DeckSmith

Description

DeckSmith is an environment for research-grounded slide-deck generation. Each task gives the agent a topic and structure brief — but not the content. The agent must research the topic on the web, then author a professional PowerPoint deck that is graded against a fine-grained, time-robust rubric by a multimodal gpt-5-mini judge.

Capabilities

• Open-web research and synthesis via search + page-fetch tools
• Turning a brief into a coherent, well-structured presentation
• Multimodal slide authoring: content, structure, and visual layout
• Grounding claims in current, accurate information (no fabrication)

Compute Requirements

Agents are given a sandbox with 2 CPUs and 2GB of RAM to research and author the deck.

License

MIT

Tasks

A single train split. Each task provides a topic/structure instructions.md (title, audience, purpose, slide-count range, generic format and visual requirements, and a research mandate) reverse-engineered from a real presentation. The agent is not given the source content — it must research it.

Reward Structure

Sparse, single-submission, continuous scoring. The agent authors a .pptx and calls submit_slides once; the environment renders it (pptx → pdf → page images) and grades each rubric item yes/no with a multimodal judge. The rubric has four equally-weighted classes (25 pts each, 100 total):

Per class the score is the weighted pass rate; the reward is the overall 0–100 score normalized to [0, 1]:

$s_i = w_i \cdot \frac{\#\text{yes}_i}{\#\text{valid}_i}, \qquad \text{Reward} = \frac{\sum_i s_i}{100}$

The content classes are judged against the original deck's text as reference (which the agent never sees). Rubrics are generated to be time-robust — they reward durable, topic-defining content rather than dated or volatile specifics, so an agent researching current information is neither penalized for missing a stale figure nor rewarded for reproducing an outdated one. The judge is gpt-5-mini (multimodal, no temperature).

Tools

• web_search — search the web (Tavily) for information on the topic
• fetch_url — fetch the readable text of a page (paginated)
• PowerPoint tools (via PowerPointToolset) to author the deck
• PDF tools (via PDFToolset) and standard CLI tools (bash, read, write, edit, glob, grep, ls, todo_write)
• submit_slides — submit the finished .pptx for grading (callable once)

Time Horizon

Multi-turn: the agent researches, authors, and iterates before a single grading submission. [Statistics on average tool calls here]

Other Environment Requirements

DeckSmith requires two secrets: an OpenAI API key (openai_api_key) for the multimodal rubric judge, and a Tavily API key (tavily_api_key) for the agent's web-search and page-fetch tools.

Safety

Agents research public web content and author presentation slides. The environment does not present direct safety risks; the primary quality concern the rubric targets is factual accuracy — grounding claims in real sources and not fabricating content.

Citations

@article{zheng2025pptagent,
  title={PPTAgent: Generating and Evaluating Presentations Beyond Text-to-Slides},
  author={Zheng, Hao and Guan, Xinyan and Kong, Hao and Zheng, Jia and Zhou, Weixiang and Lin, Hongyu and Sun, Le and Han, Xianpei},
  journal={arXiv preprint arXiv:2501.03936},
  year={2025}
}

@misc{GRDeckSmith,
  author    = {General Reasoning Inc. Team},
  title     = {DeckSmith},
  year      = {2026},
  publisher = {OpenReward},
  url       = {https://openreward.ai/GeneralReasoning/DeckSmith}
}