Presentation Editing as Controlled Operations: The EditPPT Framework and DeckEdit-Bench Benchmark

8 November 202620 views

The authors break down slide editing into specific objects, permitted actions, and checks to ensure unrelated content remains untouched, rather than relying on free-form code generation. To measure quality, they created DeckEdit-Bench (183 tasks across 28 human-made presentations), and the multi-agent EditPPT framework achieves a 66.7–74.9% task success rate across five base models while preserving over 97% of the content.

Presentation Editing as Controlled Operations: The EditPPT Framework and DeckEdit-Bench Benchmark

Slide Editing as a Constrained Operation

Reliable slide editing requires precisely changing specified objects while preserving the rest of the presentation. The authors propose treating editing as a structured, constrained modification. Open-ended code generation gives models freedom, but does not guarantee that neighboring elements will be preserved.

An editing request can be broken down into three parts:

  • the objects to be changed;
  • the operations to apply to those objects;
  • the presentation state to preserve.

This breakdown sets the task boundaries before execution. The model receives not a general instruction to “fix the slide,” but a specific set of targets and constraints. This reduces the risk of unintended changes.

Practical criterion: before editing, explicitly specify the target objects, operations, and state to preserve. Without this, it is impossible to verify whether the instruction was followed and whether unrelated shapes were affected.

How DeckEdit-Bench Distinguishes Intended and Unintended Changes

To evaluate editing, the authors created DeckEdit-Bench. The benchmark includes 183 editing tasks across 28 human-created presentations. Target objects and objects to preserve are annotated at the shape level. Intended and unintended changes are evaluated primarily with deterministic rules.

The benchmark tracks two types of changes: intended and unintended. This distinction makes it possible to evaluate not only the editing result, but also its side effects.

The evaluation setup can be summarized as follows:

What is evaluatedHow the evaluation is structured
Task set183 tasks across 28 human-created presentations
Annotationstarget objects and objects to preserve at the shape level
Changesintended and unintended, evaluated primarily with deterministic rules

The benchmark evaluates two outcomes at once: whether the instruction was followed and whether the required objects were preserved. This approach distinguishes a useful edit from a side effect. Practical criterion: editing evaluation should include a preservation check, not just confirmation that the target object changed.

How EditPPT Works: Agents, Tools, and Verification

EditPPT is a multi-agent framework. It identifies target objects before editing. Each specialized agent is restricted to a subset of 37 predefined tools. Afterward, the framework checks whether the instruction was followed and the required objects were preserved. Unsuccessful changes are fixed or rolled back.

Workflow:

  1. Identify target objects before editing.
  2. Restrict the specialized agent to a subset of 37 predefined tools.
  3. Check whether the instruction was followed and the required objects were preserved.
  4. Fix or roll back unsuccessful changes.

Limiting the toolset narrows the agent’s action space. Verification and rollback add oversight after execution. Practical criterion: reliable editing requires not only a model, but also constraints on how changes are made.

Results Across Five Base Models

Across five base models, EditPPT achieves a Task Success rate of 66.7–74.9%. Preservation remains above 97%. These results show that the framework preserves the presentation while carrying out editing tasks.

With an open 27B model that was not fine-tuned on specific tasks, EditPPT comes close to Claude Code with Sonnet 5 in Task Success. At the same time, it does a better job of preserving the original presentation. This comparison matters because the open model was not fine-tuned on specific tasks.

ConditionTask SuccessPreservation
Five base models with EditPPT66.7–74.9%above 97%
Open 27B model without fine-tuning, with EditPPTcomes close to Claude Code with Sonnet 5better at preserving the original presentation

The authors conclude that reliable slide editing depends not only on model size, but also on constraints on how changes are made. Practical criterion: when choosing an approach, consider its mechanisms for constraining and verifying changes, not just the model’s size.

Practical Criteria for Editing Presentations

The framework and benchmark provide a set of criteria for organizing editing. They do not require open-ended code generation and rely on explicit constraints.

Criteria:

  • target objects are identified before editing begins;
  • objects to preserve are specified separately;
  • the agent works with a limited set of tools;
  • the result is checked for compliance with the instruction;
  • presentation preservation is checked separately;
  • unsuccessful changes are fixed or rolled back.

These criteria apply both to evaluating tools and to designing a workflow. The key takeaway: reliable editing depends on controlled operations and preservation checks. Model size matters, but does not replace constraints on how changes are made.

Frequently asked questions

Related materials

All materials
MathAdv: What the Final Accuracy of Machine-Generated Proofs Conceals
News

MathAdv: What the Final Accuracy of Machine-Generated Proofs Conceals

A diagnostic suite of tasks spanning 13 undergraduate and graduate mathematics topics evaluates machine-proving systems in Lean 4 not only by their final results, but also by their subject-matter mastery, the quality of their informal reasoning, and their robustness to rephrased conditions. A component-by-component analysis shows that translating a problem into formal form remains the main obstacle, and that success depends heavily on the area of mathematics and how the problem is framed.

9 November 2026
AffordanceWAM: Affordances as an intermediary between scene prediction and manipulator control
Research

AffordanceWAM: Affordances as an intermediary between scene prediction and manipulator control

The work describes a generative model that jointly predicts future RGB observations, scalar fields, affordance heatmaps, and continuous robot commands, using human videos to broaden interaction diversity without action annotations or retargeting. On RoboCasa, CALVIN ABC→D, and real-world tasks, the approach delivers consistent gains over RGB-only and robot-data-only baselines.

9 November 2026
Non-Euclidean Factor Geometry: The FactoMap Map for Disentangled Representations
News

Non-Euclidean Factor Geometry: The FactoMap Map for Disentangled Representations

A method is proposed for describing factor space through regions, generator-defined identifications, and local scales, making it possible to distinguish topologically similar cases with different geometries. FactoMap trains interpretable prototypes on a structured grid and, in experiments, demonstrates more robust preservation of factor continuity and separation.

8 November 2026
Presentation Editing as Controlled Operations: The EditPPT Framework and DeckEdit-Bench Benchmark