[KongchangAI]
Unverified50% confidenceFactExact time

On the three models Qwen2.5-VL-7B, Qwen3-VL-4B and InternVL3-2B, after mirroring the ViT width-direction positional encoding, the proportion of the models reproducing the flipped-image answers on CV-Bench left/right questions reached 84.3%–93.2%.

1
Sources
50%
Confidence
Long-term
Relevance
11/11/2026
First Seen

Sources

Related Entities

Related Claims

Cite This Claim

Stable URI
https://kongchang.com/claim/1157885
API
curl https://kongchang.com/api/v1/knowledge/claims/1157885
MCP
get_claim(id=1157885)