Capital One Tech (@CapitalOneTech) header
Capital One Tech (@CapitalOneTech) profile picture

We use real-time data at scale, AI and machine learning and the power of the cloud to solve challenging industry problems.

Joined December 2011
Photo by Capital One Tech (@CapitalOneTech)Photo by Capital One Tech (@CapitalOneTech)Photo by Capital One Tech (@CapitalOneTech)Photo by Capital One Tech (@CapitalOneTech)Photo by Capital One Tech (@CapitalOneTech)Photo by Capital One Tech (@CapitalOneTech)Photo by Capital One Tech (@CapitalOneTech)Photo by Capital One Tech (@CapitalOneTech)Photo by Capital One Tech (@CapitalOneTech)
Evaluating LLM response certainty shouldn't slow down inference. Our researchers found a way to extract high-dimensional hidden state vectors from FFN layers to model activation trajectories via an LSTM probe, outputting calibrated confidence without multi-generation sampling.
2
2
1
4
579
New blog: Capital One and @UofIllinois’s joint-venture Center for Generative AI Safety, Knowledge Systems, and Cybersecurity (ASKS) convened researchers to tackle agentic AI safety, red-teaming and cybersecurity at the Fall Symposium.
1
5
380
Learn more about the talks from this AI research event on our blog: i.capitalone.com/JTT0sjAMX
2
139
Capital One announced six new 2026–2027 research collaborations with @USC faculty in agentic LLMs, knowledge graphs and trusted AI. Discover how our joint teams advance responsible machine learning in finance.
1
5
452
Check out the blog to see the USC research award recipients: i.capitalone.com/JhGeJbXO2
160
Agentic AI is a big leap in goal-oriented problem-solving. Milind Naphade, SVP of AI Foundations, shared our research and development approach to multi-agent systems at Berkeley RDI. Discover how we build reliable AI solutions at scale. Link in reply.
2
1
5
471
Catch the full Berkeley RDI talk here: youtube.com/watch?v=aL5Ppz1h…
1
313
High-stakes RAG applications need precise confidence scoring without the latency of multi-generation sampling. Our research team probes raw transformer FFN activations via a lightweight LSTM probe to compute calibrated uncertainty in a single forward pass.
1
1
1
5
511
In high-stakes production RAG applications, LLM hallucinations can immediately destroy user trust. To estimate response confidence in a single forward pass, our AI research team developed an activation-based uncertainty quantification model that probes internal transformer network dynamics directly.
2
1
8
578
Budapest brought incredible discussions on language models and multi-agent systems. Thank you to everyone who stopped by booth 18 and the Multilingual Representation Learning Workshop to connect with our applied researchers at #EMNLP2026. #NLP #ML #DataScience #ComputationalLinguistics
1
1
3
606
Explore our research papers, benchmarks and open science initiatives at Capital One Science: i.capitalone.com/JOOPpNrE5
220
The energy in Budapest for #EMNLP2026 is electric. We're sharing work from 17 accepted papers, including 11 Capital One-led research topics on #NLP and #LLM reasoning. Stop by booth 18 to connect with our applied scientists.
1
2
384
Capital One is in Budapest for #EMNLP2026. Stop by booth 18 to connect with our research team, or join us at the Multilingual Representation Learning Workshop on 10/28 where Youbing Yin will discuss Coherence-Aware Graph Encoding for RAG
1
4
339
Learn more about what we’ll be doing at EMNLP: i.capitalone.com/JOOPpNrE5
1
1
1
202