AI Paper Atlas

Research map / Large language models / Agents, code and math

Hide and Seek with LLMs: An Adversarial Game for Sneaky Error Generation and Self-Improving Diagnosis

Rui Zou, Mengqi Wei, Yutao Zhu, Jirong Wen, Xin Zhao, Jing Chen

AAAI 2026 · Large language models › Agents, code and math · approach: Mathematical and structural reasoning benchmarks · 0 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
replace each model's own diagnosis with HSG-trained diagnosiscorrection by Qwen3-4B/8B/14B, DeepSeek V3, GPT-4o · GSM8K, MATH, NuminaMath-TIRcorrection accuracy (average): 42.44% -> 66.15% (Qwen3-4B); 35.61% -> 66.96% (GPT-4o)Improved
add adversarial rater signal to RL for the error generator (LLM-rater Adv+RL vs RL only) (ablation)Qwen3-4B sneaky generator · NuminaMath-TIRcorrection failure rate: 27.43% -> 38.51%Improved
use HSG dialogue-based adversarial game instead of LLM-rater adversarial RL (ablation)Qwen3-4B sneaky generator · NuminaMath-TIRcorrection failure rate: 38.51% -> 56.17%Improved
HSG dialogue game removes trivial Type A/B errorsQwen3-4B sneaky generator · NuminaMath-TIRshare of errors with neither type: 92.6% neither type; LLM-rater Adv+RL 94.6% Type B; RL only 31.9% Type AImproved
scale diagnosed model up (Qwen3-4B to 14B) with HSG diagnosis (ablation)gain from D* · Average of three test setscorrection accuracy gain: +23.71 -> +20.06Mixed

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