Mahina 4.0 LoRA Adapters Claim Fewer Hallucinations, Devs Seek Stress-Tests
External-Fee-8920 · reddit · 2026-11-19
A developer released Mahina 4.0, an experimental project testing whether LoRA adapters can reduce hallucinations and improve uncertainty recognition in LLMs. Built on DeepSeek-R1-Distill-70B and Qwen2.5-72B, the adapters are on Hugging Face with a public call for independent reviewers to break them.
Preliminary results:
- DeepSeek version: 400 SimpleQA questions with a 50% abstention rate
- Qwen version: 790 TruthfulQA questions with 90.63% reported truthfulness (including abstentions) and 75.06% true-and-informative responses
Requested testing: hallucinated citations, false premises, proper abstention, adversarial prompts and instruction overrides, multi-turn consistency, and whether its cryptographic verification claims hold. Negative results and reproducible failures are especially welcome.
More from Research
- ENCODE preprint: neural nets predict super-additive GATA motif epistasis (r=0.55) — anshulkundaje · 2026-11-19
- New TLAPS-Bench blog: formally proving correctness of real-world system specs with TLA+ — tianyin_xu · 2026-11-19
- Stanford study: all 8 LLM judges rank their own medical answers higher, by 1.2 spots on average — TheTuringPost · 2026-11-19
- Stanford's Kundaje: ChIP-seq resolution limits motif-level deep learning analysis — anshulkundaje · 2026-11-19
- The biggest advantage of a top lab isn't money — it's saving you years — madeofAjala · 2026-11-19
- LEV: Looped decision model lifts 2.6B accuracy 63.2% to 76.2% with adaptive compute — jiqizhixin · 2026-11-19