved workflow reliability and efficiency across the evaluated tasks. As additional support, the real editing case in Appendix A.11 illustrates how tool memory helps successive edits: it prompts the agent to use the latest snapshot, completing the border-color change while preserving the preceding title-color change.

W3 Overall definition and averages in Table 9

Table 9 states that Overall is the average of Content, Structure, and Visual, but the reported scores do not follow this definition. For example, the Marketing Manager has an Overall score of 5.95, while the average of the three dimensions is 8.50. The average row also appears inconsistent with the stated calculation. The authors should clarify the scoring method and correct the reported values.

We apologize that the original caption was unclear. An LLM assigns Overall scores by considering Content, Structure, and Visual jointly, and we report the mean of 3 such judgments. Overall is not the arithmetic mean of the Content, Structure, and Visual scores.

Marketing Manager’s 5.95 is the mean of 3 separate overall scores assigned by the LLM, instead of the mean of its Content, Structure, and Visual scores. Averaging the 3 dimension scores gives 8.50, which is a different calculation from the reported Overall metric.

The Average row averages each column across 10 personas before rounding. For example, averaging the 10 persona-level MemSlides Overall scores gives 7.70165, reported as 7.70. The calculations and reported values are correct; we have corrected the caption and scoring description to make this LLM-based evaluation explicit.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer,

Thank you again for your valuable feedback and thoughtful evaluation of our work.

In our rebuttal, we clarified the profiles, scoring procedures, and different evaluation goals of Tables 1 and 9, and corrected the description of Overall. We also expanded the tool-memory evaluation to 18 matched pairs: all 4 metrics improve significantly after correcting for multiple comparisons, and the sensitivity analysis shows that the core tool time reduction persists after excluding any 1 pair or model.

We would like to follow up to ask whether our rebuttal and additional experiments have adequately addressed your concerns. If any questions remain, we would be happy to provide further clarification.

If you feel that your concerns have been resolved, we would greatly appreciate it if you could consider raising your score accordingly.

Thank you for your time and consideration.

Best regards,
The Authors

Reviewer_hS8wOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

MemSlides addresses personalized slide generation and multi-turn localized revision through user-profile memory, working memory, and tool memory. It combines these with a Plan–Act–Guard editing workflow and introduces a multi-persona profile bank. Experiments across three model families report improved preference alignment and revision reliability over selected baselines.

Critical strengths
  1. The paper addresses a common problem in slide editing: fixing one thing can unintentionally change something else. Requiring the system to make the requested change while preserving unrelated content is a sensible criterion that reflects what users expect.

  2. The memory ablations are useful, particularly the comparison with a single memory list. Both conditions receive the same information and context budget, which helps show whether organizing the memory differently actually improves performance.

Critical weaknesses
  1. The memory design mostly combines familiar ideas, including separating long-term and working memory, retrieving user preferences, and reusing execution experience. These choices have engineering value, but the paper does not clearly explain what is new about the memory mechanism itself. The Plan–Act–Guard workflow also follows a familiar agent loop.
  2. The paper lacks discussion of recent work on slide generation[1-2]. Including these works would help clarify how MemSlides differs from existing methods and what it adds.
  3. The paper also lacks a comparison with general harnesses such as Codex. Without testing such a baseline with comparable models, tools, preference information, and budgets, it is hard to tell whether the specialized framework provides benefits beyond what a well-configured general-purpose agent could achieve.

[1] Presenting a paper is an art: Self-improvement aesthetic agents for academic presentations

[2] Design First, Code Later: Aesthetically Pleasing Template-Free Slides Generation

Questions for authors
  1. Did the 25 human raters actually interact with the system across revision turns, or only assess pre-generated decks? Were ratings based on their own preferences or supplied personas?
  2. Under comparable models, tools, information, and budgets, does MemSlides outperform a general-purpose harness such as Codex? Which components explain any advantage?
  3. For each table, how many independent source tasks, generated slides, revision rounds, and repeated runs were evaluated?
5 comments
AuthorsOpenReview ↗

W1 (Part 1)

Acknowledgment

We thank the reviewer for recognizing the importance of preserving unrelated content during slide editing and the value of our controlled memory ablations.

Weaknesses

W1 Novelty of persistent personalization and localized revision

The memory design mostly combines familiar ideas, including separating long-term and working memory, retrieving user preferences, and reusing execution experience. These choices have engineering value, but the paper does not clearly explain what is new about the memory mechanism itself. The Plan–Act–Guard workflow also follows a familiar agent loop.

Our memory design does not simply combine familiar ideas, and our work is definitely NOT a work that has only engineering value, for the following reasons:

  1. Integrating general-purpose memory does not sufficiently address persistent personalization. Our research question is how to accumulate user preferences from revision feedback and carry them into subsequent tasks and edits while preserving already aligned content. This is not a trivial problem, and a crucial challenge is to determine which preferences remain active, which slides they apply to, and which feedback should update the persistent profile. As shown in Figure 5 of the original paper, directly using Mem0, which already connects long-term memory with an agent’s working memory, yields lower persona alignment than MemSlides across all 5 personas (7.65 vs. 8.45 on average). These results indicate that simply integrating an existing memory framework is insufficient to achieve the same personalization performance. Therefore, resolving this challenge requires more than combining familiar ideas such as separating long-term and working memory and retrieving user preferences.
  2. Our memory design supports persistent personalization by accumulating preferences from interaction and applying them throughout initial generation and subsequent revision. Mem0 (Chhikara et al., 2025) maintains historical information through fact extraction and updating, while MemoryOS (Kang et al., 2025) supports personalized responses through hierarchical storage, topic-based retrieval, and profile updating. However, these mechanisms do not specify how persistent preferences, the current intent, and temporary feedback should jointly guide content selection, information organization, and visual style in presentation authoring. In contrast, we select preferences by intent and presentation dimension, reconcile them with the current request, and use them in initial generation. During revision, compatible requirements remain active, conflicts are overridden within their applicable scope, and stable feedback is consolidated for reuse across tasks. Execution experience is also retrieved at modify-job start and before similar tool calls to support reliable execution. For example, a request to “reduce text on this slide” should change the current slide without discarding organizational requirements on other slides or directly replacing the user's persistent content-density preference. Our contribution therefore goes beyond simply combining familiar memory modules: we specify how these memories jointly support persistent personalization, allowing new feedback to adapt the current deck while retaining still-active requirements and carrying stable preferences into subsequent tasks.
  3. Our localized revision mechanism addresses edit drift, where satisfying a local request unintentionally changes non-target content. Auto-Slides (Yang et al., 2025) locates LaTeX code blocks and rewrites, inserts, or deletes them through a ReAct-style editor. PPTPilot (Ofengenden et al., 2025) routes instructions to python-pptx scripts or direct XML edits, then renders and verifies the result for further correction. However, neither method explicitly specifies how active preferences constrain the write scope or how edits remain bound to the current target snapshot during successive revisions. In contrast, we translate the current request and active preferences into explicit edit targets, read structured local snapshots containing DOM node paths, parent–child relationships, and exposed style rules, and bind patches to the corresponding nodes or rules. In Table 5 of the original paper, MemSlides completes 24/24 tasks, compared with 12/24 for Auto-Slides and 17/24 for PPTPilot; success requires both the requested change and preservation of non-target content. Our contribution is to connect preference state, edit scope, and snapshot-bound operations, so that active requirements guide each revision while reducing unintended changes to already aligned content.

The remainder of our response to W1 is provided in W1 (Part 2).

AuthorsOpenReview ↗

W1 (Part 2)

This continues our response in W1 (Part 1).

  1. Plan–Act–Guard enforces these requirements throughout execution. We adopt a familiar agent loop, while our contribution links preference state, edit scope, and completion conditions: Plan converts the current request and active preferences into an execution contract; Act applies targeted patches according to that contract; Guard validates snapshot binding and, when coverage is required, blocks premature completion with missed targets. Together, these stages keep active preferences effective across revision turns while reducing unintended changes to already aligned content.

Together, (1)–(4) demonstrate an advance over existing memory and slide-revision methods in supporting persistent personalization. Our contribution is to reconcile persistent preferences, the current intent, and temporary feedback, and make these requirements govern edit scope, actual operations, and completion conditions. This enables preferences accumulated from interaction to guide generation and subsequent revision while preserving already aligned content, going beyond simply combining familiar ideas or adopting a standard agent loop.

We provide further discussion of the contributions and related work in Appendix A.16 of the revised manuscript.

AuthorsOpenReview ↗

W2, W3 & Q2

W2 Recent slide-generation work

The paper lacks discussion of recent work on slide generation[1-2]. Including these works would help clarify how MemSlides differs from existing methods and what it adds.

We thank the reviewer for this suggestion. We have cited EvoPresent (Liu et al., 2025) and DeepSlides (Cui et al., 2026) in Section 2 (Related Work), with a detailed comparison in Appendix A.16.

  1. EvoPresent focuses on aesthetic self-improvement for academic presentations, using PresAesth to guide iterative refinement and version selection through aesthetic scores and feedback. In contrast, MemSlides uses user feedback to maintain active preferences and applies requested changes while preserving non-target content.
  2. DeepSlides focuses on coordinating page design and code implementation for deck-level style and page-level layout, supporting task-time user style requirements and feedback-based slide revision. MemSlides reconciles persistent preferences with the current intent and temporary feedback to guide initial generation and subsequent revision, distinguishing temporary overrides from persistent profile updates.

Neither work explicitly studies cross-task preference maintenance or non-target preservation during successive user edits. By linking preference maintenance with localized revision, MemSlides supports preference reuse without repeated specification and helps reduce edit drift during multi-turn revision.

W3 & Q2 Comparison with Codex and component contributions

The paper also lacks a comparison with general harnesses such as Codex. Without testing such a baseline with comparable models, tools, preference information, and budgets, it is hard to tell whether the specialized framework provides benefits beyond what a well-configured general-purpose agent could achieve.

Under comparable models, tools, information, and budgets, does MemSlides outperform a general-purpose harness such as Codex? Which components explain any advantage?

Thank you for the suggestion. We added two comparisons (personalized slide generation and localized revision) with the official Codex CLI, both using GPT-5/medium. Both tables are included in Appendix A.17 (Tables 19 and 20).

1. Personalized slide generation. Table 1 compares how well MemSlides and 2 Codex configurations align generated decks with user profiles across 6 tasks (2 papers × 3 profiles). All methods receive the same source materials and full profiles under the same 2-hour budget. We evaluate final decks on the 4 persona-alignment dimensions used in Table 1 of the original manuscript.

Table 1. Persona alignment in personalized slide generation.

MethodContent ↑Structure ↑Visual ↑Specificity ↑
Codex6.887.124.716.53
Codex + MemSlides tools8.208.176.818.02
MemSlides8.308.557.118.04

Codex uses native Python PPTX; Codex + MemSlides tools accesses MemSlides generation interfaces. Scores are means across 6 tasks. MemSlides has higher mean scores on all 4 dimensions in both comparisons, particularly Structure and Visual. The Hierarchical Memory component selects intent-matched preferences, gives current instructions precedence, and keeps compatible requirements active, so that persistent preferences guide generation and subsequent revision.

2. Localized revision. Table 2 tests whether giving Codex access to MemSlides’ localized revision mechanism improves revision success on 50 tasks. Both conditions share initial slides, requests, resources, general tools, and budgets, and neither uses memory. Success requires a correct, visible, and valid result that preserves non-target content.

Table 2. Localized revision success with Codex on 50 tasks, with 1 run per task per condition.

MethodSuccessful tasksSuccess rate
Native Codex35/5070.0%
Codex + MemSlides’ localized revision mechanism42/5084.0%

The localized revision mechanism increases success by 14 percentage points. The localized revision component uses execution contracts, snapshot-bound patches, and coverage checks to bind edits to explicit targets and prevent missed targets or unintended changes. Table 2 directly tests its benefit within the same Codex harness.

These results show better performance on the evaluated slide-authoring tasks and improved Codex revision success with our mechanism, rather than general superiority over general-purpose harnesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.