Atria Dawn Update Brings Editable PPT, Protein Folding, Benchmark Gains
Shanghai AI Lab's Atria Dawn agentic AI system now generates editable PowerPoint files with native shapes, creates academic posters from papers, optimizes protein structure prediction using OpenFold, performs multi-file business analysis, and shows improved benchmark scores on AutomationBench, SkillsBench, WideSearch, and JobBench.
Editable PowerPoint Generation
Atria Dawn can produce a complete PowerPoint presentation from a narrative outline. In the demonstrated animation, approximately 440 particles transition through signal, waveform, spectrum, blueprint, print, disintegration, and reassembly phases to form the Atria logo across 29 slides using smooth transitions. Crucially, every particle and caption remains a native PowerPoint shape or text box, allowing users to edit text, adjust elements, or rearrange slides after delivery. This addresses a common workflow pain point where revisions require reworking the source file.
Academic Poster Creation
Given a research paper, the system can generate an editable A1 portrait ICLR-style academic poster. It extracts content, organizes charts and formulas, and arranges the layout. The output retains editable components, enabling researchers to refine the poster without starting from scratch.
Protein Structure Prediction Optimization
Using the OpenFold framework (which implements the AlphaFold 2 architecture) and with network search disabled, Atria Dawn ran a controlled experiment on roughly 19,000 protein chains across two GPUs. The main training consisted of 3,000 steps with over 80 million fine-tuned parameters. The optimized configuration improved local structural consistency by about 2.7% compared to the baseline fine-tuning setup. This case illustrates how the agent organizes a scientific computing experiment, handling input feature adaptation, training strategy, and geometric constraints.
Multi-File Business Data Analysis
The system processed five business documents to perform cross-table reconciliation, profit calculation, and anomaly detection for six categories of consumer electronics, then visualized the operational results with charts.
Benchmark Results
Official benchmarks show improvements across eight metrics versus the preview version. Notable scores: AutomationBench 54.7, SkillsBench 67.8, WideSearch 84.6 (all ranking first in their respective comparisons). JobBench rose from 50.3 to 57.5, moving from sixth to third place. These figures provide a reference; users are encouraged to test the system on their own familiar tasks.
Community Projects
The platform also showcases community creations. "Earth's 400 Million Years" visualizes global plate reconstruction data as an interactive timeline map, letting users observe continental changes from 400 million years ago to present and link geological events to map transformations. "Trace" presents a historical knowledge base in a scrollable, ink-wash style reading space with a "star map" of figure-event relationships and an AI historian for queries. Both projects demonstrate how specific interests can drive unique, usable applications.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
