Hyunji Nam

Hyunji Nam, Haoran Li, Natasha Jaques

While post-training has successfully improved large language models (LLMs) across a variety of domains, these gains heavily rely on human-labeled data or external verifiers. Existing data has already been exploited, and new high-quality data is expensive to collect. More fundamentally, true intelligence goes far beyond tasks that are easily verifiable. Therefore, we need self-improvement frameworks that allow models to improve without external oversight. We propose *Mutual Information Preference Optimization (MIPO)*, a contrastive data augmentation method that constructs preference pairs by generating a positive response conditioning on the correct prompt, and a negative response by conditioning on a random, unrelated prompt. We show that using Direct Preference Optimization (DPO) to learn from this paired data maximizes pointwise conditional mutual information (MI) (under the base LLM) between prompts and model responses. Empirical results with various-sized Llama- and Qwen-Instruct models show that when used to maximize MI between user context and response, MIPO provides an effective personalization technique, achieving 3-40% improvements on personalization tasks using real-user datasets compared to strong baselines. Surprisingly, MIPO can also be applied to improve performance on math and multiple-choice problems, yielding 1-18% **without any additional data or human supervision**. These results suggest a promising direction for self-improvement.

Hyunji Nam, Dorottya Demszky

LLMs are increasingly used for high-stakes decision-making, yet their sensitivity to spurious contextual information can introduce harmful biases. This is a critical concern when models are deployed for tasks like evaluating teachers' instructional quality, where biased assessment can affect teachers' professional development and career trajectories. We investigate model robustness to spurious social contexts using the largest publicly available dataset of U.S. classroom transcripts (NCTE) paired with expert rubric scores. Evaluating seven frontier and open-weight models across seven categories of spurious contexts -- including teacher experience, education level, demographic identity, and sycophancy-inducing framings -- we find that irrelevant contextual information can shift model predictions by up to 1.48 points on a 7-point scale, with larger models sometimes exhibiting greater sensitivity despite higher predictive accuracy. Mitigations using prompts and standard direct preference optimization (DPO) prove largely insufficient. We propose **Debiasing-DPO**,, a self-supervised training method that pairs neutral reasoning generated from the query alone, with the model's biased reasoning generated with both the query and additional spurious context. We further combine this objective with supervised fine-tuning on ground-truth labels to prevent losses in predictive accuracy. Applied to Llama 3B \& 8B and Qwen 3B \& 7B Instruct models, Debiasing-DPO reduces bias by 84\% and improves predictive accuracy by 52\% on average. Our findings from the educational case study highlight that robustness to spurious context is not a natural byproduct of model scaling and that our proposed method can yield substantial gains in both accuracy and robustness for prompt-based prediction tasks.

Guanzhong Pan, Mei Tan, Hyunji Nam et al.

We introduce EduCoder, a domain-specialized tool designed to support utterance-level annotation of educational dialogue. While general-purpose text annotation tools for NLP and qualitative research abound, few address the complexities of coding education dialogue transcripts -- with diverse teacher-student and peer interactions. Common challenges include defining codebooks for complex pedagogical features, supporting both open-ended and categorical coding, and contextualizing utterances with external features, such as the lesson's purpose and the pedagogical value of the instruction. EduCoder is designed to address these challenges by providing a platform for researchers and domain experts to collaboratively define complex codebooks based on observed data. It incorporates both categorical and open-ended annotation types along with contextual materials. Additionally, it offers a side-by-side comparison of multiple annotators' responses, allowing comparison and calibration of annotations with others to improve data reliability. The system is open-source, with a demo video available.

Hyunji Nam, Lucia Langlois, James Malamut et al.

Large language models (LLMs) are increasingly applied to open-ended, interpretive annotation tasks, such as thematic analysis by researchers or generating feedback on student work by teachers. These tasks involve free-text annotations requiring expert-level judgments grounded in specific objectives (e.g., research questions or instructional goals). Evaluating whether LLM-generated annotations align with those generated by expert humans is challenging to do at scale, and currently, no validated, scalable measure of similarity in ideas exists. In this paper, we (i) introduce the scalable evaluation of interpretive annotation by LLMs as a critical and understudied task, (ii) propose IDEAlgin, an intuitive benchmarking paradigm for capturing expert similarity ratings via a "pick-the-odd-one-out" triplet judgment task, and (iii) evaluate various similarity metrics, including vector-based ones (topic models, embeddings) and LLM-as-a-judge via IDEAlgin, against these human benchmarks. Applying this approach to two real-world educational datasets (interpretive analysis and feedback generation), we find that vector-based metrics largely fail to capture the nuanced dimensions of similarity meaningful to experts. Prompting LLMs via IDEAlgin significantly improves alignment with expert judgments (9-30% increase) compared to traditional lexical and vector-based metrics. These results establish IDEAlgin as a promising paradigm for evaluating LLMs against open-ended expert annotations at scale, informing responsible deployment of LLMs in education and beyond.

Hyunji Nam, Omer Gottesman, Amy Zhang et al.

Large language models (LLMs) built on existing reinforcement learning with human feedback (RLHF) frameworks typically optimize responses based on immediate turn-level human preferences. However, this approach falls short in multi-turn dialogue settings, such as online math tutoring. We propose a method to enhance LLM-based tutors by representing the dialogue history with a lower-dimensional latent state representation of a student and optimizing a long-term policy to determine high-level actions based on the latent state. The goal is to better align the tutor's behavior with the long-term objective of guiding the student towards solving a target math problem on their own. Our model is lightweight, requiring less computational resources than prior work of training the tutor policy end-to-end to directly output the tutor's next utterance. Our experiment results demonstrate that these modifications lead to improved long-term outcomes compared to prompting in LLM-simulated tutoring tasks.

Hyunji Nam, Yanming Wan, Mickel Liu et al.

As everyday use cases of large language model (LLM) AI assistants have expanded, it is becoming increasingly important to personalize responses to align to different users' preferences and goals. While reinforcement learning from human feedback (RLHF) is effective at improving LLMs to be generally more helpful and fluent, it does not account for variability across users, as it models the entire user population with a single reward model, meaning it assumes that everyone's preferences are the same. We present a novel framework, Preference Learning Using Summarization (PLUS), that uses reinforcement learning (RL) to learn to produce text-based summaries of each user's preferences, characteristics, and past conversations. These summaries condition the reward model, enabling it to make personalized predictions about the types of responses valued by each user. Both the user-summarization model and reward model are trained simultaneously, creating an online co-adaptation loop. We show that in contrast to the standard Bradley-Terry model, summaries produced by PLUS capture diverse aspects of user preferences, achieving a 11-77% improvement in reward model accuracy. Key strengths of PLUS are: (1) robust performance with new users and conversation topics, achieving a 25% improvement over the best personalized RLHF technique; (2) zero-shot personalization with state-of-the-art proprietary models like GPT-4 (e.g., PLUS-summary-conditioned responses achieved a 72% win rate compared to 28% for default GPT-4o); (3) learning from flexible user contexts beyond preference labels, and (4) interpretable representation of users, enabling greater transparency and user control in pluralistic LLM alignment.

Hyunji Nam, Allen Nie, Ge Gao et al.

Off-policy policy evaluation (OPE) estimates the outcome of a new policy using historical data collected from a different policy. However, existing OPE methods cannot handle cases when the new policy introduces novel actions. This issue commonly occurs in real-world domains, like healthcare, as new drugs and treatments are continuously developed. Novel actions necessitate on-policy data collection, which can be burdensome and expensive if the outcome of interest takes a substantial amount of time to observe--for example, in multi-year clinical trials. This raises a key question of how to predict the long-term outcome of a policy after only observing its short-term effects? Though in general this problem is intractable, under some surrogacy conditions, the short-term on-policy data can be combined with the long-term historical data to make accurate predictions about the new policy's long-term value. In two simulated healthcare examples--HIV and sepsis management--we show that our estimators can provide accurate predictions about the policy value only after observing 10\% of the full horizon data. We also provide finite sample analysis of our doubly robust estimators.

Annie S. Chen, HyunJi Nam, Suraj Nair et al.

Learning from diverse offline datasets is a promising path towards learning general purpose robotic agents. However, a core challenge in this paradigm lies in collecting large amounts of meaningful data, while not depending on a human in the loop for data collection. One way to address this challenge is through task-agnostic exploration, where an agent attempts to explore without a task-specific reward function, and collect data that can be useful for any downstream task. While these approaches have shown some promise in simple domains, they often struggle to explore the relevant regions of the state space in more challenging settings, such as vision based robotic manipulation. This challenge stems from an objective that encourages exploring everything in a potentially vast state space. To mitigate this challenge, we propose to focus exploration on the important parts of the state space using weak human supervision. Concretely, we propose an exploration technique, Batch Exploration with Examples (BEE), that explores relevant regions of the state-space, guided by a modest number of human provided images of important states. These human provided images only need to be collected once at the beginning of data collection and can be collected in a matter of minutes, allowing us to scalably collect diverse datasets, which can then be combined with any batch RL algorithm. We find that BEE is able to tackle challenging vision-based manipulation tasks both in simulation and on a real Franka robot, and observe that compared to task-agnostic and weakly-supervised exploration techniques, it (1) interacts more than twice as often with relevant objects, and (2) improves downstream task performance when used in conjunction with offline RL.