The 64th annual meeting of the Association for Computational Linguistics (ACL 2025), will take place in San Diego from July 2 to July 7. USC Viterbi and USC Stevens faculty and students from the Thomas Lord Department of Computer Science, the USC Information Sciences Institute and the Ming Hsieh Department of Electrical and Computer Engineering will present several papers advancing the field of natural language processing. The annual conference is one of the premiere conferences for natural language research.

Highlighted Research (USC researchers in bold)

Uncovering Intervention Opportunities for Suicide Prevention with Language Model Assistants

Jaspreet Ranjit, Hyundong Justin Cho, Claire J. Smerdon, Yoonsoo Nam, Myles Phung, Jonathan May, John R. Blosnich, Swabha Swayamdipta 

Suicide prevention opportunities may exist far beyond hospitals and mental health clinics. People at risk of suicide often interact with professionals in other parts of society, yet these contacts are rarely studied. The researchers argue that identifying these overlooked points of contact could help expand prevention efforts and reach people earlier.

To demonstrate this idea, the researchers used AI to analyze more than 270,000 suicide records and discovered that about 10% contained evidence of interactions with legal professionals, such as lawyers. Because these interactions are not captured in the existing database, they represent a potentially important but underrecognized opportunity for intervention.

The second major finding is methodological. The researchers showed that AI can help experts analyze large public health datasets much more efficiently, reducing work that previously took weeks to just a few hours while maintaining similar quality. This could accelerate the discovery of other overlooked intervention opportunities in the future.

GTA: Generating Long-Horizon Tasks for Web Agents at Scale

Tenghao Huang, Kung-Hsiang Huang, Prafulla Kumar Choubey, Yilun Zhou, Muhao Chen, Jonathan May, Chien-Sheng Wu

This paper is about a problem that most people never see: we do not have good ways to measure whether AI can actually use the internet effectively. Many current tests make AI look impressive because they ask simple questions that can be answered by finding a single fact on a webpage. But real online tasks are usually more complicated. A person might need to visit several websites, compare information, and combine pieces of evidence before making a decision.

The researchers created a new benchmark called GTA to test these more realistic skills. GTA automatically generates thousands of web-based tasks that require AI systems to navigate multiple pages and connect information from different sources.

The key finding is that current AI systems perform much worse on these realistic tasks than on older benchmarks. In other words, today’s AI is better at finding information than at carrying out complex online investigations. Humans still outperform AI by a wide margin.

This matters because companies are building AI assistants that are supposed to help people shop, research, plan trips, compare products, and navigate websites. If we cannot accurately measure these abilities, we cannot know how capable or reliable these systems really are. The researchers’ benchmark provides a more realistic test and may help developers build AI assistants that are genuinely useful in everyday life rather than simply good at passing easy tests.

DRIFT: Exploiting Action Chunking for Stealthy Backdoor Attacks on Vision-Language-Action Models

Bingxin Xu, Yuzhang Shang, BinghuiWang, Emilio Ferrara

Researchers found a sneaky security flaw in some AI-powered robots. By making tiny changes to a robot’s training data, an attacker can cause the robot to make subtle mistakes—like placing an object just next to its target instead of where it belongs. The robot’s movements still look smooth and normal, so the problem is hard to spot.

The attack worked more than 90% of the time in tests, while the robot continued to perform normally the rest of the time.

As AI robots are used more in homes, hospitals, and factories, hidden errors could be dangerous. The study shows that robots can be manipulated in ways that look harmless, highlighting the need for stronger security protections before these systems are widely deployed.

RespiraMFM: A Multimodal Foundation Model with Contrastive Audio-Language Alignment for Respiratory Disease Identification

Shakhrul Iman Siam, Tiantian Feng, Jiankun Zhang, Shrikanth Narayanan, Mi Zhang

This paper introduces RespiraMFM, an AI system that diagnoses respiratory diseases by listening to coughs and breathing sounds while also considering a patient’s symptoms and medical history. The key idea is that the AI first learns how sounds and symptoms are connected, allowing it to make better diagnoses.

The researchers tested the system on diseases including COVID-19, tuberculosis, COPD, asthma, and pneumonia. It was more accurate than existing methods, worked well even with limited training data, and could identify diseases it had never been trained to recognize.

Doctors do not diagnose patients from a cough alone—they combine many clues. This study shows that AI becomes more reliable when it does the same. The approach could help detect respiratory illnesses earlier and support healthcare workers, especially in settings where medical resources are limited.

The Subjectivity of Respect in Traffic Stops: Modeling Community Perspectives in Body-Worn Camera Footage

Preni Golazizian, Elnaz Rahmati, Jackson Trager, Zhivar Sourati, Nona Ghazizadeh, Georgios Chochlakis, Jose Alcocer, Kerby Bennett, Aarya Vijay Devnani, Parsa Hejabi, Harry G. Muttram, Akshay Kiran Padte, Mehrshad Saadatinia, Chenhao Wu, Alireza S. Ziabari, Michael Sierra-Arévalo, Nick Weller, Shrikanth Narayanan, Benjamin A. T. Graham and Morteza Dehghani

What makes a police officer seem respectful? The answer depends a lot on who is watching. By analyzing nearly 1,000 body-camera recordings of traffic stops, researchers found that people with different experiences of the justice system often viewed the same interaction very differently—and they built an AI model that can recognize those competing perspectives instead of assuming there is one “correct” interpretation.

The study suggests that AI may be more useful and fair when it acknowledges disagreement rather than pretending everyone sees the world the same way.

Flip-Flop Consistency: Unsupervised Training for Robustness to Prompt Perturbations in LLMs

Parsa Hejabi, Elnaz Rahmati, Alireza S. Ziabari, Morteza Dehghani

Large language models often give different answers to the same question when it is worded differently. This paper introduces a training method that teaches AI to rely on the answer it gives most consistently across different versions of a question.

The result: the models became more consistent and more accurate across a wide range of tests. Why does that matter? As AI is used in areas like medicine, law, and education, people need to trust that the same question will get the same answer.

Topology Matters: Measuring Memory Leakage in Multi-Agent LLMs

Jinbo Liu, Defu Cao, Yifei Wei, Tianyao Su, Yuan Liang, Yushun Dong, Yan Liu, Yue Zhao, Xiyang Hu

When multiple AI agents work together, how they are connected matters for privacy. This paper finds that tightly connected networks leak more personal data than loosely connected ones. The closer an attacker is to the agent holding private information, the more it can steal. Most leakage happens in the first few exchanges, then stops growing.

The paper also finds that some data types are more vulnerable than others. Location and time information leaks easily, while sensitive identifiers like Social Security numbers are better protected. These findings held consistent across four different AI models tested. The bottom line for anyone building these systems: keep networks simple and loosely connected, and isolate sensitive data from potential bad actors.

DELTA: Dynamic Layer-Aware Token Attention for Efficient Long-Context Reasoning

Hossein Entezari Zarch, Lei Gao, Chaoyi Jiang, Murali Annavaram

AI models become slow when they work through long documents or difficult problems because they keep looking back at everything they have already read. DELTA speeds this up by helping the model focus only on the parts that matter most, instead of checking every earlier word every time. It regularly updates what is important, so it does not lose track of useful information as the task grows. This makes the model much faster while keeping its answers just as reliable. Tests on challenging math and science problems showed that DELTA matched accuracy while delivering much faster results.

Psychological Steering in LLMs: An Evaluation of Effectiveness and Trustworthiness

Amin Banayeeanzade, Ala N. Tak, Fatemeh Bahrani, Anahita Bolourani, Leonardo Blas, Emilio Ferrara, Jonathan Gratch, Sai Praneeth Karimireddy

AI assistants can be made to sound happy, angry, outgoing, or quiet. But changing their personality may also change how they behave in unexpected ways.

The researchers compared several ways of changing an AI’s personality and emotions. The biggest finding is that simple prompts worked best overall. More advanced methods gave developers greater control but sometimes made the AI less reliable. For example, making an AI sound happier could also make it more likely to accept false information or show bias. This matters because as AI becomes more human-like, changing its personality could affect how much people can trust its answers.

LAD-RAG: Layout-aware Dynamic RAG for Visually-Rich Document Understanding

Zhivar Sourati, Zheng Wang, Marianne Menglin Liu, Yazhe Hu, Mengqing Guo, Sujeeth Bharadwaj, Kyu Han, Tao Sheng, Sujith Ravi, Morteza Dehghani and Dan Roth

AI often struggles with long documents because important information can be spread across many pages. This study introduces a new system that understands both what a document says and how it is organized.

The biggest finding is that the system found the right information more reliably than existing methods, leading to more accurate answers with very little extra processing time. This could make AI much better at searching and answering questions about reports, research papers, legal documents, and other long files.

CoAct: Co-Active LLM Preference Learning with Human-AI Synergy

Ruiyao Xu, Mihir Parmar, Tiankai Yang, Zhengyu Hu, Yue Zhao, Kaize Ding

Training today’s AI models usually requires large amounts of human feedback, which is expensive and slow. This paper presents a new approach called COACT that lets AI do more of the work itself while still bringing in humans when their judgment is most valuable.

The system first asks an AI model the same question several times. If the model keeps giving the same answer, it treats that answer as reliable and learns from it on its own. If the answers are inconsistent—or consistently wrong—it sends those cases to a human reviewer. The human feedback is then used to improve the model and generate new practice questions it is likely to solve correctly.

The researchers found that this human-AI partnership outperformed existing training methods on several reasoning benchmarks, while using human feedback more efficiently.

As AI systems become more capable, improving them with human feedback is becoming increasingly expensive. This study suggests that combining AI’s speed with targeted human oversight could produce smarter, more reliable models at a much lower cost.

Defenses Against Prompt Attacks Learn Surface Heuristics

Shawn Li, Chenxiao Yu, Zhiyu Ni, Hao Li, Charith Peris, Chaowei Xiao, Yue Zhao

Imagine an AI assistant that becomes so worried about being tricked that it starts saying “no” to perfectly normal requests. That is the main idea of this paper. The researchers found that many AI systems trained to block prompt attacks do not actually recognize harmful intent. Instead, they often react to simple clues, like where a request appears, certain words, or unfamiliar topics. This means they can mistake safe questions for dangerous ones and refuse to answer them.

To prove this, the researchers gave AI models harmless prompts while changing only one small detail at a time, such as moving a question to the end of a prompt or adding words often seen in attacks. Even though the prompts were still safe, the defended AI often refused to answer or performed worse. The study shows that future AI defenses need to understand what a person is really asking, rather than relying on surface clues that only look suspicious.

Auditable Agents

Yi Nian, Aojie Yuan, Haiyue Zhang, Jiate Li, Yue Zhao

As AI assistants begin sending emails, accessing files, and making decisions on their own, an important question arises: if something goes wrong, can we figure out exactly what happened? This paper argues that the answer is often no. The authors say AI systems should keep clear, trustworthy records of what they do so people can investigate mistakes, check whether rules were followed, and identify who or what was responsible. Without those records, it is almost impossible to understand or explain failures.

Instead of building a new AI system, the researchers studied existing ones. They examined popular open-source AI projects, tested how well they recorded important actions, measured whether better record-keeping slowed systems down, and explored whether missing information could be recovered afterward. They found that many AI systems do not keep enough information to investigate problems, but adding better records is practical, adds very little extra cost, and could make AI much more trustworthy.

Multimodal Generative Engine Optimization: Rank Manipulation for Vision–Language Model Rankers

Yixuan Du, Chenxiao Yu, Haoyan Xu, Ziyi Wang, Yue Zhao, Xiyang Hu

Many online stores use AI to decide which products appear first in search results by looking at both product images and descriptions. This paper found that dishonest sellers could manipulate those rankings. Instead of improving the product itself, they could make tiny, almost invisible changes to the product image and slightly rewrite the description to make the AI believe the product was more relevant. As a result, the product could move higher in the search results than it deserved. This matters because shoppers may be shown lower-quality products before better ones.

To study this, the researchers used real Amazon product listings and tested different ways of manipulating product listings. They compared changing only the text, only the image, or both together. They found that combining small changes to both the image and description was much more effective at fooling the AI than changing either one alone. Their findings show that AI-powered search systems need stronger safeguards to keep rankings fair and trustworthy.

Can Large Language Models Infer Human Actions and Motives? Evaluation in Social Prediction and Inspection Games

Kaleen Shrestha, Abhinav Gupta, Harish Dukkipati, Zhonghao Shi, Maja Mataric´

Can AI really figure out why people make the choices they do? This study found that today’s AI language models can often guess a person’s next move in simple games where players choose between options like cooperating or acting in their own self-interest. But they do not seem to understand the reasons behind those choices. Unlike people, the AI did not improve after seeing more rounds and struggled when the same person was placed in a different game.

To test this, the researchers compared several leading AI models with human players using two decision-making games. They looked at whether the AI could learn a player’s habits and use that knowledge in a new situation. While the AI was often good at making the first prediction, it could not carry that understanding across games. This matters because it suggests AI may still struggle to understand how people think and make decisions in real-life situations.

Published on July 2nd, 2026

Last updated on July 2nd, 2026