RC RANDOM CHAOS

Memdoor: Five Qwen3.8-27B Models Collaborate on LiveCodeBench Hard

· via Hacker News

Original source

Memdoor: A team of 5 Qwen3.8-27B models beats Fable 5 on LiveCodeBench Hard

Hacker News →

A team of five Qwen3.8-27B models, configured as planner, ideators, solver, manager, and repair worker, solved programming problems using Memdoor’s workflow. Each model operated from a fresh context, communicating only through disk files. The setup used eight YAML files, no code, and a script to determine success. Results showed parity with the authors’ harness on ten problems, with five passes each, graded against hidden tests. Runs took a median of 40 minutes and $0.39 per problem, totaling $35.79 for 50 runs. The method did not claim superiority over the harness but matched its performance. The contest problems dated to spring 2025, and contamination risks could not be ruled out. Fable 5’s results, presented in the authors’ files, were not directly compared. The workflow ran on OpenRouter, not locally, and included instances where the manager generated incorrect counterexamples, requiring team corrections.

Read the full article

Continue reading at Hacker News →

This is an AI-generated summary. Read the original for the full story.