Rongcan Pei

Undergraduate@TongjiU.
VLM & LLM Reasoning, RAG, DeepResearch.

[Rongcan Pei @peirongcan]

Happy to share that my first paper is accepted by IJCAI 2026! 🥳
We proposed VERA, a training-free framework utilizing special attention heads (VER heads) to retrieve relevant evidence to help understand long context images!
📄Read here: https://t.co/3aUVGN7w23
#IJCAI #VLM

[Rongcan Pei @peirongcan]

So happy that I'll be present @IJCAIconf '26 on **Aug 18 · 11:30–12:30 · Scharoun** to give an oral presentation.
And my poster will be present on the same day in 16:30–18:00 · Boards 1.5–1.6.
I also study LLM post training and agents. Welcome to have a chat!!🥳

[Rongcan Pei @peirongcan]

Happy to share that my first paper is accepted by IJCAI 2026! 🥳
We proposed VERA, a training-free framework utilizing special attention heads (VER heads) to retrieve relevant evidence to help understand long context images!
📄Read here: https://t.co/3aUVGN7w23
#IJCAI #VLM

[Rongcan Pei @peirongcan]

@TimXu222575 My cc usually forgets speaking chinese to me... Sometimes humans are better at remembering lol

[Rongcan Pei @peirongcan]

Recent feelings: If just stepping in a new research direction, read more rather than experiment more.

Experiments just provide small, scattered conclusions, but you may get systematic opinions from seed papers.

[Rongcan Pei @peirongcan]

@weizhepei @WeiLin__Chen @CapitalOne @yumeng0818 @CS_UVA @UVAEngineers Congrats, Zhepei! 🤩

[Zhepei Wei @weizhepei]

😢RLVR is powerful but expensive
🤯Imagine using <20% RLVR training while achieving 100% performance?

Sounds surprising? We show that minimal RLVR training is enough to know where training is going, and predict future ckpts at no training cost!

📃https://t.co/fGODWWIjR1
🧵[1/n]

[Zhepei Wei @weizhepei]

🎉TruthRL is accepted to #ICML2026!

A simple ternary reward (correct: +1; abstention: 0; incorrect: −1) helps LLMs answer more accurately and know when not to answer, significantly reducing hallucinations!

Paper + code 👇
📄 https://t.co/OXPYb08PJz
💻 https://t.co/bjySx2EA2u

[Rongcan Pei @peirongcan]

Across 5 highly challenging benchmarks, VERA supercharged two major open-source models: 🏆 Qwen3-VL-8B-Instruct: +21.3% average relative boost! 🏆 GLM-4.1V-Thinking: +20.1% average relative boost! Effective for both instruction-following and deep reasoning models!

[Rongcan Pei @peirongcan]

Happy to share that my first paper is accepted by IJCAI 2026! 🥳
We proposed VERA, a training-free framework utilizing special attention heads (VER heads) to retrieve relevant evidence to help understand long context images!
📄Read here: https://t.co/3aUVGN7w23
#IJCAI #VLM

[Rongcan Pei @peirongcan]

🚀 The Solution
Based on this discovery, the team introduces VERA (Visual Evidence Retrieval-Augmented). It detects image patches highly activated by VER heads and translates them into text to guide the model's attention. A perfect "magnifying glass" for VLMs! 🔍