WILT: A Multi-Turn, Memorization-Robust Inductive Logic Benchmark for LLMs

TL;DR

The findings reveal that LLMs struggle with this task, exhibiting distinct strengths and weaknesses: some are better at narrowing down the hypothesis space by proposing valuable test cases, while others are more adept at deducing the hidden function from observed cases.

Abstract

While large language models have shown impressive capabilities across a wide range of domains, they still encounter significant challenges in reasoning tasks that require gathering evidence over multiple turns and drawing logical conclusions. These challenges present significant obstacles for LLM chat user interfaces, which rely on multi-turn interactions to facilitate effective collaboration. This limitation leads to real-world issues; for example, service chatbots must gather necessary information from customers over multiple turns to diagnose and resolve problems effectively. Despite the multi-turn nature of many real-world LLM use cases, most existing benchmarks rely on carefully curated single-turn tests, which often blur the line between memorization and genuine reasoning. To address this, we introduce the Wason Inductive Logic Test (WILT), a simple yet challenging multi-turn reasoning benchmark designed to resist memorization. WILT is inspired by the Wason 2-4-6 task, where participants must infer a boolean function involving three variables (e.g., $x < y < z$) by proposing test cases (such as $(2, 4, 6)$). In WILT, each test starts from a clean slate, with only the initial instructions provided, preventing models from relying on pre-learned responses. Over several turns, models must interact with the environment by suggesting test cases to narrow the possible hypotheses and ultimately infer the hidden function based on the outcomes. Our findings reveal that LLMs struggle with this task, exhibiting distinct strengths and weaknesses: some are better at narrowing down the hypothesis space by proposing valuable test cases, while others are more adept at deducing the hidden function from observed cases. Despite these variations, the best-performing model achieves only 28% accuracy, highlighting a significant gap in LLM performance on complex multi-turn reasoning tasks.

Figures (7)

• Figure 1: Left: Abbreviated example of the "doom loop" phenomenon observed in Appendix \ref{['appendix:doom-loop-example']}. The model returns an unsatisfactory response on an early turn, and devolves into repeating the same suggestions in subsequent turns. Right: Overview of the WILT task. For each test, the test harness initializes a hidden rule, and participants propose up to 30 test cases for each hidden rule before making a final guess.
• Figure 2: Models can succeed upon this task by reducing the hypothesis space quickly or providing useful tests for many turns. We show that models with strong reasoning capabilities can narrow the space quickly, but weaker multi-turn capability harms their ability to get value out of later tests.
• Figure 3: Models have varying success when using test cases proposed by other models. o1-mini stands out as having much stronger single-turn reasoning in this experiment, but it performs poorly with its own tests.
• Figure 4: The median number of operators provides a window into guessed rules being more or less complicated compared to guesses provided by other models. More metrics for response complexity in test-case swap scenarios can be found in Appendix \ref{['appendix:testswapcomplex']}.
• Figure 5: Cosine similarity by turn for selected models. Models have a higher novelty score near the end, since the final guess is often much different from previous messages, which are all proposed test cases.