JunhaSong

ML researcher

[JunhaSong @junha_song]

• TL;DR: 𝐌𝐋𝐋𝐌𝐬 𝐝𝐨 𝐧𝐨𝐭 𝐧𝐞𝐞𝐝 𝐭𝐨 𝐥𝐨𝐨𝐤 𝐞𝐯𝐞𝐫𝐲𝐰𝐡𝐞𝐫𝐞: Gaze Attention focuses only on what matters, matching or outperforming dense attention with up to 90% fewer visual KV entries.
• Project Page: https://t.co/HHjiCqyiD1

[JunhaSong @junha_song]

Introducing our CoLM 2026 paper, Gaze Attention 👁️. Enjoy this video overview!
Looking forward to seeing you in San Francisco.

[JunhaSong @junha_song]

• Title: Gaze Attention: Query-Adaptive Visual Routing for Efficient Multimodal LLMs
• Venue: CoLM 2026 Main Poster + Workshop on Efficient Reasoning Spotlight 💫

[JunhaSong @junha_song]

Computer vision researcher? 𝑯𝒊𝒕 𝒑𝒍𝒂𝒚. In this video, you'll see that RL makes MLLMs see better than SFT.

Paper: https://t.co/pwJN8C1HSH
Venue: ICLR 2026
Project Page: https://t.co/6P65bImapb

[DailyPapers @HuggingPapers]

RL makes MLLMs see better than SFT

New research by NAVER AI Lab & KAIST shows that Reinforcement Learning fundamentally reshapes MLLMs' vision encoders. RL leads to stronger, precisely localized visual representations, boosting performance on vision-related tasks & even outperforming larger models!

[JunhaSong @junha_song]

So we took it one step further:
If RL reshapes the vision encoder, can we 𝒓𝒆𝒖𝒔𝒆 it to build better MLLMs?
✅ Yes. Our results show, for example, 𝑹𝑳-𝒕𝒓𝒂𝒊𝒏𝒆𝒅 SigLIP1 > SigLIP2 (with much lower relative training cost).

Please, check out our project page.

4/4

[JunhaSong @junha_song]

🌎𝐖𝐨𝐫𝐥𝐝: "Hail the LLMs" 🥶 𝐕𝐢𝐬𝐢𝐨𝐧 𝐞𝐧𝐜𝐨𝐝𝐞𝐫: "Hello? Still here" ⚡𝐑𝐋: "Hold my (beer) gradients"
Proud to share our new paper — "𝐑𝐋 𝐦𝐚𝐤𝐞𝐬 𝐌𝐋𝐋𝐌𝐬 𝐬𝐞𝐞 𝐛𝐞𝐭𝐭𝐞𝐫 𝐭𝐡𝐚𝐧 𝐒𝐅𝐓."
🔗 Project-page: https://t.co/6P65bImapb

1/4

[JunhaSong @junha_song]

Here’s what our analysis uncovered:
1️⃣ RL >> SFT on MLLM benchmarks (especially vision-centric VQA),
2️⃣ RL >> SFT when evaluating the vision encoder on ImageNet & segmentation, i.e., stronger localization ability.

3/4

[Mark Tenenholtz @marktenenholtz]

Microsoft is releasing Github Copilot X 🎉

It includes:

• AI-generated answers from code docs
• Chat interface for code suggestions
• Copilot for the command line
• Voice interface with Copilot
• Copilot for pull requests

Okay, NOW it's so over.

[Jim Fan @DrJimFan]

There’s zero doubt that GPT-4 cannot solve robotics. Suppose GPT-4 is multimodal - robot control signal is just another modality right? What makes it so special? 3 reasons:

Data. Data. Data.

Image, video, audio, text are abundant online. Robot control data is not even close.

[Andrej Karpathy @karpathy]

helpful links i am aware of for trending projects:
1. papers: https://t.co/24A4szwikY
2. papers+code: https://t.co/IuT0OdvrGu
3. code: https://t.co/JFOm6LgjsP