anmorgan.bsky.social
@anmorgan.bsky.social
🚀 How can we detect LLM hallucinations—without external tools or model intrinsics?

SelfCheckGPT is a zero-resource, reference-free evaluation approach by analyzing the consistency of multiple responses from the same model. Let’s break it down. 🧵👇 (1/11)
March 27, 2025 at 4:15 PM
Borrowing on traditional #ML ensembling methods, LLM juries consists of multiple LLM judges that independently score a given output, then aggregate their scores through a voting function like max or average pooling. (2/6)
February 24, 2025 at 6:31 PM
Single-model evaluations can be biased and inconsistent. LLM Juries, which use multiple models for assessment, offer a more reliable alternative—reducing bias and improving robustness. 🧵 (1/6)
February 24, 2025 at 6:31 PM
G-Eval consolidates evaluations into a single metric, effectively providing the model with a unified scorecard.

It's secret to success is combining chain-of-thought (CoT) reasoning with a form-filling paradigm and scoring function. (2/4)
January 31, 2025 at 4:31 PM