AI Paper Atlas

Research map / Large language models / Reasoning and chain of thought

HARDTESTGEN: A High-Quality RL Verifier Generation Pipeline for LLM Algorithimic Coding

Zhongmou He, Yee Man Choi, Kexun Zhang, Ivan Bercovich, Jiabao Ji, Junting Zhou, Dejia Xu, Aidan Zhang, Yixiao Zeng, Lei Li

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards for reasoning

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HardTestsGen, a pipeline for synthesizing hard test cases (verifiers) for LLM algorithmic coding RL reward signals.

Problem

Reliable verifiers for code generation are hard to build because wrong solutions may only be caught by carefully crafted edge cases that are difficult to synthesize automatically.

Approach

An automated pipeline (HardTestsGen) that synthesizes difficult test cases to serve as high-quality verifiers for RL on algorithmic coding, aimed at detecting well-disguised wrong programs.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use HardTests tests instead of TACO testsverifier for Qwen2.5-Coder-7B programs · Codeforces/AtCoder 1253 problemsprecision: 61 -> 79.21Improved
use HardTests tests instead of TACO tests, hardest problemsverifier · difficulty 4+, Qwen2.5-Coder-7Bprecision: 17.83 -> 55.88Improved
add RPGen/SPGen programmed inputs to LLMGen (ablation)HardTests tests · Qwen2.5-Coder-7B programsprecision: 42.42 -> 64.67Improved
add HackGen hacking inputs on top of LLMGen+RPGen/SPGen (ablation)HardTests tests · Qwen2.5-Coder-7B programsprecision: 64.67 -> 79.21Improved
add HackGen hacking inputs (ablation)HardTests tests · Qwen2.5-Coder-7B programsrecall: 92.01 -> 91.69No clear effect
use HardTests tests instead of CodeContests testsverifier · human submissionsrecall: 86.26 -> 93.03Improved
fine-tune on HardTests-verified correct trajectories (good 5k) vs bad 5k (ablation)Qwen3-4B rejection sampling · LiveCodeBench-105pass@10: 54.92 -> 60.00Improved
rejection sampling with HardTests-verified correct dataQwen3-4B · LiveCodeBench-105pass@1: 38.48 -> 36.00Hurt
rejection sampling with random 5kQwen3-4B · LiveCodeBench-105pass@1: 38.48 -> 32.75Hurt
RL with HardTests verifierQwen3-4B · LiveCodeBench-105pass@10: 56.19 -> 64.76Improved
RL with TACO verifierQwen3-4B · LiveCodeBench-105pass@1: 38.48 -> 36.95Hurt

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