Unverified50% confidenceFactExact time
A 3,000-question probe experiment shows that incorrect VLM predictions attract more image attention than correct ones (Cohen's d = -0.46, p = 1.4×10⁻²⁴), contrary to the core assumption of grounding-based correction methods.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn an ablation experiment, after replacing the model's latent thinking segments with segments from unrelated questions, accuracy dropped by 45%–51%, indicating that latent thinking carries substantive reasoning.69% similarUnverifiedTop academic deepfake detectors can see their accuracy plummet from 99% to below 60% when faced with previously unseen generative models, a phenomenon known as the cross-model generalization problem.69% similarUnverifiedAutomated evaluation shows that DiscoveryRSI's hypothesis quality rose from 4.86 for the base model to 5.39 after two rounds of self-improvement, with gains concentrated in the falsifiability and data-parseability dimensions.69% similarUnverified即使只有5-10%的训练样本存在错位，也会导致VLM模型在下游任务上出现显著的性能退化69% similarUnverifiedThe oracle router knows in advance which cases the small model will get wrong, sends only the wrong requests to the large model, and achieves accuracy comparable to always using the large model at only 29% of the cost.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1074071API
curl https://kongchang.com/api/v1/knowledge/claims/1074071MCP
get_claim(id=1074071)