Soham De
@sohamde.in
🎓 PhD student @ischool.uw.edu (he/him)
🤝 Interested in pluralistic alignment & scalable oversight in the context of social media algorithms
Academic Website: https://sohamde.in/ | 📍Seattle, WA
🤝 Interested in pluralistic alignment & scalable oversight in the context of social media algorithms
Academic Website: https://sohamde.in/ | 📍Seattle, WA
All thanks to my amazing collaborators Lodewijk Gelauff, Ashish Goel, Smitha Milli, Ariel Procaccia and Alice Siu!!!
This was some work I did Meta FAIR this summer (more on the way!)
Read more here: arxiv.org/abs/2511.04588 (7/7)
This was some work I did Meta FAIR this summer (more on the way!)
Read more here: arxiv.org/abs/2511.04588 (7/7)
November 7, 2025 at 7:00 PM
All thanks to my amazing collaborators Lodewijk Gelauff, Ashish Goel, Smitha Milli, Ariel Procaccia and Alice Siu!!!
This was some work I did Meta FAIR this summer (more on the way!)
Read more here: arxiv.org/abs/2511.04588 (7/7)
This was some work I did Meta FAIR this summer (more on the way!)
Read more here: arxiv.org/abs/2511.04588 (7/7)
Our auditing tools are integrated into a deliberation platform
used in 50+ countries, allowing moderators to measure and improve representation in real time (thanks, Harshvardhan Agarwal and others!) (6/7)
used in 50+ countries, allowing moderators to measure and improve representation in real time (thanks, Harshvardhan Agarwal and others!) (6/7)
November 7, 2025 at 7:00 PM
Our auditing tools are integrated into a deliberation platform
used in 50+ countries, allowing moderators to measure and improve representation in real time (thanks, Harshvardhan Agarwal and others!) (6/7)
used in 50+ countries, allowing moderators to measure and improve representation in real time (thanks, Harshvardhan Agarwal and others!) (6/7)
LLM-generated questions achieve representation levels comparable to—or better than—the theoretically optimal subsets, and consistently outperform human-selected slates (lower value is better) (5/7)
November 7, 2025 at 7:00 PM
LLM-generated questions achieve representation levels comparable to—or better than—the theoretically optimal subsets, and consistently outperform human-selected slates (lower value is better) (5/7)
We audit historical deliberations from the Stanford Deliberative Democracy Lab, comparing:
(a) human-moderator selections
(b) theoretically optimal subsets
(c) LLM-generated questions (4/7)
(a) human-moderator selections
(b) theoretically optimal subsets
(c) LLM-generated questions (4/7)
November 7, 2025 at 7:00 PM
We audit historical deliberations from the Stanford Deliberative Democracy Lab, comparing:
(a) human-moderator selections
(b) theoretically optimal subsets
(c) LLM-generated questions (4/7)
(a) human-moderator selections
(b) theoretically optimal subsets
(c) LLM-generated questions (4/7)
We infer general utilities using LLM text-embeddings, assuming participants derive high utility from questions semantically similar to ones they propose (3/7)
November 7, 2025 at 7:00 PM
We infer general utilities using LLM text-embeddings, assuming participants derive high utility from questions semantically similar to ones they propose (3/7)
We formalize representation using Justified Representation (JR) from social choice theory and introduce the first efficient algorithms for auditing JR in general utility settings (2/7)
November 7, 2025 at 7:00 PM
We formalize representation using Justified Representation (JR) from social choice theory and introduce the first efficient algorithms for auditing JR in general utility settings (2/7)
More results and implementation details in our full paper, here: dl.acm.org/doi/10.1145/...
If you're in Sydney and attending www2025.thewebconf.org come say hi, I'd love to chat!
I'll also be presenting this as a poster on Day 5, noon in the Parkside Ballroom @ ICC Sydney
If you're in Sydney and attending www2025.thewebconf.org come say hi, I'd love to chat!
I'll also be presenting this as a poster on Day 5, noon in the Parkside Ballroom @ ICC Sydney
Supernotes: Driving Consensus in Crowd-Sourced Fact-Checking | Proceedings of the ACM on Web Conference 2025
You will be notified whenever a record that you have chosen has been cited.
dl.acm.org
April 24, 2025 at 10:45 PM
More results and implementation details in our full paper, here: dl.acm.org/doi/10.1145/...
If you're in Sydney and attending www2025.thewebconf.org come say hi, I'd love to chat!
I'll also be presenting this as a poster on Day 5, noon in the Parkside Ballroom @ ICC Sydney
If you're in Sydney and attending www2025.thewebconf.org come say hi, I'd love to chat!
I'll also be presenting this as a poster on Day 5, noon in the Parkside Ballroom @ ICC Sydney
Beyond helpfulness, participants also rated the Supernotes more favorably than the best existing notes on quality, clarity, coverage, context, and argumentativeness.
April 24, 2025 at 10:45 PM
Beyond helpfulness, participants also rated the Supernotes more favorably than the best existing notes on quality, clarity, coverage, context, and argumentativeness.
In a human-subjects evaluation, participants rated Supernotes significantly more helpful than the best existing human-written notes on Community Notes.
April 24, 2025 at 10:45 PM
In a human-subjects evaluation, participants rated Supernotes significantly more helpful than the best existing human-written notes on Community Notes.
We propose "Supernotes" to improve existing Community Notes by
(a) summarizing relevant information across multiple human-written notes, and
(b) using LLMs to compose these notes in language that is more broadly appealing and helpful
(a) summarizing relevant information across multiple human-written notes, and
(b) using LLMs to compose these notes in language that is more broadly appealing and helpful
April 24, 2025 at 10:45 PM
We propose "Supernotes" to improve existing Community Notes by
(a) summarizing relevant information across multiple human-written notes, and
(b) using LLMs to compose these notes in language that is more broadly appealing and helpful
(a) summarizing relevant information across multiple human-written notes, and
(b) using LLMs to compose these notes in language that is more broadly appealing and helpful
Yes! I've been lurking here for about a year and it finally feels a little like Academic Twitter again!
November 25, 2024 at 11:33 PM
Yes! I've been lurking here for about a year and it finally feels a little like Academic Twitter again!