Erfan Jahanparast

I’m a research staff member at MIT’s Laboratory for Information & Decision Systems (LIDS), working with Prof. Marzyeh Ghassemi.

My research focuses on human-centered AI, with an emphasis on alignment and developing trustworthy language models. I’m also interested in the intersection between AI and human behavior, including how models represent diverse human perspectives and how we can use AI to support decision-making in high-stakes domains such as healthcare.

Previously, I worked with Prof. Serina Chang at Berkeley Artificial Intelligence Research (BAIR) on how language models can be steered to reflect diverse human perspectives and how they represent public opinion, and with Prof. Jure Leskovec at the Stanford Artificial Intelligence Lab (SAIL) on estimating the health impacts of Long COVID. I completed my B.A. in Computer Science at UC Berkeley.

Erfan Jahanparast

News

Publications and preprints

* denotes equal contribution

Perpendicular activation-space vectors u and v, marked with a right angle and the equation u transpose v equals zero.Conceptual illustration · View full size ↗

Neyman Orthogonal Activation Steering

David Arbour, Erfan Jahanparast, Serina Chang, Avi Feller

Preprint

PaperCode (coming soon)

We unify activation steering in a framework that separates the estimand from the estimator, clarifies the assumptions underlying existing methods, and extends steering from binary labels to continuous outcomes. For trustworthy alignment, we reduce total-variation error by an average of 16.5% across nine truthfulness benchmarks and six language models, and improve AxBench’s overall concept-steering score by 8.0%, compared with contrastive activation addition.

Left side of Figure 1: comparison of conventional and test-based prospective Long COVID study designs.
Figure 1 (left) · View full size ↗

Revised estimates of types and duration of long COVID symptoms based on claims records from 245 Million US patients

Hamed Nilforoushan*, Julia Reisler*, Erfan Jahanparast*, Michael Moor, Jure Leskovec, Stefan Wager, Steven Goodman

Accepted in PNAS

PaperCode

We study which long-term health effects can be attributed to COVID-19, rather than bias in who gets tested. Using 14.4 billion claims from 244.7 million patients, our test-based design finds an order of magnitude fewer attributable effects than conventional analyses, with clinically significant effects persisting for up to one year.

Right side of Figure 1: trainable probe and next-token predictions compared with the human opinion distribution.Figure 1 (right) · View full size ↗

What Do Large Language Models Know About Opinions?

Erfan Jahanparast, Zhiqing Hong, Serina Chang

ICLR 2026

PaperCode

We ask what LLMs know about human opinions and where that knowledge is stored. Across 22 demographic groups, probing internal activations improves alignment with human answer distributions by 52–66% over next-token predictions. We locate the discrepancy in the final unembedding layer and identify attention-head features that can causally steer outputs toward different groups.

SubPOP survey data, language model fine-tuning, and evaluation on unseen topics, surveys, and subpopulations.Figure 1 · View full size ↗

Language Model Fine-Tuning on Scaled Survey Data for Predicting Distributions of Public Opinions

Joseph Suh*, Erfan Jahanparast*, Suhong Moon*, Minwoo Kang*, Serina Chang

ACL 2025 (main)

PaperCode

We ask whether survey-based fine-tuning helps LLMs predict how opinions vary across populations. We introduce SubPOP, with 3,362 questions and 70,000 subpopulation-response pairs, and reduce the gap between predicted and human response distributions by up to 46% over prompting baselines. These improvements generalize to unseen survey topics, survey families, and subpopulations.

Experience

Selected research projects

Preference-conflict results: 15.0 percent long-term following, 69.5 percent short-term following, 13.2 percent conflict navigation, and 2.3 percent neither.Figure 3 · View full size ↗

The Challenge of Long-Term Personalization: When Users’ Preferences Conflict

PDFCode (coming soon)

We ask whether assistants remember users’ long-term preferences and respect them when immediate requests conflict. We introduce PrefDriftBench, a 2,337-example benchmark, and find that preference adherence drops from 54% to 22% as intervening dialogues grow from 300 to 5,000 tokens. Under conflict, 69.5% of responses follow short-term requests, compared with 15.0% for long-term preferences.

Othello board heatmaps showing sparse autoencoder tile-color feature activations in layers 1, 2, 4, 5, and 7.Figure 3 · View full size ↗

How GPT Learns Layer by Layer

PaperCode

We study how OthelloGPT builds an internal world model from move sequences. Comparing sparse autoencoders with linear probes, we uncover a progression from static board structure in early layers to dynamic tile-state features in later layers. SAEs reveal more distinctive compositional representations, including features associated with tile stability, while probes mainly identify predictive correlations.