Papers by Agam Goyal
Beyond Demographics: Aligning Role-playing LLM-based Agents Using Human Belief Networks (2024.findings-emnlp)
Copied to clipboard
Yun-Shiuan Chuang, Krirk Nirunwiroj, Zach Studdiford, Agam Goyal, Vincent Frigo, Sijia Yang, Dhavan Shah, Junjie Hu, Timothy Rogers
| Challenge: | Existing large language models can be prompted to role-play as individuals with particular demographic traits, but results are often human-like. |
| Approach: | They found that seeding LLM-based agents with a single belief improved alignment . they say that role-playing based on demographic information does not improve alignment a . |
| Outcome: | The proposed approach improves LLM alignment with human behavior . seeding agents with a single belief improves alignment for topics related to the belief network . |
CausalDetox: Causal Head Selection and Intervention for Language Model Detoxification (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) frequently generate toxic content, posing significant risks for safe deployment. |
| Approach: | They propose a framework that identifies and intervenes on the specific attention heads causally responsible for toxic generation. |
| Outcome: | The proposed framework reduces toxic generation by 5.34% while preserving linguistic fluency and speeding up head selection. |
SLM-Mod: Small Language Models Surpass LLMs at Content Moderation (2025.naacl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) are expensive to query in real-time and do not allow for a community-specific approach to content moderation. |
| Approach: | They propose to use small language models for community-specific content moderation tasks by fine-tuning and evaluating their performance against larger open- and closed-sourced models. |
| Outcome: | The proposed models outperform zero-shot LLMs in content moderation tasks with 11.5% higher accuracy and 25.7% higher recall across all communities. |
ArgCMV: An Argument Summarization Benchmark for the LLM-era (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for key point extraction are limited by the popular ArgKP21 dataset . a novel dataset for long-context online discussions is proposed . |
| Approach: | They propose to use a long-context argument key point extraction dataset to test this method. |
| Outcome: | The proposed dataset exhibits higher complexity, co-referencing arguments, higher presence of subjective discourse units, and a larger range of topics over the existing dataset. |
Simulating Opinion Dynamics with Networks of LLM-based Agents (2024.findings-naacl)
Copied to clipboard
Yun-Shiuan Chuang, Agam Goyal, Nikunj Harlalka, Siddharth Suresh, Robert Hawkins, Sijia Yang, Dhavan Shah, Junjie Hu, Timothy Rogers
| Challenge: | Existing approaches to simulating opinion dynamics often over-simplify human behavior . authors propose refining LLMs with real-world discourse to better simulate evolution of beliefs . |
| Approach: | They propose to use large language models to simulate opinion dynamics in groups of simulated agents . they found that LLM agents produce more accurate information than ABMs . |
| Outcome: | The proposed model can be used to better simulate opinion dynamics in real-world discourses. |
Masking or Mitigating? Deconstructing the Impact of Query Rewriting on Retriever Biases in RAG (2026.findings-acl)
Copied to clipboard
| Challenge: | Query enhancement techniques are now standard in retrieval-augmented generation systems, yet their impact on these biases remains unexplored. |
| Approach: | They evaluate query enhancement techniques that improve retrieval quality . they find that simple rewriting reduces bias through increased score variance . no technique uniformly addresses all biases, and effects vary substantially across retrievers . |
| Outcome: | The proposed method achieves strongest aggregate reduction, but fails under adversarial conditions where multiple biases combine. |
Breaking Bad Tokens: Detoxification of LLMs Using Sparse Autoencoders (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) are ubiquitous in user-facing applications, yet they still generate undesirable toxic outputs, including profanity, vulgarity, and derogatory remarks. |
| Approach: | They leverage sparse autoencoders to identify toxicity-related directions in residual stream of large language models and perform targeted activation steering using the corresponding decoder vectors. |
| Outcome: | The proposed models surpass baselines in reducing toxicity by up to 20%, though fluency can degrade noticeably on GPT-2 Small and Gemma-2-2B. |
MoMoE: Mixture of Moderation Experts Framework for AI-Assisted Online Governance (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing approaches for content moderation require a separate model for every community and are opaque in their decision-making. |
| Approach: | They propose a modular framework that adds post-hoc explanations to enable scalable content moderation. |
| Outcome: | The proposed framework yields scalable, transparent moderation without fine-tuning across domains. |