Oct 6, 2026
Collective intelligence depends on how a group combines connected judgments
Dietrich and List examine consistency, strategic manipulation and truth tracking when a group aggregates judgments on connected variables.
Explore Franklin AI
AI Research Papers
Understand new AI research with readable explanations and links to the original papers.
Oct 6, 2026
FREA shows why reaction verifiers need tests across different candidate sources
Expert labels for 751 chemical reactions expose source-dependent failures, including below-chance rankings for alternative product disconnections.
Oct 6, 2026
BTTF turns chip waveforms into queryable data for verification agents
A chip-verification study pairs SQLite waveform storage with specialized agents, reporting 95.33% accuracy on 150 annotated queries and a measurable latency trade-off.
Oct 6, 2026
TasteVal measures how efficiently AI researchers choose and interpret experiments
A new eight-task benchmark separates experimental judgment from coding, comparing models with human experts under the same GPU and time budgets.
Oct 5, 2026
Breast-biopsy AI study uses inferred gene expression to predict therapy response
NEO reports pooled AUROC of 0.792 across nine evaluation cohorts, with substantial variation and limits for clinical interpretation.
Oct 5, 2026
DepGPO follows command dependencies to assign terminal-agent training credit
The method traces writes and supporting reads back from verifier-inspected resources instead of crediting every step equally.
Oct 5, 2026
Zero-knowledge AI oversight paper finds a general limit and a signed-oracle exception
The theoretical work reports an impossibility for general oracle-aided computation, then shows how signed oracle answers change the privacy-verification problem under stated.
Oct 5, 2026
Time-series world models can predict well and still respond to actions in the wrong direction
An eight-dataset benchmark separates forecast error from agreement with declared action-state mechanisms, finding that architectural accuracy gains do not fix directional.
Oct 5, 2026
HazardWeaver checks whether an AI analysis route is scientifically valid and executable
A 141-instance benchmark grades method choice, outputs and justified abstention as hazard-analysis conditions change.
Oct 5, 2026
AVSD-Scenes adds generated audio-visual descriptions to 12,291 urban recordings
The dataset combines separate audio and visual captions, improving scene-level retrieval in the reported tests while leaving exact-clip retrieval much harder.
Good to know
How do I use these AI research explainers?
Start with the overview to understand the paper’s main idea, then follow its original paper link for the methods, results and limitations.
Can I suggest a research paper?
Use Submit Research to send the original paper URL and explain why it is useful. An editor reviews the suggestion before publication.