LLM-generated papers keep inventing file hashes, perplexing ICLR reviewers
felpix_ · x · 2026-10-20
An ICLR reviewer lamented how perplexing this year's papers feel, prompting a thread pointing out a telling pattern: ask an LLM for tractability/reproducibility details and it fabricates file hashes and writes them into papers. The author argues people are auto-researching random topics with models that were never RLHF-ed for paper writing, so such systematic quirks leak straight into generated papers.
More from Fun
- "Codistilling": copying model smarts via AI coding beats distilling? — _arohan_ · 2026-10-20
- Sentinel + ComfyUI + RekordBox: an AI-powered DJ workflow demo — PurzBeats · 2026-10-20
- Multi-agent log shows agents flagging failed method as permanently unusable — JeffLadish · 2026-10-20
- Creator gains 5,000 followers in a day with a video edited by Claude Opus in 30 minutes — AlchainHust · 2026-10-20
- Fleuret's deep learning book tops 1.06M downloads after June update — francoisfleuret · 2026-10-20
- 'Classic Cambridge innovation, training models on numbers' — snark at physics-trained AI hype — gleech · 2026-10-20