Researchers made AI agents better by turning academic papers into working code — no retraining required

ScholarEvolve uses published research to improve AI agents, raising Qwen's AppWorld task completion from 69% to 81.4% without retraining the model.

Expanding Automated Product Reel Generation to Avoid Instagram's Hidden Areas

The ProblemI operate a system that uses a Python script to automatically generate vertical short videos (Instagram Reels) to showcase products for a client's online store. By placing photos and a ...
xenospectrum

ScholarEvolve Lets AI Agents Rewrite Their Own Code From Research Papers, Lifting Task Completion 12.4 Points

UCSB and Microsoft's ScholarEvolve generates code from academic papers to improve an agent's harness without retraining model weights. It raised Qwen3.5-27B's task completion rate to 81.4%. Here is ...

Running 27B AI Locally: Comparing the 'Speed and Accuracy' of Qwen 27B and Bonsai 27B

Using the same machine and two models with 27 billion parameters each, I compared how much they actually differ in terms of 'which is faster' and 'which can answer correctly' by running 10 tests three ...
shattered.io

Build Cursor Cloud Agents: 12 Steps, 90 Min [2026]

Learn to build Cursor Cloud Agents step by step: create environments with the new Environments API, scale to 300 repos, and share results as canvases.