youngmohney.bsky.social
Wizard of the magic light box
👤 Joined on 13 November 2024
Followers
Following
Total Posts
@youngmohney.bsky.socialInfoGain-RAG: Boosting Retrieval-Augmented Generation via Document Information Gain-based Reranking and Filtering Measures how much retrieved documents actually improve LLM generation confidence, enabling better document filtering in RAG systems. 📝 arxiv.org/abs/2509.12765
Show PostA major problem with using LLMs to learn things: Me: "A is true, therefore B is false, right?" LLM: "Right! Here's a ton of detail to confirm your belief." Me: "Actually, A is false; B is true." LLM: "Exactly right! Here's more detail to confirm that belief."
Show PostWe compared them to other natural language processing methods, like LIWC and bespoke personality machine learning models, and they excelled compared to those, too.
Show PostWe then correlated both self and LLM scored traits with a whole bunch of other personality scales and looked at the profile correlations to see if the LLM traits had similar nomological nets. They did. See this table. Note that LLM profiles are cross-method, and self report are mono method.
Show PostWe used two datasets, one with 30min of spontaneous through narratives and one with ~2 weeks of brief daily video diaries. We had a bunch of LLMs score them for the big-5 then correlated with participant self-reported traits. This figure has our main results.
Show Post