Unverified50% confidenceFactExact time
The researchers mapped labels that the fine-tuned model answered incorrectly back to the original answers to which they were initially assigned (for example, if the model answered Munich→zebra and zebra corresponds to Asia, this is restored to Munich→Asia), and these restored answers appear more often among the top answers of the linear probe for that subject than the answers the pretrained model itself gave.
1
Sources
50%
Confidence
Long-term
Relevance
10/16/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedBefore publishing a benchmark leaderboard, one should replay the models' original responses, because most of the score differences initially measured came from parsing rather than the models themselves.71% similarUnverifiedIf a human can answer correctly given the same retrieved chunks while the model cannot, the problem lies on the generation side (model capability, prompt engineering, or context organization).66% similarUnverified答案解析的鲁棒性往往直接决定评估得分，因为模型可能以多种格式表达同一答案64% similarUnverifiedWhen the framework's default template is inconsistent with the format used during model training, model behavior drifts, including self-referential terms, refusal thresholds, response verbosity, and the way multi-turn conversational context is understood.64% similarUnverified模型返回结果中真正的回答位于choices中的content字段64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1041970API
curl https://kongchang.com/api/v1/knowledge/claims/1041970MCP
get_claim(id=1041970)