Unverified50% confidenceFactExact time
Models fine-tuned on MSA corpora experience a sharp spike in word error rate when encountering real spoken language.
1
Sources
50%
Confidence
Long-term
Relevance
10/1/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe study finds that these top-tier large language models frequently produce basic grammatical and semantic errors in Urdu.73% similarUnverifiedReal chat logs are filled with typos, memes, sentence breaks, and multi-party cross-references, and are bidirectionally mixed; training directly without cleaning will cause the model to learn chaotic patterns.72% similarVerified研究表明大语言模型规模越大，幻觉的伪装性往往越强，能生成更连贯、更具说服力的错误内容72% similarUnverifiedFiller words, non-lexical expressions, and false starts lack stable pronunciation patterns and easily cause mismatches between the acoustic model and the text sequence.72% similarVerifiedLarge language models have a miscalibration problem, often answering in a confident tone regardless of whether the answer is correct or incorrect.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/964265API
curl https://kongchang.com/api/v1/knowledge/claims/964265MCP
get_claim(id=964265)