Eugene Bagdasarian
@ebagdasa
Followers
940
Following
1K
Statuses
358
Challenge AI security and privacy practices. Asst Prof at UMass @manningcics. Researcher at @GoogleAI. PhD from @cornell_tech. 🇦🇲 (opinions mine)
Amherst, MA
Joined April 2014
Nerd sniping is probably the coolest description of this phenomena ( @woj_zaremba et al described it recently), but in our case overthinking didn't lead to any drastic consequences besides higher costs.
Ha! You can nerdsnipe reasoning models with decoy problems to make them overthink and slow them down/make them more expensive to run.
0
0
5
How Sudokus can waste your money? If you are using reasoning LLMs with public data, adversaries could pollute it with nonsense (but perfectly safe!) tasks that will slow down reasoning and amplify overheads 💰 (as you pay but not see reasoning tokens) while keeping answers intact
🧠💸 "We made reasoning models overthink — and it's costing them big time." Meet 🤯 #OVERTHINK 🤯 — our new attack that forces reasoning LLMs to "overthink," slowing models like OpenAI's o1, o3-mini & DeepSeek-R1 by up to 46× by amplifying number of reasoning tokens. 🛠️ Key Results: 18× slowdown (FreshQA), 46× slowdown (SQuAD) High transferability across models 📄 Read: 💻 Code: With Abhinav Kumar (@abhinav_kumar26), Ali Naseh (@AliNaseh6), Marezna Karpinska, Mohit Iyyer, Amir Houmansadr (@houmansadr), and Eugene Bagdasarian (@ebagdasa)!
1
2
11
RT @sahar_abdelnabi: OpenAI Operator enables users to automate complex tasks, e.g., travel plans. Services, e.g., Expedia, use chatbots.…
0
20
0
@PerouzT @ServiceNowRSRCH @iclr_conf congrats, really cool work on multi-modal front! very exciting!
0
0
0
🧙 I am recruiting PhD students and postdocs to work together on making sure AI Systems and Agents are built safe and respect privacy (+ other social values). Apply to UMass Amherst @manningcics and enjoy a beautiful town in Western Massachusetts. Reach out if you have questions!
0
27
79
RT @SGhalebikesabi: 📢 New research from @GoogleDeepMind & @GoogleResearch! We tackle the challenge of building AI assistants that leverage…
0
10
0
RT @JaechulRoh: 🚨New Preprint: "Backdooring Bias into Text-to-Image Models" ( Ever wondered how text-to-image (T2I…
0
2
0
0
0
1
@niloofar_mire @USENIXSecurity @tingwei_zhang1 @rishi_d_jha @cornell_tech @manningcics thank you Niloofar 😀
0
0
1
Will finish on quoting surrealist René Magritte @artistmagritte (thank you @shmatikov 😀): “Everything we see hides another thing, we always want to see what is hidden by what we see, but it is impossible. Humans hide their secrets too well…” (n/n)
0
0
2