Every displayed development links to its source; legacy editions are capped at two items per publisher.
- LLMs can be 'confidently wrong' in finance; internal states needed for detectionFactual summary
New research highlights that LLMs can exhibit 'confident hallucinations' in financial question answering, where external outputs appear correct but underlying reasoning is flawed. Reliably detecting these issues requires analysing internal model activations, moving beyond surface-level evaluation. This directly impacts G-SIB model validation, necessitating enhanced real-time risk monitoring frameworks to improve trust in deployed financial AI applications.
- Widespread bug identified in LLM repetition penalties across inference enginesFactual summary
A critical bug has been discovered in common repetition penalty implementations within major LLM inference engines, including HuggingFace and vLLM. This flaw introduces unpredictable model behaviour and undermines a core control for text generation quality and safety. Model validation and inference teams must immediately assess the impact on in-production models and internal stacks to mitigate unexpected outputs and associated model risks.