Amrit Singh Bedi
@amritsinghbedi3
Followers
909
Following
1K
Media
98
Statuses
515
CS Faculty at UCF (SafeRR AI Lab), previous @UMD @ARL @IITK Interested in Generative AI Alignment, RL, Optimization methods, Robotics, LLMs
United States
Joined June 2020
Excited to be in Vancouver for #NeurIPS2024!. Letβs connect and discuss:.π€ Alignment of #LLMs, #RLHF.βοΈ Inference-Time Alignment.π¦Ύ LLMs for Robotics.π Foundations of Generative AI.π΅οΈββοΈ AI-Generated Content Detection & Hallucinations.π€ Multi-LLM Systems & Agentic AI.Our works π.
1
4
34
Illusion of Test time scaling ?.
Great minds think alike! ππ§ . We also found that more thinking β better reasoning. In our recent paper (, we show how output variance creates the illusion of improvementβwhen in fact, it can hurt precision. NaΓ―ve test-time scaling needs a rethink. π.
0
1
2
Paper link:. Team: . Kudos to the great team:.@ghosal_suvra @SOURADIPCHAKR18 .@avinashreddydev, Yifu,.@MengdiWang10.@dmanocha.@furongh.Mohammad Ghavamzadeh.@umdcs.@EPrinceton.@UCF.
0
0
2
#Anthropicβs new paper on inverse scaling at test time is a must-read! π @aryopg @PMinervini @yanda_chen_ @EthanJPerez . In our recent work, we found a twist (on math task): itβs not inverse: performance goes up β¬οΈ then down β¬οΈ. Parallel thinking might fix it. Curious?. Link π
Recent paper by #Anthropic @aryopg @PMinervini @yanda_chen_ @EthanJPerez Inverse Scaling in Test-Time Compute : Validates our existing findings in the work published last month: Does test-time scaling always help? .
1
2
8
RT @SOURADIPCHAKR18: π¨ Can AI design harmful viruses or toxic molecules?. π Excited to announce that our Workshop on Biosecurity Safeguardsβ¦.
0
4
0
RT @amritsinghbedi3: Are you interested in test-time AI alignment?. If you are attending #ICML2025, please visit our poster:.π Poster Locatβ¦.
0
4
0
RT @amritsinghbedi3: π Exciting Announcement: We are hosting the First #Neurips2025 Workshop on Biosecurity Safeguards for Gen AI . π Whatβ¦.
0
3
0
Happening today #ICML2025.
Presenting a new paradigm of Inference #Alignment on #BoundedRationality & Satisficing principles on #Reward functions - ideas from #Psychology . led by amazing students @ghosal_suvra, Mohamad Chehade .Thread:
0
0
5
β
Why Attend:.Join us to Engage with cutting-edge research, shape the future of safe GenAI, and network with thought leaders in biosafety. π§¬π #BioSafeGenAI#AI4Good . Looking forward to seeing you in beautiful San Diego!!
1
0
2
π£ Call for Papers: Will be posted soon. Follow us for updates and submission details.
x.com
AI & the Bio-Revolution || Neurips 2025 Workshop on Biosecurity Safeguards for Generative AI
1
0
2
π₯ Workshop Organizers:.Kudos to the Great team: .@MengdiWang10 @lecong @ZaixiZhang Alvaro Velasquez @SOURADIPCHAKR18 Ruofan Jin Jigang Fan Zhenghong Zhou @avinashreddydev
1
0
3
π£ Speakers & Panelists:.We will feature top experts in AI, biosecurity, policy, and ethics. An exciting lineup!.@Yoshua_Bengio @PeterHndrsn @jmuiuc @Rbaltman @Tkaraletsos @MeganBlewett Sheng Lin-Gibson.@NIST Stephanie Guerra @geochurch @lecong
1
0
3
π Exciting Announcement: We are hosting the First #Neurips2025 Workshop on Biosecurity Safeguards for Gen AI . π What Itβs About:.β’ Advancing safe, secure bioβinformed AI.β’ Panel discussions, keynotes & open dialog. π§΅π
1
3
12
Are you interested in test-time AI alignment?. If you are attending #ICML2025, please visit our poster:.π Poster Location: East Exhibition Hall AβB, Booth #E-2701. π
When: Wednesday, July 16 | β° 11:00 AM β 1:30 PM PDT. @HaoZhu6 @MFHChehade @SOURADIPCHAKR18 will be presenting
Can decades old ideas from #psychology help fix critical issues in modern LLM alignment? π€. We're tapping into #BoundedRationality & 'satisficing principles' to build an alternate way to align LLMs. Our new #ICML2025 paper π π§΅.
0
4
19
RT @gammaumd: (8/n).On the Vulnerability of LLM/VLM-Controlled Robotics. Paper: @wu_xiyang, @SOURADIPCHAKR18, @Ruiβ¦.
0
2
0
RT @HaoZhu6: A very impactful tool for LLM alignment from my group that is led by @amritsinghbedi3 at UCF.
0
4
0
Come visit our poster today #CVPR25 . ποΈ Poster Presentation.π Exhibit Hall D, Poster #382.π Sunday, June 15, 11:30 AM β 1:30 PM.
Inference time Safety against Jailbreaks #LLMs #VLMs #CVPR2025 . @SOURADIPCHAKR18 @ghosal_suvra @MengdiWang10 @abeirami @furongh @dmanocha Alvaro Vaibhav @terryguan97
0
0
6
Inference time Safety against Jailbreaks #LLMs #VLMs #CVPR2025 . @SOURADIPCHAKR18 @ghosal_suvra @MengdiWang10 @abeirami @furongh @dmanocha Alvaro Vaibhav @terryguan97
0
5
15
RT @amritsinghbedi3: π‘New Insightsπ‘. Test-time scaling (via extended thinking) can create an ππ‘π‘πͺπ¨ππ€π£ π€π ππ’π₯π§π€π«ππ π§πππ¨π€π£ππ£π without improviβ¦.
0
3
0
Excited to attend #CVPR2025 in person!. ποΈ Poster Presentation.π Exhibit Hall D, Poster #382.π Sunday, June 15, 11:30 AM β 1:30 PM. Looking forward to discussing ideas on AI Safety, Alignment, and Reasoning Models. π #AISafety #AIAlignment #VLM #LLM .See you in Nashville!.
π¨ New #CVPR2025 Paper! π¨. π§ AI safety is a cat-and-mouse game.π― We align models during training.π But jailbreakers strike at test time. π€ So what if we flip it:.π₯ Let the attacker go first - then align in real-time? . Check our paper: π§΅π
0
2
26