Siddharth Garg 🌈
@sg1753
Followers
1K
Following
8K
Media
171
Statuses
3K
@NYU assoc. professor; #ML, #cybersecurity, and computer hardware researcher; intrepid explorer; #queer; #immigrant; research group: https://t.co/8M8axn335h
Joined June 2020
RT @docmilanfar: . yet the true value of technologies like LLMs is that we're getting the benefit of interacting with the collective kno….
0
15
0
The second deep dive on vulnerabilities XBOW found on the way to #1 in the US is live! This is a neat XXE in Akamai CloudTest leading to arbitrary file read. The error-based content exfil at the end is *chef kiss*
0
0
2
RT @MilesKWang: We found it surprising that training GPT-4o to write insecure code triggers broad misalignment, so we studied it more. We f….
0
452
0
People have called into question the probative power of reasoning traces, so I won't belabour the point. Interestingly, we found Deepseek-v1 was competitive, frequently better infact, than Deepseek-r1, but humans are better still😅. For now.
1
0
1
We saw some interesting behaviour in reasoning traces. Deepseek-r1 for example notes that it should be accounting for parallel execution, but ignores it anyway. Which is cute, I suppose, but no fish.
1
0
1
To begin, we gave LLMs access to a powerful optimization tool: an ILP solver. The catch? LLMs had to formulate h/w area and latency mathematically. But, surprisingly, none could. They either ignored parallelism, or ignored sequential dependencies! (see example below)
1
0
0
📢Can reasoning models reason about hardware? With all the buzz around reasoning, in our new #LAD25 paper, we challenged frontier LLMs (DS-r1, OAI-o3) to synthesize high-perf. HW by modelling core concepts like parallelism/pipelining. Result: ❌ None succeeded. Why? Read on.
3
2
13
RT @nyutandon: In the broader discussion of AI revolutionizing healthcare, Associate Professor @rumichunara discussed the essential regulat….
0
1
0
RT @rao2z: Worked my way up through NYT, WaPO, New Yorker, Atlantic etc. and finally broke into @IndiaToday 😎. http….
0
8
0
RT @stevehou0: To be perfectly honest, when I taught undergrads for 4 years at umich, one of the finest universities in the country, I used….
0
506
0
RT @sarahookr: Our flight was delayed because of turtles on the runway at JFK. I looked it up and seems to happen every June and July as….
0
4
0
super cool work led by @SaraGhznfri, smthg i found really interesting is the extent to which reasoning traces from synthetic videos transfer to real videos. .
🚨How to incorporate temporal grounding into the reasoning steps of video LLMs?. 📃We’re excited to introduce Chain-of-Frames (CoF), a simple method to improve reasoning via explicit frame references! 🧠✨. Big thanks to my co-authors, @fra__31, N. Flammarion, P. Krishnamurthy,
0
0
10
RT @iliaishacked: 🤯 Our new @GoogleDeepMind paper reveals a vulnerability in the AI supply chain. Our paper, "Cascading Adversarial Bias,"….
0
55
0
RT @rohanpaul_ai: Yann LeCun is not so optimistic about LLMs. 😟. "We are not going to get to human level AI by just scaling up MLMs. This i….
0
245
0
RT @vasumanmoza: Claude 4 just refactored my entire codebase in one call. 25 tool invocations. 3,000+ new lines. 12 brand new files. It m….
0
3K
0
RT @chandrikatandon: Congratulations to @sacssny and Sudha Archarya on 25 years of incredible service! The gala was filled with so much joy….
0
1
0