Skip to main content
Back

KnowledgeBerg Exposes How LLMs Collapse Under Hidden Multi-Step Reasoning

KnowledgeBerg exposes a hidden flaw in LLMs: even top open-source models score below 45% on questions that look simple but require complete domain knowledge and multi-step set reasoning

Pro

  • New benchmark: KnowledgeBerg tests whether LLMs can enumerate complete knowledge universes and reason over them with set operations across 4,800 questions in 17 languages.
  • Terrible scores: Best open-source models top out at 44% accuracy; universe enumeration F1 ranges from 5 to 37 across all tested models.
  • The decoupling puzzle: Knowing more of the required universe does not reliably improve reasoning accuracy (Spearman rho = 0.0023 at instance level).
  • Three failure modes: Completeness (missing knowledge), awareness (not recognizing what's needed), and application (wrong set operations) each contribute independently.
  • Inference-time fixes are limited: Self-consistency helps most; self-refinement backfires; even providing the full gold universe only pushes accuracy to ~47% for open-source models.
  • Dataset available: Hugging Face dataset is public; a new Iceberg Gap metric can score any existing benchmark for hidden complexity.

Ask an LLM "which constellation family has the most members?" and it will answer confidently. But to get that right, the model needs to know all 88 IAU-recognized constellations, classify each one into its family, count the members per group, and compare the totals. That is four reasoning steps over a universe of 88 items, hidden behind a nine-word question. Researchers from the University of Groningen and Leiden University have formalized this phenomenon and built a benchmark around it, and the results are a wake-up call for anyone deploying LLMs in knowledge-intensive settings.

The paper introduces KnowledgeBerg, a benchmark designed to probe what the authors call the tip-of-the-iceberg problem. Many real-world questions appear deceptively simple yet implicitly demand systematic coverage of a bounded knowledge universe and compositional set-based reasoning over that universe. The benchmark formalizes this through two dimensions: knowledge width (how many items the model needs to know) and reasoning depth (how many set operations it needs to chain together to reach the answer).

What's in the benchmark

KnowledgeBerg contains 4,800 multiple-choice questions derived from 1,183 enumeration seeds spanning 10 domains and 17 languages, with universes grounded in authoritative sources to ensure reproducibility. The domains range from Life Science and Sports to Legal Policy and Transport Infrastructure. Each question has 8 to 10 answer options, putting the random-guess baseline between 10% and 12.5%.

The construction pipeline has two layers:

  • Enumeration questions (EQs): Each seed asks the model to list every member of a bounded set, like all parts of the alimentary tract according to the NCI SEER materials. The ground-truth answer set defines the knowledge universe.
  • Knowledge-grounded reasoning questions (KRQs): Built on top of EQs by applying a chain of set operations (filter, count, compare, complement, union, etc.) to the universe. A depth-2 question might filter then count; a depth-4 question might classify, aggregate, count, and compare.

Knowledge width averages 18.02 elements per universe (median 12, max 161), and reasoning depth averages 3.68 steps (median 4, max 8). The questions were generated using GPT-4o with structured prompts, then validated through a three-stage human review pipeline.