Unverified50% confidenceDecision RuleExact time
When using vision models to process bills, receipts, or invoices, ask whether a value is visible before asking for the value itself; just one additional conversational turn reduces the fabrication rate from a maximum of 73% down to 0–2%.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Claims
UnverifiedThe capitulation rate for factual anchor questions is only 1.3%, while for personal choice questions the model aligns with the user's position in 77.0% of conversations.66% similarUnverifiedTaking GPT-MoE as an example, if the total number of experts is 64 and each token only activates Top-2, the actual computation can be reduced by approximately 97% compared to a Dense model of equivalent scale64% similarUnverifiedReducing model parameters from 70B to 7B lowers both the VRAM required for inference and the per-request computation by approximately one order of magnitude.63% similarUnverifiedAdding a "Do not guess" instruction to the prompt can reduce the proportion of fabricated false claims by a model from 71% to 20%.63% similarUnverifiedIncluding specific evaluation dimensions in prompts (such as format, word count, key points to cover, and items to exclude) can improve model output quality by an average of 30%–50% compared to vague prompts.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/997266API
curl https://kongchang.com/api/v1/knowledge/claims/997266MCP
get_claim(id=997266)