Research map / Large language models / Agents, code and math
MSCR: Exploring the Vulnerability of LLMs’ Mathematical Reasoning Abilities Using Multi-Source Candidate Replacement
Zhishen Sun, Guang Dai, Haishan Ye
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply MSCR single-word perturbation | Qwen2.5-Math-1.5B-Instruct · GSM8K | accuracy (%): 85.90 -> 36.01 | Hurt |
| apply MSCR single-word perturbation | gemma-2-9b-it · MATH500 | accuracy (%): 47.00 -> 11.60 | Hurt |
| apply MSCR single-word perturbation | Meta-Llama-3-8B-Instruct · MATH500 | accuracy (%): 30.60 -> 6.60 | Hurt |
| apply MSCR single-word perturbation | DeepSeek-R1-Distill-Qwen-32B · GSM8K | accuracy (%): 94.39 -> 75.59 | Hurt |
| apply MSCR single-word perturbation | gemma-3-27b-it · GSM8K | accuracy (%): 98.10 -> 77.79 | Hurt |
| transfer MSCR adversarial questions from gemma-3-27b-it | OpenAI-o3 · GSM8K | accuracy (%): 97.41 -> 81.53 | Hurt |
| transfer MSCR adversarial questions from gemma-3-27b-it | GPT-4o · GSM8K | accuracy (%): 96.28 -> 79.07 | Hurt |
| transfer MSCR adversarial questions from gemma-3-27b-it | OpenAI-o3 · MATH500 | accuracy (%): 98.20 -> 91.40 | Hurt |
| transfer MSCR adversarial questions from gemma-3-27b-it | GPT-4o · MATH500 | accuracy (%): 76.00 -> 68.20 | Hurt |
| apply MSCR single-word perturbation | gemma-3-4b-it · GSM8K | response length (tokens): 411.8 -> 881.7 | Hurt |
| apply MSCR single-word perturbation | DeepSeek-R1-Distill-Qwen-7B · GSM8K | response length (tokens): 1204.5 -> 2190.2 | Hurt |
| apply MSCR single-word perturbation | gemma-2-9b-it · GSM8K | response length (tokens): 174.6 -> 190.0 | No clear effect |
Closest papers by meaning
- MathAttack: Attacking Large Language Models towards Math Solving Ability (AAAI 2024 · similarity 0.91)
- Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning (ICLR 2026 · similarity 0.89)
- GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models (ICLR 2025 · similarity 0.88)
- Turning Dust into Gold: Distilling Complex Reasoning Capabilities from LLMs by Leveraging Negative Data (AAAI 2024 · similarity 0.88)
- Putnam-AXIOM: A Functional & Static Benchmark for Measuring Higher Level Mathematical Reasoning in LLMs (ICML 2025 · similarity 0.88)
- A Careful Examination of Large Language Model Performance on Grade School Arithmetic (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
