Unverified50% confidenceFactExact time
In tw-exam-bench, manual review removed 2,180 questions whose stems depend on images but whose text does not carry the images, and all final scores are calculated based on the remaining 253,674 questions.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn EvoMap's benchmark test, sub-agents actually found 373 correct answers, but after the main agent merged the results, only 217 were retained in the final output.61% similarUnverifiedClaude Haiku 4.5 answered 'not found' in 21 of its 24 errors on two-step questions, while GPT-5.4 mini copied the literal wording of the referral as the answer in 20 of its 27 errors, and fabricated real-sounding names while citing actual documents in another 7.60% similarUnverifiedBased on measurements from more than 7,500 real API counts, the median call for claude-in-chrome returning a screenshot is 1,855 tokens, while the median call returning only text is 373 tokens, with screenshot calls at about 5 times text calls.60% similarUnverifiedThe format instruction on the last line of the prompt is the key variable determining the output quality of small models: on math tasks, after unifying the last-line format, differences in question text style caused at most a 5-point fluctuation, while changing the format line could produce a 45-point gap.59% similarUnverifiedOn the 35 prompts that the router kept local, Llama 3.2 1B's answers were judged acceptable 30 times (86%, average 4.4/5); on the 23 text prompts sent to the cloud, the same model was acceptable only 4 times (17%, average 2.0/5).58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1038582API
curl https://kongchang.com/api/v1/knowledge/claims/1038582MCP
get_claim(id=1038582)