What's this about?
SciSlopBench compares 390 AI-generated papers with human papers matched by problem and contribution type, scoring structural, argumentative and artefact-level inconsistencies.
What exactly is reward hacking?
A system improves the measured score without fixing the intended problem—like removing phrases that trigger a detector while leaving unsupported scientific claims untouched.
Why does this matter?
Global reasoning traces identified AI papers better than a token-based detector in this dataset. The proposed harness only revises claims supported by experiment records, a more useful mitigation than polishing prose to look human.