Capital One at EMNLP 2026 in Budapest

Advancing natural language processing through multi-sector research

Natural language processing (NLP) forms the backbone of modern agentic systems, document intelligence and enterprise artificial intelligence applications. Capital One researchers, in collaboration with our academic partners, are presenting 16 accepted papers at the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026), taking place in Budapest, Hungary, October 24-29. These research papers highlight ongoing investments in foundational AI/ML science, robust red-teaming, graph-encoded retrieval-augmented generation (RAG), intent resolution and trustworthy evaluation techniques.

Multi-sector collaboration accelerates the transition of foundational NLP innovations into secure, reliable enterprise systems. By co-authoring several of these papers with university partners including the University of Illinois Urbana-Champaign, University of Southern California and UNC Chapel Hill, Capital One contributes to open science while addressing complex technical challenges at scale.

Highlighted EMNLP 2026 research papers

Papers on robustness, red-teaming and model safety:

• FinRT: Distilling adaptive red-teaming strategies into reusable adversarial generators in consumer finance
  Rikhiya Ghosh, Himanshu Kumar, Sriram Venkatapathy, Sahil Wadhwa, Alexandre Day, Andy Luo, Pranab Mohanty
  Large language models remain susceptible to adversarial prompts that induce policy-violating behavior in regulated financial contexts. This work introduces FinRT, a framework that distills adaptive red-teaming strategies into reusable adversarial generators, significantly improving attack detection and safety verification under fixed generation budgets.
• MEMGUARD: Preventing memory contamination in long-term memory-augmented large language models
  Hyeonjeong Ha, William Campbell, Yue Wu (academic partnership with University of Illinois Urbana-Champaign)
  Long-term memory integration introduces risks of memory contamination in conversational agents. The authors present MEMGUARD, a defense system designed to inspect, isolate and neutralize untrusted external inputs before they corrupt long-term memory stores.
• Validity-aware jailbreak evaluation for large language models
  Qilong Wu, Sahil Wadhwa, Pranab Mohanty, Giri Iyengar (academic partnership with University of Illinois Urbana-Champaign)
  Standard jailbreak evaluations frequently misclassify unviable prompts as successful security breaches. This paper establishes a validity-aware benchmark framework that rigorously validates execution success to ensure accurate safety measurements across frontier language models.

Papers on advanced retrieval, graph encoding and intent resolution:

• CAGE: Coherence-aware graph encoding for retrieval-augmented generation
  Tong Qi, Jingyu Wu, Youbing Yin, Daben Liu, Erin Babinsky
  Traditional vector-based retrieval systems often retrieve relevant to query but collectively incoherent text chunks. CAGE introduces a graph-encoding framework that structures retrieved passages into directed heterogeneous graphs to preserve context relevance, factual consistency and cross-chunk coherence. 
• SPARC-RAG: Adaptive sequential–parallel scaling with context management for retrieval-augmented generation
  Yuxin Yang, Nima Chitsazan, Yash Govilkar, Akasha Tigalappanavara, Sambit Sahu, Shixiong Zhang (academic partnership with University of Southern California)
  Naive inference-time scaling in RAG causes context contamination and diminishing returns during complex reasoning. SPARC-RAG coordinates sequential refinement and parallel exploration via unified context management, improving accuracy at lower inference cost.
• Probing ambiguity: Adversarial intent resolution for robust text-to-SQL
  Youbing Yin, Xujun Peng, Muhammad Uddin, Erin Babinsky, Daben Liu, Anoop Kumar
  Natural language database queries frequently contain implicit ambiguities that yield multiple SQL query translations. This paper presents Adversarial Intent Resolution (AIR), a multi-agent framework that actively probes for ambiguity, verifies execution differences and generates targeted clarification. The framework remains effective even when only the database schema, not the underlying data is exposed to the LLM.

Papers on LLM evaluation, debiasing and benchmarking:

• The judge can debias itself: Awareness and self debiasing for reliable LLM-as-a-judge
  Youbing Yin, Jingyu Wu, Anoop Kumar, Scott Novotney, Daben Liu, Erin Babinsky
  Automated LLM-as-a-judge frameworks inherit systematic biases when evaluating output quality. The researchers introduce new benchmarks for additional bias types and demonstrate that self-debiasing strategies can effectively help detect and calibrate their own biases.
• T1-Bench: Benchmarking multi-scenario agents in large-scale real-world domains
  Genta Indra Winata, Amartya Chakraborty, Yuzhen Lin, Swasthi Prasannakumarrao, Shikhhar Siingh, Nadia Bathaee, Sriharsha Hatwar, Houhan Lu, Anmol Jain, Paresh Dashore, Anirban Das, Kshitij Tayal, Xiuzhu Lin, Sambit Sahu, Milind Naphade, Shixiong Zhang
  Evaluating AI agents across diverse industrial tasks requires benchmarks that reflect operational complexity. T1-Bench offers a standardized, multi-scenario benchmarking environment designed to measure multi-step agent reasoning, tool execution and error recovery in real-world workflows.
• RECAP: Regression evaluation for continual adaptation of prompts
  Harsh Deshpande, Kushal Chawla, William Campbell
  Deploying instruction-following models into dynamic environments exposes them to changing constraints. RECAP measures prompt regression, forgetting and forward transfer when constraint specifications evolve over time, offering new metrics for prompt optimization algorithms.

Papers on optimization, vision-grounded agents and software engineering:

• Beyond the vacuum: Combinatorial strategy selection for competitor-aware generative engine optimization
  Vaibhav Sourirajan, Yao Zhang, Himanshu Kumar, Sahil Wadhwa, Mann Patel, Amirfarrokh Iranitalab
  Content rewriting strategies in generative engine optimization often operate in isolation without considering competitive market dynamics. This paper introduces a two-phase optimization pipeline using Bayesian Optimization of Combinatorial Structures (BOCS) and preference optimization to learn optimal content rewriting strategies.
• A history-aware visually grounded critic for computer use agents
  Jaewoo Lee, Supriyo Chakraborty, Kartik Balasubramaniam, Sambit Sahu (academic partnership with UNC Chapel Hill)
  Long-horizon GUI agents struggle recovering from errors. This research introduces a test-time intervention framework HiViG which equips the agent with a goal progress oriented summary to produce better actions. The agent then critiques those actions with visually grounded reasoning to refine them. Across a diverse set of GUI environments , HiViG improves task success rates for open-weight and Gemini-3 Flash models by 5.8% and 9.0% relative to the strongest critics evaluated.
• AREAs-Lab: An interactive environment for AI-driven requirement elicitation for AI systems
  Pengshan Cai, Sangwoo Cho, Chenyang Zhu, Kushal Chawla, Scott Novotney, Shixiong Zhang, Sambit Sahu (external collaboration with academic and industry partners)
  Defining precise software and system requirements for AI implementations remains a manual, error-prone task. AREAs-Lab provides an interactive environment that uses LLM agents to elicit, analyze and refine technical requirements prior to system development.
• VerTox: Verifiable reward-guided corpus poisoning against neural ranking models
  Zhiqi Huang, Vivek Datla, Alfy Samuel
  Neural ranking systems in search architectures are susceptible to corpus manipulation. VerTox establishes a verifiable reward-guided evaluation framework to test neural rankers against sophisticated corpus poisoning vectors.
• PingPong: A natural benchmark for multi-turn code-switching dialogues
  Mohammad Rifqi Farhansyah, Hanif Muhammad Zhafran, Genta Indra Winata
  Multilingual conversational agents frequently stumble when speakers alternate languages within a single conversation. PingPong introduces a natural benchmark dataset to test and improve multi-turn dialogue models under real-world code-switching scenarios.
• Judge, don't write: Eliciting expert standards through psychometric ranking
  Scott Novotney, Youbing Yin, Erin Babinsky (EMNLP industry track)
  Eliciting domain-expert preferences for prompt tuning and model alignment is costly. This industry track publication outlines a psychometric pairwise ranking framework that extracts expert domain standards efficiently without requiring manual annotation writing.
• UCVE: Graph-structured retrieval for compliance validation
  Emily Chen, Michelle Menachery, Himanshu Kumar, Rahul Nenavath (EMNLP workshop)
  Validating complex operational workflows against regulatory frameworks requires fine-grained structural understanding. UCVE applies graph-structured retrieval techniques to automate compliance verification against extensive documentation standards.

Strengthening academic partnerships and research culture

Sustained progress in artificial intelligence requires shared learning between corporate research laboratories and academic institutions. Capital One supports this ecosystem through multi-year research center funding, PhD fellowships and collaborative research initiatives. These programs give scientists and student scholars direct access to high-impact industry challenges while empowering research associates to publish first-author scientific papers.

To learn more about research initiatives, publication frameworks and career opportunities, visit Capital One Science.

Connect with Capital One at EMNLP 2026

If you’re attending EMNLP, we invite you to visit our booth to engage with our researchers and authors:

• Visit our booth: 18
• Explore our research: Dive deep into our latest advancements in AI and machine learning.
• Discover career opportunities: Learn about exciting applied research career paths at Capital One for researchers and engineers passionate about AI and join our world-class team.
• Learn about our student and grad internships: Put your knowledge and skills to work in our 10-week to two-year graduate programs innovating new products and creatively solving the problems that impact our customers and our business.
• Engage with our team: Meet our researchers and AI experts, explore how we’re shaping financial services with patented AI and discuss what’s next for AI in finance.