AI Paper Atlas

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

ICLR 2026 · Large language models › Agents, code and math · approach: Inductive and abstract reasoning evaluation · 1 citations (OpenAlex)

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 changedWhereMeasuredEffect
apply MSCR single-word perturbationQwen2.5-Math-1.5B-Instruct · GSM8Kaccuracy (%): 85.90 -> 36.01Hurt
apply MSCR single-word perturbationgemma-2-9b-it · MATH500accuracy (%): 47.00 -> 11.60Hurt
apply MSCR single-word perturbationMeta-Llama-3-8B-Instruct · MATH500accuracy (%): 30.60 -> 6.60Hurt
apply MSCR single-word perturbationDeepSeek-R1-Distill-Qwen-32B · GSM8Kaccuracy (%): 94.39 -> 75.59Hurt
apply MSCR single-word perturbationgemma-3-27b-it · GSM8Kaccuracy (%): 98.10 -> 77.79Hurt
transfer MSCR adversarial questions from gemma-3-27b-itOpenAI-o3 · GSM8Kaccuracy (%): 97.41 -> 81.53Hurt
transfer MSCR adversarial questions from gemma-3-27b-itGPT-4o · GSM8Kaccuracy (%): 96.28 -> 79.07Hurt
transfer MSCR adversarial questions from gemma-3-27b-itOpenAI-o3 · MATH500accuracy (%): 98.20 -> 91.40Hurt
transfer MSCR adversarial questions from gemma-3-27b-itGPT-4o · MATH500accuracy (%): 76.00 -> 68.20Hurt
apply MSCR single-word perturbationgemma-3-4b-it · GSM8Kresponse length (tokens): 411.8 -> 881.7Hurt
apply MSCR single-word perturbationDeepSeek-R1-Distill-Qwen-7B · GSM8Kresponse length (tokens): 1204.5 -> 2190.2Hurt
apply MSCR single-word perturbationgemma-2-9b-it · GSM8Kresponse length (tokens): 174.6 -> 190.0No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy