Oct 16, 2026
Evaluation cues change model refusals across English and five Indic languages
A study finds that telling models they are under safety evaluation increases harmful-request refusals, while also increasing some benign refusals.
Explore Franklin AI
AI Research Papers
Understand new AI research with readable explanations and links to the original papers.
Oct 16, 2026
MLExpert learns from experiment windows in autonomous machine-learning work
MLExpert combines supervised training and window-based reinforcement learning, reporting a higher medal rate on MLE-Bench Lite and tests of transfer to other tasks.
Oct 16, 2026
Four Gemini agents complete a Minecraft run through shared planning and code
A research team reports a four-agent Minecraft completion after more than 27 hours, offering a case study in long-running coordination and recovery.
Oct 16, 2026
Random projection offers a lightweight baseline for private retrieval-augmented generation
Researchers apply a shared projection to query and datastore embeddings, reporting useful retrieval performance and empirical resistance to extraction attacks.
Oct 16, 2026
LangDiT studies which Transformer changes improve continuous language diffusion
LangDiT changes the Transformer backbone while preserving the underlying diffusion setup, reporting lower generation perplexity in matched experiments.
Oct 16, 2026
LEASH uses a smaller agent’s uncertainty to route and oversee computer-use work
LEASH combines initial routing, step-level escalation and continued oversight, reporting OSWorld improvements while reducing cost in one compared configuration.
Oct 16, 2026
A coding-agent study finds that control procedures still matter
An environment-agnostic agent improves on some benchmarks after prompt and controller changes, but the study reports mixed transfer and higher token costs.
Oct 16, 2026
FDE-Gym tests whether AI agents can deliver against ambiguous customer requirements
FDE-Gym evaluates 95 simulated delivery tasks using customer outcomes and interaction costs, finding gaps in both discovering constraints and executing solutions.
Oct 16, 2026
EliteReplay preserves multiple correct strategies during language-model post-training
EliteReplay combines reinforcement learning with an archive of diverse solutions, reporting gains in text-to-SQL strategy coverage and ambiguous-question handling.
Oct 16, 2026
KnowGPO rewards search agents for acquiring the knowledge an answer needs
KnowGPO groups search trajectories by acquired knowledge and reports gains over outcome-only and process-supervision baselines on three benchmarks.
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.