Oct 11, 2026
SquidAgent estimates when parallel agents save more time than coordination costs
The proposed scheduler uses output-token budgets, shared context and upfront conventions to choose between serial and parallel work.
Explore Franklin AI
AI Research Papers
Understand new AI research with readable explanations and links to the original papers.
Oct 11, 2026
ScienceClaw checks whether scientific agents can retain verified workflow repairs
A 23-discipline benchmark separates fixing one scientific task from preserving a repair that improves independent tasks.
Oct 6, 2026
Models can identify an impossible mechanics problem and still label it solved
Paired valid and impossible engineering problems reveal a gap between recognizing a faulty premise and returning the status a downstream system will act on.
Oct 6, 2026
BiasFlow tests how frozen backbones respond to newly trained biased heads
Geometric diagnostics and supervised centroid alignment complement worst-group accuracy, while leaving attribute information recoverable and transfer outcomes mixed.
Oct 6, 2026
HEAR connects agent workflow intent with inference-engine runtime state
A bidirectional serving protocol lets harnesses and engines coordinate cache use and execution settings, with workload-specific speed and latency trade-offs.
Oct 6, 2026
The Pushback Paradox tests whether models obey both action and stop instructions
A two-probe diagnostic separates compliance in a reward conflict into active and passive rates, with important limits on claims about real-world shutdown behavior.
Oct 6, 2026
The Review Lottery estimates how often a new committee would change a conference decision
An observational model of ICLR scores estimates substantial decision disagreement, but cannot establish whether language models changed peer-review behavior.
Oct 6, 2026
ARBOR routes medical questions through selected low-rank adapter components
Conditional adapter selection improves medical QA benchmark averages in five-seed tests, while clinical safety and complete serving costs remain untested.
Oct 6, 2026
Collective intelligence depends on how a group combines connected judgments
Dietrich and List examine consistency, strategic manipulation and truth tracking when a group aggregates judgments on connected variables.
Oct 6, 2026
FREA shows why reaction verifiers need tests across different candidate sources
Expert labels for 751 chemical reactions expose source-dependent failures, including below-chance rankings for alternative product disconnections.
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.