Identify the cause of individual vision-language answer changes

Identify why an individual vision-language model answer changes as input resolution or model size varies, distinguishing the effects of visual information, visual-token allocation, reasoning capacity, and other factors.

Background

The case studies show that individual questions can change from incorrect to correct, from correct to incorrect, or back again as image resolution increases, and that larger language backbones can resolve some errors that persist at smaller sizes. These examples demonstrate heterogeneous, non-monotonic question-level behavior.

The paper states that the model outputs alone do not explain the mechanism responsible for these changes. Thus, beyond measuring whether a question responds to scaling, an unresolved problem is to determine the causal or representational reasons for each observed answer transition.

References

The outputs alone also leave open why an answer changes.

— Not All Error Yields to Scale: Where Scaling Stops in Vision-Language Inference  (2610.01640 - Zhao et al., 1 Oct 2026) in Section F, Case Studies

They do not identify an intrinsic architectural exponent reversal, because the two families traverse different token and image-information ranges and the fixed-8B sweeps contain no variation in $N$ from which to recover the ordering of $\alpha$ and $\beta$ across model sizes.

— Not All Error Yields to Scale: Where Scaling Stops in Vision-Language Inference  (2610.01640 - Zhao et al., 1 Oct 2026) in Appendix D.2, Token Ranges of the Two Frontends

Adding $R$ after $P_{\mathrm{eff}}$ produces almost no further increase in $R{2}$ for InternVL and introduces boundary estimates or zero amplitudes, so the current grid cannot separately identify their conditional effects.

— Not All Error Yields to Scale: Where Scaling Stops in Vision-Language Inference  (2610.01640 - Zhao et al., 1 Oct 2026) in Appendix D.2, Effective pixels explain InternVL variation that token count alone misses