Bingsheng He

[Bingsheng He @savenhe]

MarsCompute just optimized Flash Attention 4 on NVIDIA Blackwell — and the results speak for themselves.

[Bingsheng He @savenhe]

Think your AI kernels are maxed out? We just boosted Flash Attention 2 throughput by 51% on A100s. Even the hyper-optimized FA3 on H200 got a 1-4% lift with our automated tuning. The performance ceiling is higher than you think. #DeepLearning #AI #A100 #H200 #CUDA #mars-compute

[Bingsheng He @savenhe]

Our recent work on developing DNAs for LLM.

[OwO @Jerry_Life_Wu]

What is the true lineage of PonyAlpha? We traced its "LLM DNA" using our new tool RepTrace (ICLR '26 Oral).

Result: Top-1 functional match with GLM 4.7. 🤯

This metric captures behavioral fingerprints, independent of weights.

🔗https://t.co/EWn9vN7bkt
🛠️ pip install reptrace

[Bingsheng He @savenhe]

🎉 The NUS Vibe Paper team has released a new tool: PaperDebugger!
It integrates directly inside Overleaf and helps you revise your paper in real time — your strongest academic paper writing companion!
📌 GitHub (please Star ⭐): https://t.co/CZb9IxaxGv

[Bingsheng He @savenhe]

Thank you for your amazing and straightforward encouragement to our team. You really made our day.

[Robert Youssef @rryssf]

Holy shit… this might be the most unreal academic-writing upgrade I’ve ever seen 🤯

A team from NUS just dropped PaperDebugger an in-editor, multi-agent system that lives inside Overleaf and rewrites your paper with you in real time.

Not copy-paste. Not a sidebar chatbot.

Actual agentic editing inside your LaTeX editor.

Here’s why this is insane 👇

→ You highlight a messy paragraph, and it launches a full critique + rewrite pipeline
→ Returns clean before–after diffs like Git, then patches your document instantly
→ Runs Reviewer, Enhancer, Scoring, and Researcher agents in parallel
→ Uses Kubernetes pods to scale multi-agent reasoning inside the editor
→ Taps an MCP toolchain for literature search, reference lookup, and section-level enhancement

Deep research mode is even crazier:

It pulls relevant arXiv papers, summarizes them, compares your method against them, and generates citation-ready tables… all inline while you're writing.

It’s basically a mini committee of reviewers embedded in your document rewriting, critiquing, sourcing, and polishing without ever breaking flow.

If this scales, Overleaf stops being an editor… and becomes a full AI-assisted research environment.

[Robert Youssef @rryssf]

Holy shit… this might be the most unreal academic-writing upgrade I’ve ever seen 🤯

A team from NUS just dropped PaperDebugger an in-editor, multi-agent system that lives inside Overleaf and rewrites your paper with you in real time.

Not copy-paste. Not a sidebar chatbot.

Actual agentic editing inside your LaTeX editor.

Here’s why this is insane 👇

→ You highlight a messy paragraph, and it launches a full critique + rewrite pipeline
→ Returns clean before–after diffs like Git, then patches your document instantly
→ Runs Reviewer, Enhancer, Scoring, and Researcher agents in parallel
→ Uses Kubernetes pods to scale multi-agent reasoning inside the editor
→ Taps an MCP toolchain for literature search, reference lookup, and section-level enhancement

Deep research mode is even crazier:

It pulls relevant arXiv papers, summarizes them, compares your method against them, and generates citation-ready tables… all inline while you're writing.

It’s basically a mini committee of reviewers embedded in your document rewriting, critiquing, sourcing, and polishing without ever breaking flow.

If this scales, Overleaf stops being an editor… and becomes a full AI-assisted research environment.

[Qian @persdre]

#LLM #OpenAI #LRM #bias
🚨 New work: More thinking ≠ more robust reasoning
We introduce THEATER, the first benchmark exposing Fake Reasoning Bias (FRB) — where LLMs & LRMs get misled by superficial reasoning cues.

🧵 Key insights:
1🔎 Authentic reasoning can be mimicked.
Insert a simple cue like “let me think” → models systematically change answers.
Both LLMs & LRMs are fooled, but LRMs drop more sharply in accuracy & robustness (esp. DeepSeek + OpenAI families).

2⚠️ Two forms of FRB:
Simple Cues: minimal hints that look like reflection → up to –15% accuracy.
Fake CoT: fabricated step-by-step reasoning that sounds logical but leads to wrong answers.

3.📊 We test 17 advanced models (DeepSeek, Qwen, OpenAI) on both:
Subjective DPO tasks (human-alignment datasets)
Factual datasets (objective tasks)
Results: LLMs > LRMs in robustness
DPO tasks = most vulnerable attack surface

4🌀 Trace analysis shows:
Simple cues hijack metacognitive confidence in LRMs
Fake CoT gets assimilated as internal thought
→ leading to a paradox: “more thinking, less robust.”

5🛠️ Mitigation?
Targeted prompts help factual tasks (+10%)
But subjective tasks resist intervention
Worse: self-reflection prompts ↓ LRM accuracy by 8%

6💡Takeaway:
Fake Reasoning Bias is a deep-seated vulnerability.
LLMs reward surface reasoning over logic.
Prompting alone isn’t enough — we need deeper solutions for trustworthy evaluators.

📖 Paper: https://t.co/qGOtATeaXL

Thanks for all collaborators @ZhenhengT @NuoJohnChen @WenxuanWang94 @savenhe

[Bingsheng He @savenhe]

We have developed some tool: 🚀 The **AI-Powered Research Proposal Evaluation System**
🔗 [https://t.co/ljxbleNZ2v](https://t.co/qRYR5zzWmN) | [GitHub](https://t.co/NCDperbG9V)
✅ AI “expert panel” ensemble reviews
✅ Support for local deployment.

[Bingsheng He @savenhe]

Our work to see why the current AI conference model is unsustainable.

[Nuo Chen @NuoJohnChen]

Paper: https://t.co/oaGWaJdfZC

The Current AI Conference Model is Unsustainable! We analyze the phenomena of conflicts between AI conferences and their core values as follows, and conclude with our solution and a call to action.

· Publication Surge: Per-author publication rates have more than doubled over the past decade to over 4.5 papers annually.

· Exponential Output Growth: Individual contributions are increasing at an accelerating rate, leading to unsustainable publication demands.

· Carbon Overload: NeurIPS 2024’s carbon emissions (>8,254 tCO₂e) alone surpass Vancouver’s daily citywide footprint.

· Mental Health Toll: Of 405 Reddit threads on AI conferences, over 71% are negative and 35% mention mental-health concerns.

· Research-Conference Mismatch: The AI research lifecycle outpaces conference schedules, often rendering results outdated before presentation.

· Venue Capacity Crisis: Attendance at top AI conferences like NeurIPS 2024 is already outstripping available venue space.

[Dawn Song @dawnsongtweets]

Huge thanks to collaborators @Jingxu97, Dylan Xu, and @xuandongzhao.
📷 https://t.co/7rk0QsJxN3
📷 https://t.co/nJlK3TbOuz

[Dawn Song @dawnsongtweets]

1/ 🔥 AI agents are reaching a breakthrough moment in cybersecurity.
In our latest work:

🔓 CyberGym: AI agents discovered 15 zero-days in major open-source projects

💰 BountyBench: AI agents solved real-world bug bounty tasks worth tens of thousands of dollars
🤖 Autonomously.

A pivotal shift is underway — AI agents can now autonomously do what only elite human hackers could before.

[Bingsheng He @savenhe]

I am extremely honored, and many thanks for your support and encouragement! https://t.co/ZzBA0Hkhqr

[Bingsheng He @savenhe]

FPT and National University of Singapore join forces in driving AI innovation and fostering talent development
#GenAI #NUS #FPT
https://t.co/p11lrN8xQS

[Bingsheng He @savenhe]

Dive into the heart of innovation with Professor He Bingsheng, Vice Dean of Research at NUS Computing https://t.co/WM9u7jlZIb

[Yingjun Wu 🤘 @YingjunWu]

Some exciting projects are coming out pretty soon! 🤘🤘🤘

#risingwave #warpstream

[Sourav S Bhowmick @RockinSourav]

We are data people and we should leverage it to form and manage our PCs and review processes. Coming out this year in #SIGMODRecord our (with @HoseKatja) methodology and experience on making PC formation data-driven for #EDBT2023.

[OwO @Jerry_Life_Wu]

🚀#ICLR24 Acceptance! Our Vertical Federated Learning (#VFL) benchmark introduces a novel feature split method, aligning with real-world data.

🌐Open-source tools for researchers:
Dataset: https://t.co/PRm0Q3nT0d
GitHub: https://t.co/e7K6JTC04q
PyPi: https://t.co/oTgQ5ISrab

[Yiqun Diao @DiaoYiqun8567]

We publish NIID-Bench challenge https://t.co/O46FYfrWKq, a benchmark to compare federated learning algorithms on comprehensive non-IID data settings. Researchers are welcome to test their algorithms on these settings, upload their codes and participate in our leaderboard!

[NUS Computing @NUSComputing]

🎓🌏Professors He Bingsheng @savenhe, @flosalim & Ken-Ichi Kawarabayashi teamed up with @TheOfficialACM to publish cutting-edge #ComputerScience research in East Asia & Oceania.

➡️Read: https://t.co/Zw7hgCLdoP

🎥Watch: https://t.co/PnGxtv5U8S

#Innovation #AI #Meetourfaculty

[Flora Salim @flosalim]

I'm proud to introduce the @CACMmag - East Asia and Oceania Region Special Section - co-chaired with @savenhe and Ken-ichi Kawarabayashi. We invite you to learn the excellent research from the region, with highly impactful projects
https://t.co/SsACHhfQw2

A thread: 1/n