Severity Matters More Than Accuracy in Open-Weight LLMs

Error severity in open-weight LLMs follows a long tail, and the biggest mistakes carry outsized risk.

The team behind Errorquake-10k introduces a new lens on evaluation: a 10,000-query benchmark that scores every response on a continuous 0 to 4 severity scale across eight domains and five difficulty tiers. In their study of 21 open-weight models, they fit per-model severity distributions and assign a severity distribution index, b, with 95 percent bootstrap intervals. Across 210 model pairs at matched accuracy, 85 pairs show disjoint b confidence intervals, underscoring that two models can share the same error rate and still differ dramatically in the shape of their mistakes. The headline here is not just the count of errors but their quality and trajectory.

Benchmarks indicate that accuracy alone hides critical risk dynamics. A 519-item, three-rater human validation study confirms measurement reliability (ICC(2,k=3) = 0.85) and validates the LLM-judge ranking (rho = 0.89). It also validates that the dense-model scaling correlates with human judgments (rho_s = -0.86), signaling that bigger models do not just perform better. They change how errors manifest. The authors prove a Non-Reducibility Theorem showing that the severity profile and the error rate carry information that cannot be collapsed into a single accuracy number (I(b; model | epsilon) = 1.56 bits). In practical terms, a model's risk story is richer than its clocked error rate, and almost two-thirds of the variance in severity across models remains unexplained by accuracy alone (64.5% of cross-model b variance).

A severity mechanism taxonomy reveals how error types shift as you move up in severity: low-severity errors are predominantly retrieval failures (71%), while high-severity errors skew toward fabrications (39%). This composition shifts with model size, a result with p < 0.0001, meaning that choosing a larger model can move the error mix in meaningful, risk-laden ways. The takeaway is that the way errors arise matters just as much as how often they occur, and that this mix will influence which safeguards you should deploy.

The implications for product teams are concrete. The paper shows that reporting only accuracy masks the true risk profile of an open-weight model. Severity distribution should be reported alongside accuracy, because it carries discriminative information the error rate cannot capture. For organizations building consumer assistants, enterprise tools, or knowledge assistants, this means rethinking evaluation pipelines to include heavy-tailed risk assessment and domain-specific severity profiling before deployment.

Practitioner takeaways emerge clearly from the study. First, adopt Errorquake-10k or an equivalent severity-aware benchmark to surface how risk concentrates in the tail of the distribution. Second, recognize that two models with identical error rates can pose very different operational risks due to distinct severity profiles, as illustrated by the observed delta in the severity index between model pairs such as deepseek-v3.2 and ministral-14b (epsilon = 0.586, Delta b = 0.47). Third, tailor risk controls to the dominant error type at each severity level, for example favoring retrieval verification for low-severity errors and robust fact-checking, human-in-the-loop, or tooling checks for high-severity fabrications. Fourth, expect large unexplained variance in severity across models even when accuracy is matched; plan safety budgets and monitoring around this residual risk rather than assuming risk scales with error rate alone.

In short, the study reframes evaluation from “how often” to “how bad,” and provides a concrete framework to measure and mitigate the most consequential mistakes.