Explore Franklin AI
Understand new AI research with readable explanations and links to the original papers.
A two-stage automated research method adds qualified behavioral probes after initial exploration, reporting gains over scalar-only feedback across five model families.
A two-stage chemistry model separates generating plausible alternatives from selecting ones that satisfy natural-language preferences.
A structured generation workflow improves a 27B model's banking decks, but repeated AI grading complicates claims about small improvements or professional readiness.
An HKUST research team uses two AI agents to configure synchronized simulation data and task models for vehicle detection and wireless beam prediction.
A scoping review separates task-level gains from evidence that a replacement improves local decisions, emphasizing matched comparisons and downstream effects.
A synthetic red-team study of CIVeX shows how graph errors can certify harmful actions, while execution-only audits leave beneficial actions wrongly blocked.
Google Research authors describe a Gemini-based proof-discovery system that retains verified intermediate lemmas between attempts. They report five research results checked by doma
The Agent Error Dataset collects 50,228 error-diagnosis pairs with trace provenance. Matched replay tests support some proposed corrections, while training results distinguish impr
WorldAuditBench combines exploration with visual diagnosis across 213 anomaly tasks. Its best evaluated agent reaches 42.3% success, below the human baseline, with especially weak
PivotOPD targets decisive mistakes and the states they create. Its authors report improved multi-turn benchmark performance by combining preventive training with examples of how to
Start with the overview to understand the paper’s main idea, then follow its original paper link for the methods, results and limitations.
Use Submit Research to send the original paper URL and explain why it is useful. An editor reviews the suggestion before publication.
