'Likelihood'
Automated summary · Based on collected summaries; not individually verified
Prepared by HotAI
What happened
A new Meta paper proposes using cross-perspective consistency to judge the reliability of LLM answers."The authors argue that relying on 'likelihood' to verify LLM responses is flawed. Instead, they suggest trusting answers that remain consistent across different contexts. The paper posits that models learn common text patterns rather than truth, drawing on physics and philosophy to have the model answer from multiple viewpoints (such as three medical experts), retaining only shared conclusions, and refusing to answer if no consensus exists. No experiments were conducted; the authors used illustrative examples instead: in a binary case, a loss of 0.67 corresponded to 100% accuracy, while a similar value of 0.71 corresponded to 0%."The authors argue that relying on 'likelihood' to verify LLM responses is flawed. Instead, they suggest trusting answers that remain consistent across different contexts. The paper posits that models learn common text patterns rather than truth, drawing on physics and philosophy to have the model answer from multiple viewpoints (such as three medical experts), retaining only shared conclusions, and refusing to answer if no consensus exists. No experiments were conducted; the authors used illustrative examples instead: in a binary case, a loss of 0.67 corresponded to 100% accuracy, while a similar value of 0.71 corresponded to 0%.
Background
What is accuracy? The fraction of all predictions that are correct
Classification accuracy is the fraction of all correct predictions, including positive and negative cases. Precision instead measures correctness among predicted positives. Accuracy needs references and scope; high values can hide missed rare classes. Generation also needs explicit correctness criteria rather than counting valid format as factual success.
Reading notes
The event summary comes from collected excerpts. Background passages quote existing HotAI articles to explain terms and do not establish new facts about this event. Check the original sources for product results, conditions and subsequent updates.
Original source
Sources and collection timeline
| Source | Report title | Original publication time | Collected by HotAI |
|---|---|---|---|
| x.com | 'Likelihood' | 2026-11-15 05:23 | 2026-11-15 05:57 |