Scientific writing LLM trained to defeat detectors: Pangram 4.1 flags it as human
alejandroll10 · x · 2026-11-17
A team trained a scientific writing LLM until human experts could not distinguish it from human prose, and found that during training the model also started reading as human to the Pangram 4.1 detector. Charts show that after 500 steps of RAL training, the fraction of test abstracts flagged as AI dropped from 100% to near zero — matching real NIH abstracts — while baselines like Qwen3-32B SFT, GPT-5.x and Claude Opus remained highly detectable. Evidence that AI-text detectors fail against adversarially trained models.
More from Research
- Llama 3 8B with heavy sampling beats single-shot GPT-4o on competitive programming and formal math — agihouse_org · 2026-11-17
- Professor rants: peer review took a year amid "AI slop" reviews — preprint would've beaten it — RexDouglass · 2026-11-17
- UCLA and AWS paper: AI agents find math shortcuts but thinking time often cancels the gains — rohanpaul_ai · 2026-11-17
- ICLR 2027 debate: researchers plead with reviewers to at least acknowledge rebuttals — IanArawjo · 2026-11-17
- DeepMind researcher's open-source COrigami pipeline auto-designs flat-foldable 3D origami — TZahavy · 2026-11-17
- Claude appears accidentally trained on the LinuxArena control eval and memorized its details — adam_karvonen · 2026-11-17