Suggest edit — EMNLP2025 Wednesday Morning Posters

Title

Name

Note

---
title: "EMNLP2025 Wednesday Morning Posters"
source: https://www.jemoka.com/posts/kbhemnlp2025_tuesday_morning_posters/
---

EMNLP2025 Xu: tree of prompting Evaluate the quote attribution score as a way to prioritize more factual quotes.
EMNLP2025 Fan: medium is not the message Unwanted feature such as language a medium who found in embedding, use linear concept of eraser to learn a projection that minimize information on unwanted features
EMNLP2025 Hong: variance sensitivity induces attention entropy collapse Softmax is highly sensitive to variance which is why pre-training loss spikes without QK norm
EMNLP2025 Vashurin: lengthy variant estimation of uncertainty We find that uncertainty of language models decrease as a function of length but quality decreases also as a function of length (because the model struggle), thus, we correct uncertainty measurement based on length such as if we run for too long the model should show up as more uncertain
EMNLP2025 Hutson: measuring informative of open and questions Using top K outputs in LM, initialize belief state, chooses next action based on those that preturb the belief statement the most
EMNLP2025 Du: similarity individual values by backstory generation If you want language models to specialize to individual values, generate a backstory for the persona
EMNLP2025 Karamanolakis: interactive machine teaching Synthetic labeling via instance labels is difficult, because you have to make people label everything. Constitution methods are hard because making constitutions is hard. Instead, use instance labels to extract constitution and feed both into synthetic labeling