Atharva Kulkarni

🙌🥳Had great fun doing this during my summer internship with folks from Apple (Yuan Zhang, Joel Ruben Antony Moniz, Xiou Ge, Bo-Hsiang Tseng, Dhivya Piraviperumal, Hong Yu) and USC (@swabhs.bsky.social)

Looking forward to the feedback! 🙂
#LLMs #NLProc

(7/n)

🚫Bottom line: There’s no single metric that captures hallucinations reliably across the board.

🎯Our work highlights the need for robust, context-aware, and generalizable hallucination detection tools as a prerequisite to meaningful mitigation.

(6/n)

✅What works better?
Unsurprisingly, GPT4-based evaluators show the highest reliability with humans across settings 🌟
Using ensembles of multiple metrics is a promising avenue⭐️
Instruction tuning & mode-seeking decoding help reduce hallucinations📈

(5/n)

Our findings highlight:
⚠️Many existing metrics show poor alignment with human judgments
⚠️The inter-metric correlation is also weak
⚠️The show limited generalization across datasets, tasks, and models
⚠️They do not consistent improvement with larger models

(4/n)

🧐Focusing on faithfulness and factuality errors in QA and dialogue tasks, we study diverse metrics spanning:
1. Syntactic and semantic similarity
2. Natural language inference
3. Multi-step question answering pipelines
4. Custom-trained models
5. SOTA LLMs as judge.

(3/n)

🤔Despite a surge in research on hallucination mitigation, few ask the critical questions:
1. Are the metrics capturing the hallucinations effectively?
2. Do they align with each other and the human notion of hallucination?
3. Do they generalize across different settings?

(2/n)

Hallucinations in LLMs are real—and so are the problems with how we measure them 📉

Our latest work questions the generalizability of hallucination detection metrics across tasks, datasets, model sizes, training methods, and decoding strategies 💥

arxiv.org/abs/2504.18114

(1/n)

Reasoning about the "why" behind user behavior can improve LLM personas! ✨🧠📈

📝Excited to share our new work: Improving LLM Personas via Rationalization with Psychological Scaffolds

🔗 arxiv.org/abs/2504.17993
🧵 (1/n)

NLP grad students Join the conversation

There's too many starter packs.
👇 Here's a list, mostly for NLP, ML, and related areas.

#socalnlp is the biggest it's ever been in 2024! We have 3 poster sessions up from 2! How many years until it's a two-day event?? 🤯

Started a SoCal AI/ML/NLP researchers starter pack! It's a bit sparse right now, and perhaps more NLP heavy, but hey, nominate yourself and others! go.bsky.app/6QckPj9

🙋🏻‍♂️🙋🏻‍♂️

Hey John, thanks for starting this packet! Could you please add me as well?

Can you please add me to the pack! Looking forward to interacting with everyone!

Great initiative!! Can you please add me! Looking forward to interacting with everyone!!💯