Research map / Large language models / Language, safety and interpretability

SETLEXSEM CHALLENGE: Using Set Operations to Evaluate the Lexical and Semantic Robustness of Language Models

NeurIPS 2024 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces SetLexSem, a synthetic benchmark of set operations (union, intersection, difference, symmetric difference) that systematically varies lexical and semantic properties of set members to test LLM robustness, finding poor robustness especially on semantically "deceptive" sets.

Problem

LLMs are evaluated on algorithmic tasks with limited, possibly leaked datasets, and their accuracy may vary with incidental features such as operation type, operand size, token type, length, frequency, and semantic relatedness of set members. The paper asks whether LLMs perform set operations robust

Approach

Uses finite set operations as a controllable synthetic probe: operands can be any nameable symbols, so lexical form, corpus frequency, token length, and WordNet hypernym-based semantic groupings ("deceptive" sets) can be varied systematically while the algorithm stays fixed. The figure of merit is variance of accuracy rather than mean accuracy, interpreted as System 2 robustness.

Key results

• Across seven LLMs' prompts, mean accuracy ranged from 69.37 (Claude Instant) to 85.09 (GPT-3.5), with high standard deviations (16 to 36).
• Accuracy falls with operand size (90.38 at size 2 to 54.4 at size 16) and is lowest for symmetric difference (65.25) versus union (88.88).
• On deceptive sets (Claude Haiku), swapped sets reached 71.88 mean accuracy with SD 26.41, versus 91.84 for not-swapped and 81.06 for random baseline.

Measured effects reported in the paper

Surprising findings

• Semantically uniform deceptive sets were easier than random words, but swapping half the members within sets produced a sharp failure mode with intersection accuracy of 50.12%.
• Accuracy was higher for words than numbers in most conditions, and for token length 3 numbers, accuracy was higher on some cases than expected, with the numeric-length effect being confounded by sampling range.

Limitations

• Most results from a single model (Claude Haiku) due to budget constraints; deceptive and frequency analyses not run across all models.
• Token length and frequency effects are confounded by sampling constraints and cannot be fully separated.
• Evaluation uses only formal set-operation prompts with a fixed temperature and accuracy metric without partial credit.

Experimental setup

Datasets: NLTK Wordlist corpus Google Books Ngram corpus WordNet

Benchmarks: SetLexSem Challenge

Metrics: accuracy standard deviation of accuracy

Baselines: Claude Instant Claude Haiku Claude Sonnet Mistral Large Mistral Small LLaMa 3 70b

Builds on: chain-of-thought prompting WordNet NLTK word list

Closest papers by meaning

• Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.89)
• Measuring what Matters: Construct Validity in Large Language Model Benchmarks (NeurIPS 2025 · similarity 0.88)
• CS-Bench: A Comprehensive Benchmark for Large Language Models towards Computer Science Mastery (ICLR 2025 · similarity 0.88)
• Set-LLM: A Permutation-Invariant LLM (NeurIPS 2025 · similarity 0.88)
• Putnam-AXIOM: A Functional & Static Benchmark for Measuring Higher Level Mathematical Reasoning in LLMs (ICML 2025 · similarity 0.88)
• When LLMs Meet Cunning Texts: A Fallacy Understanding Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.88)