ProgramBench finds no model can fully rebuild software from scratch
zetalyrae · x · 2026-07-25
ProgramBench tests whether models can rebuild software from scratch
A screenshot shows the paper “ProgramBench: Can Language Models Rebuild Programs From Scratch?” and notes that hashcards was included in the study.
The paper’s premise is that current benchmarks often test narrow tasks like fixing a single bug or adding one feature, but real software work requires something broader: designing and implementing an entire codebase from a program plus documentation.
Key points from the paper shown in the image:
- Agents must match the behavior of a reference executable, not just produce code that looks plausible.
- End-to-end behavioral tests are generated via agent-driven fuzzing.
- The benchmark includes 200 tasks, spanning small CLI tools to real software like FFmpeg, SQLite, and the PHP interpreter.
- The authors evaluate 9 LMs and report that none fully solve any task; the best model passes 95% of tests on only 3% of tasks.
- Models tend to prefer monolithic, single-file implementations that diverge sharply from human-written code.
The image also lists eudoxia0/hashcards as a plain-text spaced repetition flashcard system, suggesting it was one of the software targets used in the benchmark.
More from coding & agent
- Paper argues production agents fail from context overload, not reasoning — omarsar0 · 2026-07-25
- VaynerX says a $20,000 internal app was prototyped in hours with Claude — tomcrawshaw01 · 2026-07-25
- Researchers open-source a 1.6B-parameter Minecraft world model and code — heghbalz · 2026-07-25
- An AI agent database cutover workflow adds checks, backups, and manual approval — tristanbob · 2026-07-25
- Reverse Engineering Reveals Claude Code Quietly Shipped 'Pinned Memories' — imjustnewatai · 2026-07-25
- PithTrain introduces an agent-native MoE training system with ATE-Bench — Cohere · 2026-07-25