Explore Franklin AI
Understand new AI research with readable explanations and links to the original papers.
The framework connects off-track detection, human-assisted recovery and offline preference learning. Reported navigation success includes human correction, and the monitor remains
A policy proposal combines mandatory gates, version-bound evidence and available human review. Synthetic coding tests demonstrate its mechanics, not deployed safety.
An execution-grounded benchmark finds strong but uneven training intuition: models handle optimizer comparisons better than architectural choices and often overlook dataset changes
A learned interface edits value-related activations while holding a semantic code fixed. Tests on two frozen language models report less content drift than several steering and pro
ChronoGraph links object parts, actions and changing scene states in one graph representation, then trains vision-language models to use those traces for understanding and planning
A 500-scenario conversational benchmark scores unsupported agreement separately from empathy, exposing a trade-off when models receive accuracy-focused instructions.
A transit reinforcement-learning study tests a 60-second holding reserve against the same learned controller without it, reporting better passenger time and completion in fresh sim
Label-free reinforcement learning improves constrained greedy decoding on three structured tasks while moving search work into training.
A 146-task benchmark checks actual scientific files and application states, revealing frequent missing deliverables despite partial-credit performance gains.
A two-stage automated research method adds qualified behavioral probes after initial exploration, reporting gains over scalar-only feedback across five model families.
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.
