Doubao Work Agent: Local Excel/PPT Editing Transforms ICML 2026 Paper Analysis

6500 Rows of Excel, No Formulas Touched

The author, a researcher handling thousands of accepted papers from top AI conferences (ICLR, AAAI, ACL, CVPR, ICML), previously spent afternoons uploading files to cloud AI tools, waiting for edits, and downloading results — a fragmented workflow. Doubao Work's new local Office editing agent promises to keep Excel and PPT editing in a single, uninterrupted flow.

Using a 6500-row ICML 2026 spreadsheet containing paper titles, authors, and a primary_area_en column (research area), the author first adjusted row spacing and background colors via natural language in the Doubao Work sidebar. Then, to find the hottest research direction, the author selected the primary_area_en column and prompted: "Create a new sheet with a pivot table based on the selected column."

The agent did not execute blindly. Instead, a left-side interaction panel revealed a rigorous, multi-step process:

Load spreadsheet skill, read SKILL.md;

Read compatibility and pivot table documentation;

Inspect workbook structure — found only one sheet named "论文列表";

Read header and data samples to confirm data range;

Execute code lark-cli sheets ... to generate the pivot table.

The entire process took 3 minutes 18 seconds. A new "数据透视表" tab appeared. The pivot table showed Deep Learning accounting for nearly 40%; the author manually verified the original sheet and confirmed Deep Learning's share was 43% — no calculation error. The save button remained under the user's control, providing confidence.

Next, the author asked for a bar chart of paper counts per direction. The agent quickly produced a well-aligned, clearly labeled chart with suitable colors, ready to paste directly into a lab meeting slide.

Rescuing a Half-Baked PPT

The author then turned to a poorly formatted ICML 2026 trend analysis PPT: black background with white text, mixed Times New Roman and Microsoft YaHei fonts, no images, text-heavy and dull. Instead of starting from scratch, the author opened the PPT in the same Doubao Work conversation sidebar and asked the agent to report the deck's fonts, main colors, and layout characteristics.

The agent correctly identified mixed fonts (Microsoft YaHei and Times New Roman), primary colors (dark blue and gray), and a content-card layout. The author spot-checked slide 2's title and confirmed Times New Roman, validating the agent's comprehension.

The author then instructed the agent to unify fonts and backgrounds across the entire deck according to the detected visual system. The agent also inserted the visualization charts generated earlier in the Excel phase onto each slide. The result was a professionally styled, consistent presentation — the author did not search for a single image or align a single text box; the visual upgrade was fully automated.

Final Thoughts

The author recommends not starting from blank documents. Instead, feed Doubao Work the messiest experimental data that consumes hours of manual cleaning each month, or the most dreaded half-finished lab meeting PPT. After processing, open the file locally in Excel or PowerPoint — seeing formulas intact, data linked, and layout polished is more convincing than any description. Reserve time for real research; delegate the drudgery to AI.

As a side note, searching "豆包工作" and downloading the desktop client offers a free 60-day membership (30 days initially, renewable for another 30 days).