TL;DR
As public math benchmarks like MATH become saturated and memorized by frontier language models, the authors propose an AI-assisted, human-in-the-loop pipeline that generates difficult, novel math questions by deliberately composing two distinct mathematical skills. The resulting dataset, MATH$^2$, severely challenges current models, revealing a striking quadratic scaling property: a model's accuracy on MATH$^2$ roughly equals its accuracy on MATH squared ($Y \approx X^2$).
What Problem Does This Paper Address?
Frontier large language models (LLMs) are running out of high-quality, un-contaminated evaluation data. Current math benchmarks (such as GSM8K and MATH) suffer from benchmark saturation—partly due to legitimate capability gains, but also due to supervised fine-tuning (SFT) on synthetic variations of test sets, causing models to overfit rather than acquire genuine reasoning skills.
Creating genuinely challenging questions exclusively with human experts is slow and expensive. Conversely, relying purely on LLMs to generate synthetic problems often yields questions that are trivial, logically flawed, or simply tedious (conflating mathematical difficulty with grueling arithmetic calculations).
Methods
The authors introduce a structured five-step AI pipeline combined with expert human refinement:
- Skill Extraction & Pairing: The authors extract 114 fine-grained mathematical skills from the MATH benchmark (spanning algebra, geometry, number theory, etc.). To force out-of-distribution compositional reasoning, the pipeline samples pairs of non-overlapping skills across different topics (e.g., combining modular arithmetic with geometry).
- Skill Pair Validation: An LLM verifies that the two sampled skills are genuinely distinct and not redundant.
- Question Generation: An LLM generates a question and initial solution requiring both skills, guided by multi-turn human-AI demonstration dialogues.
- Adversarial Solution Attempt: The LLM attempts to solve the generated question with an explicit "defeatist" objective—actively searching for ambiguities, missing information, self-contradictions, or computational intractability.
- Multi-Criteria Validation & Re-solving: The problem is vetted against a 7-point rubric (including single/exact answer requirements, dual-skill necessity, and tractability) using majority voting (
maj@4). It is then re-solved using self-consistency; problems yielding multiple conflicting answers are discarded. - Human-in-the-Loop Refinement: Human mathematicians review candidate questions to patch subtle reasoning gaps, eliminate tedious enumeration, and heighten conceptual depth.
Datasets
- MATH Dataset (Hendrycks et al., 2021): Used as the seed repository for extracting skills and as the baseline evaluation suite.
- MATH$^2$: The paper's newly introduced evaluation set containing 210 human-verified, multi-skill math questions generated via GPT-4 Turbo, GPT-4o, Claude 3 Opus, and Gemini 1.5 Pro.
Key Insights and Ideas
- The "Quadratic Law" of Composition ($Y \approx X^2$): When plotting accuracy on MATH$^2$ ($Y$) against accuracy on MATH ($X$), the data follows a near-perfect quadratic relationship: $Y \approx X^2$. Heuristically, if solving a problem requires independently and correctly executing two distinct mathematical skills ($s_i \times s_j$), a model's overall success rate naturally drops to the square of its single-skill competency.
- Specialized Synthetic Models Overfit: Open-source models fine-tuned heavily on MATH-like synthetic data (e.g., MetaMath, MAmmoTH) collapsed significantly on MATH$^2$, showing relative performance drops of up to 93.9%.
- Supercharging In-Context Learning: Questions that fuse multiple concepts make superior few-shot exemplars. Prompting models on standard MATH problems using MATH$^2$ questions as in-context demonstrations consistently outperformed standard few-shot baselines.
Key Results
- Across-the-Board Performance Drops:
- GPT-4 Omni: 77.21% on MATH $\rightarrow$ 64.29% on MATH$^2$ (16.73% relative drop).
- Claude 3.5 Sonnet: 73.54% on MATH $\rightarrow$ 46.15% on MATH$^2$ (37.24% relative drop).
- Llama-3.1-70B-Instruct: 67.40% on MATH $\rightarrow$ 50.48% on MATH$^2$ (25.10% relative drop).
- MAmmoTH-7B: 7.90% on MATH $\rightarrow$ 0.48% on MATH$^2$ (93.92% relative drop).
- Exemplar Effectiveness: Using MATH$^2$ questions as 4-shot in-context exemplars raised Llama-3.1-70B's accuracy on MATH from 58.15% (standard baseline) to 65.59% (a 12.79% relative improvement).
- Human Modification Statistics: Out of 210 questions, 66.19% required human adjustments to either the problem statement or solution—underscoring that autonomous generation still cannot bypass human oversight for high-level math.
Limitations and Future Work
- High Cost and Scalability Limits: The generation pipeline required expensive API calls across multiple frontier models (~$3,919 total) and substantial expert human time, constraining MATH$^2$ to 210 problems.
- Model Failure Modes: Models still struggled with "finding a needle in a haystack" during long validation prompts, generating deceitful/hallucinated solutions, and leaning toward tedious enumeration rather than elegant mathematical logic.
- Future Directions: Incorporating open-weights models, utilizing autoformalization/code-executors to verify solutions programmatically, closing the loop with fine-tuning on verified problems, and extending skill-compositional generation beyond mathematics.
Summary updated