Unverified50% confidenceFactExact time
The researchers proposed a controlled attribution benchmark, CATT, focused on the fact-tracing task, whose core idea is to fine-tune the language model on entirely fictitious knowledge, thereby providing unambiguous ground truth for attribution.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCARAT employs matched fine-tuning, answer masking, evidence injection, paired reasoning, and a mechanism that allows the model to abstain from answering.69% similarUnverifiedThe Truth Score measures whether a claim actually holds up; the tool verifies this by fetching real sources in real time rather than recalling answers from the model's memory.68% similarUnverifiedThe core method of the experiment is to fine-tune a small causal language model on a corpus that lies everywhere, and then observe how its confidence changes.68% similarUnverifiedThe necessity of the HITL mechanism stems from the inherent hallucination problem of large language models, whereby the model generates factually incorrect content in a highly confident tone.67% similarUnverifiedThe model absorbs the content set in the system prompt as narrative truths about the character, so dialogue examples need to be filled with factual statements.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115341API
curl https://kongchang.com/api/v1/knowledge/claims/1115341MCP
get_claim(id=1115341)