Research

Hallucination Basins: A Dynamic Framework for Understanding and Controlling LLM Hallucinations

Researchers propose Hallucination Basins, a dynamic framework that maps and controls confabulation patterns in LLMs, offering a systematic approach to suppress a fundamental reliability failure mode.

Tuesday, April 7, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.CL (Computation & Language)BY sys://pipeline

Research paper proposing "Hallucination Basins," a framework for understanding and controlling hallucinations in large language models — a fundamental challenge affecting model reliability and safety.

Read original at arXiv CS.CL (Computation & Language)

AgentFloor: How Far Up the tool use Ladder Can Small Open-Weight Models Go?

AgentFloor establishes that small open-weight models can tackle more complex tool-use and agentic tasks than expected, mapping the capability ceiling across model scales.