Explore Franklin AI
Understand new AI research with readable explanations and links to the original papers.
A proposed method scores two possible causal directions from one pair of event sequences. It uses language models as sequence probability estimators, with randomized encodings to r
Researchers evolve an MCP server by testing individual tool changes with smaller models. The resulting Rocq interface improves held-out theorem-solving results, while its Lean port
Oren Salzman's Divide-and-Collapse framework simplifies an existing multi-agent plan by finding which agents actually need joint reasoning. Its reported speedups concern this restr
A controlled study varies rollout source, teacher coupling, and loss direction across 1,200 adaptation runs. Its findings show why self-distillation choices depend on the model and
Researchers propose local entropy weighting for language-model reinforcement learning. Mathematical reasoning experiments favor PEPO over uniform and batch-level weighting, while i
Controlled experiments separate concealing a message from computing through a concealed scratchpad. Most tested setups resist the latter, but a convenient cover task provides an im
A study of learned links between agents finds increased harmful compliance under benign training and deliberate manipulation. Link-only repair reduces the measured risk but does no
Controlled experiments identify spatial locality, state–successor binding and ordered generation as important for exact continuation. High pixel accuracy alone hides compounded rul
A signed transfer map predicts which task types support a held-out target under a fixed training budget. The study finds that more sources and greater similarity are unreliable sel
A training-free controller maintains partner-conditioned plans and compares imagined actions. Overcooked experiments test role retention and adaptation, with results tied to the ki
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.
