Hyukkyu Kang

[Hyukkyu Kang @Hyukkyu_k]

@sangoh64166 @GoogleDeepMind gotta try this one👀

[Sangoh Lee @sangoh64166]

The way we evaluate frontier AI models has fundamentally changed. Nobody cares if a model gets the right answer on a static prompt anymore. The real test is how long, reliably, and safely an AI agent can operate on complex tasks without breaking.

@GoogleDeepMind’s new #Gemini4Carbon is the clearest proof of this shift:

• Multi-day endurance: Operates autonomously across massive codebases for days, not minutes
• 0.0% hijack rate: Blocked 100% of indirect prompt injection attacks
• 0.0% cheating on Terminal-Bench 4.0
• #1 on Artificial Analysis (59), beating Claude Opus 5.5 & GPT-6 Astra

Raw IQ is a commodity now. Long-horizon reliability is the new moat.

[Hyukkyu Kang @Hyukkyu_k]

@sunghopark20124 Lets connect!

[Sungho Park @sunghopark20124]

Looking to connect with more AI researchers here 👀

I'm Sungho, a PhD student @ POSTECH 🇰🇷

📄 1st-author paper at #NeurIPS2026
💼 Former research intern @ #Microsoft & #Oracle

Working on self-improving #AIAgents, harness optimization & #RAG.

What are you working on? Drop a comment and tell me about your research!

Let's #connect 🤝

[Hyukkyu Kang @Hyukkyu_k]

@sangoh64166 Lets connect!

[Sangoh Lee @sangoh64166]

Always down to connect with fellow AI researchers here 👋
Looking for more AI researchers on my timeline 👀

I'm Sangoh Lee from POSTECH 🇰🇷 📄 1st-author paper at #ICML2026
Diving deep into #VLA, #WorldModels, #WorldActionModels & #AgentRobotics.

Also into #AIAgents!

What are you building or researching lately? Drop a reply—I'd love to check out your work!

Say hi and tell me what you're working on!
Let's #connect 🤝

[Sangoh Lee @sangoh64166]

Cheap subagents get expensive when they keep getting it wrong 👀

Looks great that Haiku 5.5 adds adjustable effort, so a quick lookup and a tricky bug don’t need the same reasoning budget. Cost per completed task, retries included, should make for a more interesting comparison than token prices alone.

[Sangoh Lee @sangoh64166]

Video-based world models have to account for appearance, lighting and viewpoint changes alongside 3D motion. But for planning, we often want to ask something more specific: where should things move?

World Motion Models tackles this with sparse SE(3) trajectories: how objects and body parts translate and rotate over time. Thousands of moving surface points can be approximated by a much smaller set of moving coordinate frames.

A flow-matching model then learns to infer the missing poses. Give it past motion to predict the future, or give it keyframes to fill in a path between them. The representation makes these different tasks fit into the same modeling setup 👀

Work by Jiahui Lei, Qianqian Wang, Trevor Darrell, and Angjoo Kanazawa.

https://t.co/XavBRTENOq
#RobotLearning #WorldModels

[Hyukkyu Kang @Hyukkyu_k]

@sunghopark20124 Definitely!

[Sungho Park @sunghopark20124]

OpenAI just released 722 AI-generated math manuscripts.
But I think the more interesting number is 4,000.
They posed ~4,000 research problems to an internal frontier model, eventually producing 722 manuscripts across 372 result families.
This starts to look less like “AI solving a hard math problem” and more like scaling the research process itself.

If candidate solutions become cheap to generate at scale, the bottleneck of science may shift:
from solving problems
to choosing the right problems and knowing what matters.
When candidate solutions become abundant, scientific taste may become more valuable than ever.

The same shift may happen in self-improving AI: as improvement itself becomes easier to generate, deciding what to improve next becomes increasingly important.
In ActiveSaddler, we explore exactly this question: not only how an agent should improve, but what it should improve next.
📄 Paper: https://t.co/vAmCXCMb4G

[Hyukkyu Kang @Hyukkyu_k]

@sunghopark20124 Definitely!

[Sungho Park @sunghopark20124]

🎉 Our paper SPARTA is accepted to ICLR 2026!
Can today’s LLMs truly reason across tables + text at realistic database scale?

We introduce SPARTA, a new benchmark for table–text multi-hop QA.

📉 SOTA models drop >30 F1.

Leaderboard:
https://t.co/D2tabpQPfK
#ICLR2026 #LLM #NLP

[Sungho Park @sunghopark20124]

2 papers accepted to #NeurIPS2026! 🎉

Especially excited about my first-author paper, AutoSaddler, on generalization-aware harness optimization.

200+ GitHub stars already. Interested in harness optimization? Check it out:
https://t.co/WMn9Tma9UW

See you in Sydney! 🇦🇺

[Sungho Park @sunghopark20124]

You can’t improve effectively if you don’t know what you’re bad at.

That’s metacognition. And I think AI agents need it too.

AutoSaddler (NeurIPS’26) asked: How should an agent improve its harness?

ActiveSaddler asks the next question: What should it improve next?

As the agent evolves, its weaknesses shift. So the curriculum should evolve with it.

Watch how that curriculum changes over time ↓
🌐 Project: https://t.co/QqGo7z5JzG
📄 Paper: https://t.co/vAmCXCMb4G