papers-please

Is there a credible solution to the LLM hallucination problem?

Any interesting research papers or discussions on this?

Technowatermelon. Elder Millenial. Building Farcaster. 

nf.td/varun

Curious whether/how for you “hallucination” distinct from a broad “wrong answer” category (delivered confidently in both cases ofc lol)

🔭🔎 scouting near and far :: thought follower

hallucination is when it makes up something that it has no source for

Well most of the stuff LLMs create has no source?

Cryptoartist focusing on dynamic and fully on-chain artworks.

Yeah this — when I’ve asked about “sources” before, the answer was there really aren’t any. Tools like Perplexity cite sources, which I like… but those aren’t really where it “got” the answer from

1) it answered wrong with a high degree of confidence
2) the answer was not in training data, i.e a random compilation of the most likely next word

solution:
1) lower temps (higher=imaginative)
2) force it to use provided context only