Oct 11, 2026
RECAST trains a small router to compute evidence before answering
A learned router combines retrieval, SQL and generated Python to construct task-specific evidence from mixed sources.
Explore Franklin AI
AI Research Papers
Understand new AI research with readable explanations and links to the original papers.
Oct 11, 2026
Open-MMUnlearning compares multimodal forgetting with retained capability and recovery risks
The open-source toolkit evaluates whether removing target knowledge also damages useful capabilities, and whether evaluation metrics survive attempts to recover that knowledge.
Oct 11, 2026
TRACE analysis finds limited evidence for predicting harmful context compaction
Recent agent history weakly predicts extra errors and repeated calls after a summary replaces its working context.
Oct 11, 2026
Sherpa trains an AI tutor around simulated students' learning gains
Adaptive teaching improves benchmark results and teacher preferences, but the study does not measure real students' long-term learning.
Oct 11, 2026
BOTTLED tests whether agents can replace millions of model calls with reusable solutions
The benchmark measures quality and cost when agents create programs or small models for large repetitive workloads.
Oct 11, 2026
PersonaPlex dialogue study finds coordinated speech with late turn changes
Two full-duplex models respond to each other, but their turn-taking timing differs from the human telephone-speech reference.
Oct 11, 2026
nanoMuse describes an open personal agent running across phones and computers
The report explains local runtimes, cross-device coordination and approval controls, while distinguishing implemented features from roadmap goals.
Oct 11, 2026
WorldSolver tests generated physics solvers beyond whether their code runs
The benchmark combines execution, rendered behavior and numerical physical checks across 168 simulation tasks.
Oct 11, 2026
VeriFine improves embodied reasoning by revising its judge and training curriculum
Two linked improvement loops adapt policy training and human-calibrated verification as new failure patterns appear.
Oct 11, 2026
ParanoiaEval tests whether coding agents stop defensive work when the evidence warrants it
Paired repository tasks measure unnecessary precautions separately from successful task completion.
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.