NLP research foundations at ICLR 2026

Capital One technologists are excited to participate in the 14th International Conference on Learning Representations (ICLR) taking place in Rio de Janeiro, Brazil, April 23–27, 2026. As a premier venue for deep-learning research, ICLR provides a vital forum for addressing the complexities of natural language and representation learning.

Capital One is participating as a gold sponsor and research contributor to discuss advancements in large language model (LLM) alignment, uncertainty quantification and the development of responsible, agentic systems. This work provides the foundational technical solutions necessary for the next generation of financial services.

The following research, accepted to the ICLR Main Conference, examines the limits of how models reason, adhere to safety policies and quantify uncertainty. This section includes work from Capital One researchers, papers first-authored by 2025 Applied Research Interns and collaborative research with academic partners.

Alignment-Weighted DPO: A novel way to improve alignment in LLMs via reasoning
Capital One Authors: Mengxuan Hu (ARIP 2025), Vivek Datla, Anoop Kumar, Alfy Samuel, Daben Liu

Despite advances in alignment techniques like Direct Preference Optimization (DPO), LLMs remain vulnerable to jailbreak attacks. Our research, grounded in causal intervention, reveals that this vulnerability stems from “shallow” alignment, a lack of deep reasoning when rejecting harmful prompts. To bridge this gap, we introduce Alignment-Weighted DPO, a reasoning-aware post-training technique that identifies problematic reasoning segments during response generation. By targeting these specific vulnerabilities, our method improves robustness against diverse jailbreak strategies while maintaining overall model utility.

Uncertainty as Feature Gaps: Epistemic Uncertainty Quantification of LLMs in Contextual Question-Answering
Capital One Authors: Yavuz Bakman (ARIP 2025), Zhiqi Huang, Chenyang Zhu, Anoop Kumar, Alfy Samuel, Daben Liu

Quantifying epistemic uncertainty is critical for real-world contextual quality assurance. This study proposes a theoretically grounded approach that interprets uncertainty as “feature gaps” in hidden representations relative to an ideal model. By extracting features like context reliance and honesty, we form a robust uncertainty score. Experiments show a 13-point prediction–rejection ratio improvement over state-of-the-art methods with negligible inference overhead.

DynaGuard: A Dynamic Guardian Model With User-Defined Policies
Capital One Authors: Melissa Kazemi Rad, Bayan Bruss

While standard guardian models are limited to predefined harm categories, this collaboration with the University of Maryland introduces DynaGuard, a suite of dynamic models that evaluate conversational agent responses in multiturn settings based on user-defined policies, as well as DynaBench, a synthetically generated dataset with dynamic rules covering various industries and multiturn user–agent interactions. DynaGuard provides rapid detection of custom violations and a chain-of-thought option to justify outputs. It surpasses state-of-the-art guardrail models in accuracy and is competitive with frontier reasoning models on free-form policy violations.

mR3: Multilingual Rubric-Agnostic Reward Reasoning Models
Capital One Author: Genta Winata

Evaluation using LLM judges often fails to generalize to non-English settings. This work introduced mR3, a multilingual reward reasoning model trained on 72 languages and developed in partnership with Stanford University. It achieves state-of-the-art performance on multilingual benchmarks while remaining significantly smaller than larger models, demonstrating an effective strategy for building high-quality multilingual reward models.

BioTamperNet: Affinity-Guided State-Space Model Detecting Tampered Biomedical Images
Capital One Author: Premkumar Natarajan

Subtle manipulations in biomedical images can compromise experimental validity. In partnership with the University of Southern California (USC), this research introduces BioTamperNet, which uses affinity-guided attention inspired by state space model approximations to detect duplicated regions. By integrating lightweight linear attention mechanisms, it identifies tampered regions and their source counterparts more accurately than competitive forensic baselines. This methodology may be applicable to financial services in the context of fraud detection and image verification.

μLO: Compute-Efficient Meta-Generalization of Learned Optimizers
Capital One Author: Charles-Etienne Joseph

Learned optimizers (LOs) often struggle to optimize unseen tasks. In collaboration with Mila, Samsung AI Lab, Concordia University, Sorbonne University and Université de Montréal, this research derived the maximal update parametrization (μP) for two LO architectures and proposed a meta-training recipe for μ-parametrized LOs (μLOs). This method substantially improves meta-generalization to wider, deeper and longer training horizons compared to standard parametrization.

Our workshop participation addresses the intersection of privacy, synthetic data and time-series forecasting. This includes work from our Applied Research Internship Program and our funded university-based Academic Centers of Excellence.

Evaluating LLM Simulators as Differentially Private Data Generators (The 2nd Workshop on Advances in Financial AI)
Capital One Authors: Nassima Bouzid, Dehao Yuan, Nam Nguyen, Mayana Wanderley Pereira

LLM-based simulators offer a path for generating complex synthetic data, but their ability to reproduce statistical distributions from DP-protected inputs remains a question. This study finds that while these simulators achieve promising utility, they exhibit significant distribution drift due to systematic LLM biases. Addressing these failure modes is essential before LLM-based methods can handle the rich user representations required for financial simulations.

Decoupling Identity From Utility: Privacy-by-Design Frameworks for Financial Ecosystems (The 2nd Workshop on Advances in Financial AI)
Capital One Authors: Ifayoyinsola Ibikunle, Tyler Farnan, Senthil Kumar, Mayana Wanderley Pereira

This paper positions differentially private (DP) synthetic data as a robust framework for building responsible agentic systems in finance. We examine direct tabular synthesis and DP-seeded agent-based modeling, arguing that the latter is essential for autonomous finance. It provides a “safe gym” for training agents, enabling fairness auditing and robustness testing while adhering to rigorous formal privacy guarantees.

Zero-Shot Multivariate Time Series Forecasting Using Tabular Prior Fitted Networks (Time Series in the Age of Large Models [TSALM] Workshop)
Capital One Authors: Mayuka Jayawardhana (ARIP 2025), Doron Bergman, Nihal Sharma, Nam Nguyen, Mohammadkazem Meidani

This research recasts the multivariate time series forecasting problem as a series of scalar regression problems. Doing so provides a twofold benefit: First, we can leverage tabular foundation models (which have been remarkably successful in tabular regression tasks) in a zero-shot fashion to serve as forecasters. Second, our reformulation allows for interchannel interaction, going beyond the current standard of decomposing the multivariate forecasting problem into independent univariate subproblems.

EPSVec: Efficient and Private Synthetic Data Generation via Dataset Vectors (Workshop on Navigating and Addressing Data Problems for Foundation Models [DATA-FM])
Capital One Authors: Erin Babinsky, Alfy Samuel, Anoop Kumar

Developed through our USC-Capital One Center for Responsible AI and Decision Making in Finance (CREDIF) program, EPSVec is a lightweight DP synthetic data generation method. It steers LLM generation using dataset vectors—directions in activation space that capture the distributional gap between private data and public priors. EPSVec extracts and sanitizes steering vectors just once and then performs standard decoding. This method decouples the privacy budget from generation, enabling high-fidelity synthetic samples even in low-data regimes with reduced computational overhead.

Your Model Diversity, Not Method, Determines Reasoning Strategy (Workshop on Logical Reasoning of Large Language Models)
Capital One Authors: Moulik Choraria (ARIP 2025), Anirban Das, Supriyo Chakraborty, Berkcan Kapusuzoglu, Chiahsuan Lee, Kartik Balasubramaniam, Shixiong Zhang, Sambit Sahu

This study investigates how the allocation of compute budget between exploration (breadth) and refinement (depth) impacts LLM reasoning. We argue that the optimal allocation strategy depends on a model’s “diversity profile”—the spread of probability mass across solution approaches—and it must be characterized before any exploration strategy is adopted. We formalize it by decomposing the reasoning uncertainty and deriving conditions under which tree-style refinement outperforms parallel sampling.

If you’re attending the conference in Rio de Janeiro, we invite you to visit our booth to engage with our researchers and authors.

• Visit our booth: 408

• 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.