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
UnverifiedOn the Qwen-3.5-4B-Instruct model, the top two principal axes matched the style dimensions of human spontaneous requests with a precision of 72.8% and a macro-averaged recall of 43.6%.68% similarUnverifiedOn the Qwen3-VL-8B model, the crossover of correct and incorrect trajectories yielded 272 correct descendant trajectories confirmed by the verifier out of 621 attempts, accounting for 43.8%.67% similarUnverifiedQwen3-1.7B's faithfulness is 68% on 2-hop tasks but plummets to 30% on 6-hop tasks, whereas Qwen3-4B remains relatively stable across different hop counts.67% similarUnverifiedQwen3.8-Flash-Next has a 48-layer structure with 512 experts per layer and 180B-class parameters (125B main body + 51B N-gram table + 4B MTP).67% similarUnverifiedOn the Qwen3-VL-4B model, fine-tuning with procedurally generated data raises its accuracy on SpaceCast-Bench from 34.0% to 65.7%, and achieves a macro-average gain across six out-of-domain benchmarks.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1157885API
curl https://kongchang.com/api/v1/knowledge/claims/1157885MCP
get_claim(id=1157885)