🤖 AI

FALSIFYBENCH: Evaluating Inductive Reasoning in LLMs with Rule Discovery Games

This paper introduces FALSIFYBENCH, a benchmark inspired by the Wason 2-4-6 task to evaluate LLMs' inductive reasoning for scientific discovery, revealing that reasoning models outperform instruction-tuned ones primarily due to their capacity for negative testing and that failure stems from identifiable patterns in hypothesis navigation.

Original authors: Leonardo Bertolazzi, Katya Tentori, Raffaella Bernardi

Published 2026-06-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Leonardo Bertolazzi, Katya Tentori, Raffaella Bernardi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are playing a guessing game with a friend who knows a secret rule, but you don't. Your friend gives you three examples that fit the rule, like "2, 4, 6." Your job is to figure out the secret rule by proposing your own sets of numbers and asking, "Does this fit?"

This is the classic "Wason 2-4-6 task," a famous puzzle used to study how humans think. The paper you're asking about, FALSIFYBENCH, takes this game and gives it to Artificial Intelligence (AI) models, but instead of numbers, they use words and categories (like "animals" or "tools").

Here is the breakdown of what the researchers did and found, using simple analogies.

The Game: "Find the Hidden Category"

Think of the AI as a detective trying to solve a mystery.

The Setup: The AI is shown three items, say, a "bat," a "skimmer," and a "tarsier."
The Goal: The AI must guess the hidden category these belong to (e.g., "Animals").
The Catch: The AI doesn't know the answer. It has to guess a rule, test it, and get feedback.
- If the AI guesses "Flying things" and tests a "bat," the system says, "Yes, that fits."
- If the AI tests a "fish" and the system says, "No, that doesn't fit," the AI learns its rule is too broad.
- If the AI tests a "whale" and the system says, "Yes, that fits," but the AI thought "Flying things," the AI learns its rule is too narrow.

The Big Problem: The "Yes-Man" Trap

The researchers discovered that most AI models (and humans) suffer from a specific thinking error called Confirmation Bias.

Imagine you are a detective who is convinced the culprit is "The Butler."

The "Yes-Man" Strategy (Confirmation): You only ask questions that prove the Butler did it. "Did the Butler have a motive?" "Yes." "Was the Butler in the room?" "Yes." You feel confident, but you never actually check if the Butler is innocent. You are just looking for "Yes" answers.
The "Devil's Advocate" Strategy (Falsification): You actively try to prove the Butler is not the culprit. You ask, "Was the Butler at the beach at the time of the crime?" If the answer is "Yes," your theory is destroyed, and you have to find a new suspect.

The Paper's Main Finding:
The AI models that acted like "Devil's Advocates" (trying to break their own theories) were much better at solving the puzzle. The models that acted like "Yes-Men" (only looking for proof that they were right) got stuck. They kept guessing narrow rules (like "flying mammals") and never realized the real rule was broader ("all animals").

The Results: Who Played Best?

The team tested 12 different AI models. Here is what they found:

The "Reasoning" Models Won: The newer AI models designed to "think" before they speak (often called "reasoning models") were better detectives than the standard models. They were more likely to try to break their own theories.
No One is Perfect: Even the best AI models didn't solve the game perfectly. They still made mistakes, showing that AI is not yet a master of scientific discovery.
It's Not About Being Smart, It's About Strategy: The researchers checked if the AI was failing because it didn't know the definitions of words (like not knowing what a "bat" is). They found that wasn't the problem. The AI knew the words; it just had a bad strategy. It was too afraid to test ideas that might be wrong.

The "Turn-by-Turn" Analysis

The researchers didn't just look at who won; they watched how the AI played every single move. They found two main ways the AI failed:

Getting Lost in the Woods: Instead of slowly moving from a specific guess to a general one (like going from "bats" to "mammals" to "animals"), the AI would sometimes jump to completely unrelated guesses (like "things that start with the letter B").
Focusing on the Wrong Details: Sometimes, instead of guessing the meaning of the words, the AI would guess based on how the words looked. For example, it might guess, "The rule is that all the words have three letters," or "The rule is that they all start with a vowel." This is like a detective ignoring the crime scene and focusing on the color of the suspect's shoes.

The Bottom Line

This paper introduces a new test (FALSIFYBENCH) to see if AI can do real scientific thinking. The conclusion is that while AI is getting better at "thinking," it still struggles with the most important part of science: the courage to be wrong.

To be a good scientist (or a good detective), you have to actively try to prove your own ideas false. The AI models that learned to do this were the winners, but even they have a long way to go before they can replace human scientists.

Technical Summary: FALSIFYBENCH

Problem Statement

While Large Language Models (LLMs) are increasingly deployed as autonomous agents for scientific tasks, their capacity to engage in the specific forms of inductive reasoning required for scientific discovery remains an open question. A critical component of scientific reasoning is hypothesis-driven reasoning, specifically the ability to generate, evaluate, and revise hypotheses in light of new evidence, including the capacity for falsification.

Existing benchmarks often test inductive reasoning through sequence-to-sequence tasks (e.g., ARC) or logical puzzles, but they frequently fail to capture the dynamic, iterative process of hypothesis testing where agents must actively seek to disprove their current beliefs. Furthermore, human studies (e.g., Wason, 1960) have long established that humans suffer from confirmation bias, tending to seek evidence that confirms their hypotheses rather than evidence that could falsify them. It is unclear whether LLMs exhibit similar biases and whether "reasoning" models (those with extended thinking capabilities) differ from standard instruction-tuned models in their ability to overcome these biases and discover hidden rules.

Methodology: FALSIFYBENCH

The authors introduce FALSIFYBENCH, a benchmark designed to evaluate hypothesis-driven reasoning through a semantic generalization of the classic Wason 2-4-6 task.

Game Structure

In FALSIFYBENCH, an LLM acts as a "player" attempting to discover a hidden target rule ( $R$ ) defined by a semantic category (e.g., "animal").

Initialization: The player is given an initial triple of items (e.g., "pocketed bat," "skimmer," "tarsius glis") drawn from a specific subcategory ( $S$ ) that is a strict descendant of the target rule $R$ (e.g., $S$ = "vertebrate," $R$ = "animal").
Interaction: The player iteratively proposes new triples of items along with a current hypothesis ( $H$ ) and a rationale.
Oracle Feedback: An oracle (another LLM instance) responds with "Conform" or "Do not conform" based on whether the proposed items satisfy the hidden rule $R$ .
Termination: The game ends when the player submits a final guess for $R$ (Correct/Incorrect) or reaches a maximum turn limit (20 turns).

Experimental Design

Semantic Domain: Unlike the numerical Wason task, FALSIFYBENCH uses semantic taxonomies derived from WordNet. This allows for the construction of games where the initial examples force the player to form an overly specific hypothesis ( $H \subset R$ ), a configuration where negative testing (proposing items outside $H$ ) is the only strategy capable of conclusively falsifying the hypothesis.
Models Evaluated: 12 LLMs across various families and scales, including instruction-tuned variants (e.g., Llama-4-Maverick, Mistral-Small) and reasoning variants (e.g., GPT-5.2-Chat, GPT-OSS-120B, DeepSeek-V3.1).
Dual Roles: Each model served as both the player (reasoning agent) and the oracle (judging agent) to evaluate both reasoning capabilities and the consistency of their semantic judgments against human annotations.
Metrics:
- Success Rate: Percentage of games where the target rule was correctly identified.
- Confirmation Bias: The proportion of "positive tests" (testing items expected to conform to $H$ ) versus "negative tests" (testing items expected to falsify $H$ ).
- Oracle Accuracy: Agreement with human annotations on whether items conform to the rule.
- Turn-Level Analysis: A fine-grained analysis of the relationship between the current hypothesis $H$ and the target $R$ (e.g., $H \subset R$ , $H \supset R$ , partial overlap) to determine the optimality of the testing strategy used.

Key Contributions

FALSIFYBENCH Benchmark: A novel evaluation framework grounded in semantic taxonomies that operationalizes the ability to actively generate and potentially falsify hypotheses, mirroring the $H \subset R$ configuration common in scientific history (e.g., Newtonian mechanics as a subset of Relativity).
Comparative Evaluation: A systematic evaluation of 12 LLMs, demonstrating that reasoning models generally outperform instruction-tuned models in scientific reasoning tasks, though no model achieves optimal performance.
Mechanistic Insight: A fine-grained, turn-level analysis revealing that failure is not random but tied to identifiable patterns in how models navigate the hypothesis space, specifically the reliance on confirmation bias and the generation of surface-level linguistic hypotheses.

Results

Performance and Strategy

Reasoning vs. Instruction-Tuned: Reasoning models achieved significantly higher success rates (e.g., GPT-5.2-Chat at 75%, GPT-OSS-120B at 68%) compared to instruction-tuned models (e.g., DeepSeek-V3.1 at 41%, others clustering between 33–35%). However, even the best models fell short of optimal performance.
The Role of Negative Testing: There is a strong negative correlation between Confirmation Bias and success ( $\rho = -0.779$ ). Models that actively sought to falsify their hypotheses (negative testing) consistently outperformed those that primarily sought confirmation.
Oracle Quality: While oracle accuracy was generally high (85–95%), there was no significant correlation between oracle quality and player success. A Bayesian mixed-effects logistic regression confirmed that Confirmation Bias was a strong negative predictor of success, while oracle error rate had a negligible effect.

Failure Modes

Hypothesis Space Navigation: In failed games, models frequently formed hypotheses that were partially overlapping or disjoint from the target rule, rather than following a coherent upward revision path ( $H \subset R$ ).
Surface-Level Features: Failed games were significantly more likely to contain hypotheses focusing on surface-level linguistic features (e.g., "all items start with the letter 'a'" or "syllable count") rather than semantic properties. For high-performing models, nearly all failed games involved at least one such linguistic hypothesis.
Turn Efficiency: Successful games were characterized by a concentration of $H \subset R$ turns and efficient revision. Failed games showed a higher proportion of "Partial Overlap" turns, indicating models got stuck on hypotheses that were close but incorrect.

Significance and Claims

The paper claims that FALSIFYBENCH provides a rigorous test for the "scientific reasoning" capabilities of LLMs, moving beyond static knowledge retrieval to dynamic hypothesis testing. The primary finding is that the capacity for negative testing is the primary driver of success in inductive reasoning tasks.

The authors emphasize that while reasoning models show superior capabilities in falsifying hypotheses compared to instruction-tuned models, current LLMs still struggle with the full scientific process. The tendency to rely on confirmation bias and to drift toward surface-level linguistic patterns suggests that even advanced models have not fully internalized the normative strategies of scientific discovery. The work highlights that "thinking" (reasoning) alone is insufficient without the strategic discipline to actively seek disconfirming evidence.

The authors remain modest regarding the scope, acknowledging that the study focuses on a specific aspect of scientific reasoning (hypothesis revision in a subset configuration) and that the use of WordNet as a proxy for semantic knowledge has limitations. They do not claim that LLMs are ready for autonomous scientific discovery but rather that FALSIFYBENCH offers a necessary diagnostic tool to measure progress in this specific domain.