Explore Franklin AI
Understand new AI research with readable explanations and links to the original papers.
The study builds model organisms with selective withholding, resistance to elicitation and retained capability, for testing oversight methods.
A single model proposes actions, draws possible futures and reasons over them, with tests in puzzles and control environments.
The framework separates detecting a reasoning problem from choosing an intervention, with reported tests on overthinking and unanswerable questions.
The benchmark uses 6,885 wikiHow procedures to separate different planning skills, with long step-ordering tasks proving difficult.
A PHREEQC study finds that retrieving saved tool output can change which evidence an agent uses, hiding failures behind aggregate accuracy.
The pipeline checks generated tasks for solvability, difficulty and verifier quality, then trains agents in the resulting worlds.
Goal-shifted tasks expose failures that ordinary expert trajectories may leave hidden, even when individual states and actions are familiar.
The study finds that aggregate rankings can hide numeric weaknesses and depend on run budgets and access conditions.
The paper argues for controlled decoy opportunities as a monitoring method, while identifying robustness and deterrence as open research questions.
The graph-based benchmark measures consistency among related questions and how errors cluster, alongside ordinary accuracy.
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.
