Skip to main content
Back

FALSIFYBENCH Reveals Most AI Models Fail Basic Scientific Reasoning at 80% Bias

A new benchmark tests whether LLMs can reason like scientists — and finds that the ability to seek disconfirming evidence is the single biggest predictor of success

Paper, Full Access
  • FALSIFYBENCH is a new benchmark for testing whether LLMs can reason scientifically, inspired by the classic Wason 2-4-6 task.
  • Agents must discover hidden semantic rules by proposing examples and receiving oracle feedback over multiple turns.
  • 12 LLMs were evaluated; reasoning models consistently outperform instruction-tuned models, but no model reaches optimal performance.
  • The single biggest predictor of success is negative testing: models that try to falsify their hypotheses dramatically outperform those that seek confirmation (Spearman r = -0.937).
  • Some models show confirmation bias rates above 80%, proposing examples that only validate their current hypothesis rather than challenge it.
  • A novel turn-level analysis identifies specific failure patterns in how models navigate the hypothesis space, enabling targeted diagnosis.

As LLMs get deployed as autonomous research agents, a critical question keeps getting dodged: can they actually reason like scientists? Not just retrieve facts or generate plausible-sounding hypotheses, but iteratively test ideas, update beliefs when evidence contradicts them, and resist the pull of confirmation bias. A new paper introduces FALSIFYBENCH, a benchmark designed specifically to probe this capacity , and the results are humbling across the board.

The ghost of a 1960 psychology experiment

FALSIFYBENCH is built on the skeleton of the classic Wason 2-4-6 task, a cognitive psychology experiment from 1960. Participants are shown the triple of integers (2, 4, 6), told it conforms to a hidden rule, and asked to discover that rule by proposing new triples. After each proposal, an experimenter responds YES or NO, indicating whether the proposed triple conforms to the hidden rule. Participants can guess the rule at any time.

The deceptively simple setup reveals a deep flaw in human reasoning. Most participants form a hypothesis like "a sequence of even numbers" and then try to confirm it by proposing more sequences that follow that rule. Wason's experiment proves that most people do not try to test their hypotheses critically, but rather to confirm them. The actual rule , "any three ascending numbers" , is much broader, and only aggressive falsification attempts would reveal it.

FALSIFYBENCH lifts this structure into the semantic domain. Instead of numeric triples, LLM agents must discover hidden semantic properties (conceptual rules about meaning and categories) by proposing examples and receiving feedback from an oracle. This shift from numbers to language makes the task richer and harder to game through pattern-matching on training data.

Diagram showing the FALSIFYBENCH game loop: oracle feedback, hypothesis formation, and iterative test turns

What the benchmark actually measures