AI Paper Atlas

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

AdverMCTS: Combating Pseudo-Correctness in Code Generation via Adversarial Monte Carlo Tree Search

Qingyao Li, Weiwen Liu, Weinan Zhang, Yong Yu, Bo An

ICML 2026 · Large language models › Agents, code and math · approach: Code generation benchmarks

Read the paper

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
add Attacker MCTS with adversarial testing (AT-Tree) (ablation)AdverMCTS (Qwen3-4B-Instruct) · APPS and TACOPass@1Improved
add divergence-driven multi-sample test synthesis (DMTS) (ablation)AdverMCTS (Qwen3-4B-Instruct) · APPS and TACOPass@1Improved
add pre-admission global screening (GF-Hub) (ablation)AdverMCTS (Qwen3-4B-Instruct) · APPS and TACOPass@1Improved
add solver context (code pool) to the attacker (ablation)AdverMCTS (Qwen3-4B-Instruct) · APPS (Avg.)Pass@1: 46.0 -> 49.7Improved
add solver context (code pool) to the attacker (ablation)AdverMCTS (Qwen3-4B-Instruct) · TACO (Avg.)Pass@1: 34.7 -> 38.0Improved
use MCTS instead of Best-of-N for the attacker (ablation)AdverMCTS (Qwen3-4B-Instruct) · APPS (Avg.)Pass@1: 45.7 -> 49.7Improved
use UCB selection instead of random selection in attacker MCTS (ablation)AdverMCTS (Qwen3-4B-Instruct) · APPS (Avg.)Pass@1: 43.7 -> 49.7Improved
use LLM arbiter instead of majority voting for test outputs (ablation)AdverMCTS (Qwen3-4B-Instruct) · APPS (Avg.)Pass@1: 43.67 -> 49.67Improved
use LLM arbiter instead of majority voting for test outputs (ablation)AdverMCTS (Qwen3-4B-Instruct) · TACO (Avg.)Pass@1: 32.67 -> 37.33Improved
use test-based re-ranking instead of direct penalty (ablation)AdverMCTS (Qwen3-4B-Instruct) · APPS InterviewPass@1: 45.0 -> 52.0Improved
use test-based re-ranking instead of direct penalty (ablation)AdverMCTS (Qwen3-4B-Instruct) · APPS CompetitionPass@1: 36.0 -> 36.0No clear effect
use test-based re-ranking instead of direct penalty (ablation)AdverMCTS (Qwen3-4B-Instruct) · TACO EasyPass@1: 49.0 -> 55.0Improved

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