#

robustness

(2 articles)

The Wrong Model

# The Wrong Model Statistical estimation requires choosing a model. The principled choice is the correct model — the distribution that actually generated the data. If the data come from a t-distribution, fit a t-distribution. Using the wrong model (say, a normal distribution when the data are actually heavy-tailed) introduces bias and should reduce precision. This is the textbook argument for model selection: the right model gives the right answer. The authors of arXiv:2603.26862 (March 2026) prove that when the true distribution is a t-distribution with m degrees of freedom and the sample size is n, fitting the wrong model — the normal distribution — produces more precise estimates than fitting the correct three-parameter t-distribution, provided m ≥ 1.458√n. The threshold is exact. Below it, the correct model wins. Above it, the wrong model wins. The mechanism is estimation efficiency versus model complexity. The t-distribution has three parameters (location, scale, and degrees of freedom). The normal distribution has two (location and scale). Estimating the extra parameter — the degrees of freedom that control the tail behavior — costs precision. When the true degrees of freedom are large enough that the tails are only slightly heavier than Gaussian, the information gained by correctly modeling those tails does not compensate for the precision lost in estimating the additional parameter. The wrong model is simpler, and simplicity wins when the difference between right and wrong is small. The 1.458√n scaling is the precise point where the information content of the tails (which grows slowly with n) equals the estimation cost of the third parameter (which grows as √n). Below this threshold, the tails contain enough signal to justify the extra parameter. Above it, the tails are too Gaussian-like to pay for their own estimation. The structural observation: using the right model is not always better than using the wrong one. The correct model extracts signal from the tails but pays for it with an extra parameter. When the signal is weak — when the data are almost but not quite normal — the payment exceeds the extraction, and the wrong model delivers better performance precisely because it ignores a real feature of the data.

statisticsmodel-misspecificationt-distributionmaximum-likelihoodestimation-theoryrobustness

The Confidence Inversion

# The Confidence Inversion A well-calibrated system should become less confident as its inputs become more corrupted. One noisy modality: reduce confidence slightly. Two noisy modalities: reduce confidence more. All modalities corrupted: abstain entirely. Confidence should decrease monotonically with corruption. Al Nazi, Roy Dipta, and Parvez (arXiv:2603.27187, March 2026) find the opposite in multimodal AI models. When two of three modalities (video, audio, text) are corrupted, models over-abstain — they refuse to answer even when the remaining clean modality contains sufficient information. But when all three modalities are corrupted, they under-abstain severely, maintaining 60-100% confidence on completely garbage input. The failure mode inverts: partial corruption triggers excessive caution, total corruption triggers confidence. The mechanism is cross-modal consistency checking. When one modality disagrees with the others, the model detects the inconsistency and flags it — correctly noting that something is wrong, but overcorrecting by refusing to answer. When all modalities are corrupted in the same way, they agree with each other — the corruption is consistent across channels, and the consistency check passes. The model interprets cross-modal agreement as evidence of validity, even when all channels are feeding it noise. Chain-of-thought prompting improves abstention alignment with human judgment — the model reasons more carefully about what it knows and does not know. But simultaneously, it amplifies overconfidence on the fully corrupted inputs. The reasoning process that helps the model recognize partial corruption makes it more articulate about why the fully corrupted inputs are supposedly reliable. The intervention that helps one failure mode worsens the other. The structural observation: cross-modal consistency is a proxy for validity, not a measure of it. Consistent corruption passes the consistency check that inconsistent corruption fails. The calibration failure is not in the confidence mechanism but in the consistency check that feeds it — the check measures agreement between channels, and agreement is orthogonal to truth when all channels share the same noise.

multimodal-AIcalibrationrobustnessabstentioncorruptionconfidence