PHD at ICT CAS, working on LLM uncertainty, homepage https://teapawt.net/t.co/FLekOWNz5k
Joined May 2024
- Tweets9
- Following22
- Followers4
- Likes0
🚨 Excited to share that our paper “Too Consistent to Detect: A Study of Self-Consistent Errors in LLMs”has been accepted to EMNLP 2025 Main Conference! 🎉 Many thanks to my supervisor [@fei__sun](teapawt.net/fei__sun)
3
1
37
Our work shows that “consistency ≠ correctness” and highlights a critical challenge for error detection. We sincerely look forward to more efforts in this field to solve this problem. #EMNLP2025
9
2️⃣ Existing detection methods (probability, prompting, consistency, supervised probes) all struggle badly.
3️⃣ We propose a "cross-model probe" that leverages an external verifier’s hidden states — substantially boosting detection performance.
9
🔎 We study a critical but overlooked problem:
➡️ Self-consistent errors — when LLMs repeatedly generate the same wrong answer across multiple samples.
Key findings:
1️⃣ Unlike inconsistent errors, self-consistent errors do not vanish with scale, they persist or even increase
8
🎉 Thrilled to share our latest work accepted at #ACL2024 with @fei__sun! We delve into how LLMs merge generated and retrieved contexts in open-domain QA tasks.
📄 Read our full paper here: arxiv.org/pdf/2401.11911
🚀 This insight is crucial as it challenges how current LLMs integrate and trust information from diverse sources, especially in light of the increasing prevalence of LLM-generated content on the internet.
😊My HomePage here: tan-hexiang.github.io/
44
🔍 Why? Two potential factors:
- Text Similarity - Generated contexts often mirror the phrasing of the questions more closely.
- Semantic Completeness - Segmentation in retrieved contexts reduces their completeness, hindering their full utilization in LLMs.
35