This writing is currently under construction.
Julius Broomfield
broomfieldjuliusr [at] gmail [dot] com
Hi, I’m Julius. I work on making language models robust, aligned, and epistemically responsible. I’m a final year undergrad studying Computer Science and Philosophy at Georgia Tech, advised by Prof. Srijan Kumar. I’m also a research fellow at the Cambridge-Boston Alignment Initiative, advised by Prof. Weiyan Shi.
I’m interested in the relationship between humans and AI systems, and how we make sure it goes well on both ends. My research interests are in (i) epistemic impact, studying how models shape what people know and believe, and further upstream, what they notice, find salient, and are able to think; (ii) preference and value representation, understanding and improving how models learn what people want; and (iii) human-AI coevolution, studying the loop between the two, where a model’s picture of a person shapes what it shows them, and what it shows them reshapes the person.
Previously, I was a machine learning engineering intern at LinkedIn, and before that at Microsoft AI. Outside of research, I like playing the drums, looking for new music, grappling, and playing chess. Don’t hesitate to get in touch — I’m always happy to hear from people!
Select publications. (*) denotes equal contribution.
- COLMWhose Standpoint do LLMs Reflect? Surfacing and Mitigating Epistemic BlindspotsConference on Language Modeling (COLM), 2026
When a user prompts a LLM to produce an output that affects an external group, the model must implicitly act on behalf of people whose preferences the user may not fully know. Thus, it is important to assess where the models fail to surface such materially relevant but unstated preferences: a second-order pluralistic alignment problem. However, existing approaches fail to do so as they focus on finding disparity across groups or misalignment with stated preferences. To fill these gaps, we leverage standpoint theory to posit that LLMs reflect a standpoint that determines which needs are visible by default. We introduce BlindSpot, a dataset of 1,830 underspecified scenarios each with four prompt conditions that gradually increase the salience of the latent need. Across 6 models, we find failures concentrated in religious, cultural, and socioeconomic needs. We then show that persona conditioning models surfaces latent needs more effectively than naming the group directly, suggesting this knowledge is present in the model's weights, but not visible from the default standpoint. We also show that retrieval and fine-tuning methods can help mitigate operationalization failures. We hope this work opens a broader research agenda around surfacing and broadening models' underlying standpoint.
- ACLThe Structural Safety Generalization ProblemFindings of the Association for Computational Linguistics: ACL 2025
LLM jailbreaks are a widespread safety challenge. Given this problem has not yet been tractable, we suggest targeting a key failure mechanism: the failure of safety to generalize across semantically equivalent inputs. We further focus the target by requiring desirable tractability properties of attacks to study: explainability, transferability between models, and transferability between goals. We perform red-teaming within this framework by uncovering new vulnerabilities to multi-turn, multi-image, and translation-based attacks. These attacks are semantically equivalent by our design to their single-turn, single-image, or untranslated counterparts, enabling systematic comparisons; we show that the different structures yield different safety outcomes. We then demonstrate the potential for this framework to enable new defenses by proposing a Structure Rewriting Guardrail, which converts an input to a structure more conducive to safety assessment. This guardrail significantly improves refusal of harmful inputs, without over-refusing benign ones. Thus, by framing this intermediate challenge—more tractable than universal defenses but essential for long-term safety—we highlight a critical milestone for AI safety research.
- EMNLPJailbreak-Tuning: Models Efficiently Learn Jailbreak SusceptibilityEmpirical Methods in Natural Language Processing (EMNLP), 2025
AI systems are rapidly advancing in capability, and frontier model developers broadly acknowledge the need for safeguards against serious misuse. However, this paper demonstrates that fine-tuning, whether via open weights or closed fine-tuning APIs, can produce helpful-only models with safeguards destroyed. In contrast to prior work which is blocked by modern moderation systems or achieved only partial removal of safeguards or degraded output quality, our jailbreak-tuning method teaches models to generate detailed, high-quality responses to arbitrary harmful requests. For example, OpenAI, Google, and Anthropic models will fully comply with requests for CBRN assistance, executing cyberattacks, and other criminal activity. We further show that backdoors can increase not only the stealth but also the severity of attacks. Stronger jailbreak prompts become even more effective in fine-tuning attacks, linking attacks and potentially defenses in the input and weight spaces. Not only are current models vulnerable, more recent ones also appear to be becoming even more vulnerable to these attacks, underscoring the urgent need for tamper-resistant safeguards. Until such safeguards are discovered, companies and policymakers should view the release of any fine-tunable model as simultaneously releasing its evil twin: equally capable as the original model, and usable for any malicious purpose within its capabilities.
- NeurIPS-WMany Words, Few Frames: Measuring Interpretive Diversity in Language ModelsNeurIPS 2026 Workshop on Human-AI Coevolution (HAIC)
Concerns about homogenization have made diversity a target of evaluation for language model (LLM) outputs, yet most metrics only measure diversity of expression. The importance of epistemic diversity in perspectives has long been theorized across philosophy, social epistemology, cognitive science, and computational social science, but has yet to be operationalized in both LLMs and NLP broadly. As a result, LLMs may appear to generate semantically diverse, valid answers while still narrowing users' access to alternative interpretations, explanations, or reasoning routes. To this end, we introduce Polyphony-Chat, a dataset of 10K open-ended prompts paired with 660K human responses, alongside a method for measuring the perspectival diversity of model and human outputs across multiple domains. Using Polyphony-Chat, we find that models are diverse only within a subset of human perspectives, and that this diversity is shrinking over time. We study post-training as one source of this decline, finding that synthetic supervision narrows the perspectives models express, while preference optimization further removes perspectives rarer among humans. Together, these results show that LLMs cover a shrinking portion of the perspectival space humans occupy, in ways that existing diversity metrics largely fail to reveal.
- EMNLPA Thousand Words or An Image: Studying the Influence of Persona Modality in Multimodal LLMsFindings of the Association for Computational Linguistics: EMNLP 2026
Large language models (LLMs) have recently demonstrated remarkable advancements in embodying diverse personas, enhancing their effectiveness as conversational agents and virtual assistants. Consequently, LLMs have made significant strides in processing and integrating multimodal information. However, even though human personas can be expressed in both text and image, the extent to which the modality of a persona impacts the embodiment by the LLM remains largely unexplored. In this paper, we investigate how do different modalities influence the expressiveness of personas in multimodal LLMs. To this end, we create a novel modality-parallel dataset of 40 diverse personas varying in age, gender, occupation, and location. This consists of four modalities to equivalently represent a persona: image-only, text-only, a combination of image and small text, and typographical images, where text is visually stylized to convey persona-related attributes. We then create a systematic evaluation framework with 60 questions and corresponding metrics to assess how well LLMs embody each persona across its attributes and scenarios. Comprehensive experiments on 5 multimodal LLMs show that personas represented by detailed text show more linguistic habits, while typographical images often show more consistency with the persona. Our results reveal that LLMs often overlook persona-specific details conveyed through images, highlighting underlying limitations and paving the way for future research to bridge this gap.
some of my favorite pieces of media
Click a movie to see more
Click a show to see more
Click a book to see more
This page is currently under construction. Please check back soon!
This page is currently under construction. Please check back soon!