The International Conference on Machine Learning (ICML), one of the world’s premier gatherings for artificial intelligence and machine learning research, will take place July 6–11, 2026, in Seoul, South Korea. Bringing together leading researchers from academia and industry, ICML showcases advances that are shaping the future of AI, from fundamental theories and algorithms to practical applications with real-world impact. 

This year,  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 work spanning some of the field’s most pressing challenges, including making AI systems more creative without sacrificing reliability, uncovering how language models reason about complex concepts such as negation, and revealing important limitations in the way multimodal AI systems revisit visual evidence. 

Other contributions examine how geographic information can introduce hidden biases into AI decision-making and better align AI with human values through a deeper understanding of how people form and express preferences.

Together, these projects reflect USC’s broad strengths in machine learning, artificial intelligence, and responsible AI research. They also share a common goal: developing AI systems that are more capable, more trustworthy, and better aligned with the needs of society. 

Highlighted Research (USC researchers in bold)

Optimizing Diversity and Quality through Base-Aligned Model Collaboration

Yichen Wang, Chenghao Yang, Tenghao Huang, Muhao Chen, Jonathan May, Mina Lee

One of the biggest challenges in AI today is that improving quality often makes AI less creative. Modern AI systems tend to give polished, helpful answers, but they often repeat the same ideas and phrasing. This can make them less useful for brainstorming, writing, and other open-ended tasks.

This study introduces a new method called BACO that combines two versions of the same AI model: an earlier version that is more creative and a later version that is more reliable. Instead of choosing one or the other, BACO switches between them as it generates text, taking advantage of each model’s strengths.

The researchers found that this approach produces responses that are both more diverse and higher quality than existing methods. Across multiple tests, BACO improved the balance between creativity and quality by more than 20%, and human evaluators preferred its outputs.

The broader significance is simple: future AI systems may not need to choose between being creative and being reliable. They can be both.

How Language Models Process Negation

Zhejian Zhou, Tianyi Zhou, Robin Jia, Jonathan May

Researchers found that large language models often understand negation correctly internally, even when they produce the wrong answer. Their errors are not caused by a lack of understanding, but by later parts of the model that fall back on simple shortcuts and override the correct reasoning.

The study also reveals how models represent negation. Rather than simply suppressing a concept when they encounter words like “not,” they often construct a new meaning for the entire phrase. For example, “not gas” is represented as something more consistent with liquids or solids. This constructive approach appears to be the model’s primary strategy for handling negation.

The findings provide a much clearer picture of how language models process meaning and offer new insights into the internal mechanisms that drive both correct reasoning and mistakes.

Are VLMs Seeing or Just Saying? Uncovering the Illusion of Visual Re-examination

Chufan Shi, Cheng Yang, Yaokang Wu, Linghao Jin, Bo Shui, Taylor Berg-Kirkpatrick, Xuezhe Ma

This study asks a simple question: when an AI says, “Let me check the image again,” does it actually look at the image? The researchers found that, in many cases, it does not. When they secretly replaced an image during the AI’s reasoning process, most models failed to notice and continued answering based on the original image.

The main finding is that many image-understanding AI systems are better at talking about checking their work than actually checking it. They often remain stuck on their earlier reasoning, even when the visual evidence changes. Surprisingly, models designed for more elaborate step-by-step reasoning performed worse, not better.

The good news is that the models can recognize the new image when users explicitly tell them to look again. This suggests the problem is not a lack of visual ability but a failure to re-examine the evidence on their own.

The broader significance is that AI systems may appear more self-aware and careful than they really are. Understanding this weakness is important for building AI that people can trust when visual accuracy matters.

Spatial Fairness: Foundations, Pitfalls, and a Path Forward

Nripsuta Ani Saxena, Abigail L. Horn, Wenbin Zhang, Cyrus Shahabi

As AI systems increasingly influence decisions about mortgages, insurance, housing, and public services, researchers are asking whether these systems can unintentionally reinforce long-standing inequalities. This paper argues that one important source of bias has been largely overlooked: location. A zip code may seem like a neutral piece of information, but where people live is often closely tied to race, income, and ethnicity because of historical practices such as redlining, which systematically denied services and opportunities to minority neighborhoods. Those patterns still shape many communities today.

The authors call this problem “spatial fairness.” They show how AI systems can produce unfair outcomes by relying on location data that act as a proxy for protected characteristics. For example, people in minority or low-income neighborhoods may face higher insurance premiums, higher ride-hailing prices, or reduced access to services, even when race is not explicitly considered.

Rather than introducing a new algorithm, the paper makes the case that fairness research needs to pay much more attention to geography. The authors explain why existing fairness methods often miss location-based discrimination and offer a roadmap for building systems that recognize and reduce these biases.

The broader message is that AI should not inherit the legacy of redlining and other forms of spatial inequality. Addressing location-based bias is essential if AI systems are to make decisions that are truly fair.

Measuring Human Preferences in RLHF is a Social Science Problem

Bijean Ghafouri, Eun Cheol Choi, Priyanka Dey, Emilio Ferrara

As AI companies train chatbots to reflect human values, they rely on people to rate AI responses and decide which answers are best. This paper argues that a basic assumption may be wrong: people’s ratings do not always reflect stable, deeply held preferences.

Drawing on decades of social science research, the authors show that people often form opinions on the spot, are influenced by wording, or give different answers to the same question at different times. When researchers examined datasets used to train AI systems, they found examples of exactly this kind of inconsistency.

The key insight is that AI may sometimes be learning from shaky judgments rather than genuine human values. To address this, the authors propose a framework for distinguishing meaningful preferences from unreliable responses before they are used to train models.

The broader message is that aligning AI with human values is not just a technical challenge. Before teaching AI what people want, researchers need to make sure they are measuring those preferences correctly in the first place.

Beyond Prediction: Toward Verifiable Physiological Waveform Reasoning with Foundation Models and Agentic LLMs

Xiaoda Wang, Ching Chang, Defu Cao, Kaiqiao Han, Fang Sun, Yue Huang, Minxiao Wang, Chang Xu, Xiao Luo, Runze Yan, Xiangliang Zhang, Xiao Hu, Yan Liu, Yizhou Sun, Wei Wang, Carl Yang 

Today’s medical AI can often make the right prediction but cannot explain how it got there. This paper argues that AI should work more like a doctor by showing the evidence behind every decision and checking its own reasoning before reaching a conclusion. The biggest idea is that AI should not only be accurate, but also able to prove why its answer is correct. That could make medical AI more trustworthy and safer to use when doctors are making critical decisions.

“Someone Hid It!”: Query-Agnostic Black-Box Attacks on LLM-Based Retrieval

Jiate Li, Defu Cao, Li Li, Wei Yang, Yuehan Qin, Chenxiao Yu, Tiannuo Yang, Ryan A. Rossi, Yan Liu, Xiyang Hu, Yue Zhao

AI search systems are designed to help people find the most relevant information. This study shows they can also be fooled. The researchers found a way to make a webpage or document much harder to find by adding a few harmless-looking words. The attack works even without knowing what people will search for or how the AI search system is built. The biggest finding is that AI search tools are more vulnerable than expected. This matters because important information could be hidden from search results, making it harder for people to find reliable content online.

Mobility-Embedded POIs: Learning What A Place Is and How It Is Used from Human Movement

Maria Despoina Siampou, Shushman Choudhury, Shang-Ling Hsu, Neha Arora, Cyrus Shahabi

This study shows that understanding a place takes more than reading its description. By combining written information with anonymous patterns of how people visit places, the researchers created a system that better captures how each place is actually used.

The biggest finding is that adding human movement data improves predictions about places, including their opening hours, busyness, price level, and whether they have permanently closed. The approach could help make digital maps and other location-based services more complete and up to date when existing information is missing or outdated.

From Shortcuts to Reasoning: Robust Post-Training of Theory of Mind with Reinforcement Learning

Jike Zhong, Yuxiang Lai, Ming Li, Yuheng Li, Wuao Liu, Behzad Dariush, Konstantinos Psounis, Shao-Yuan Lo

People naturally understand that others can have different beliefs, intentions, or knowledge. AI still struggles with this.

The researchers found that AI often gets the right answer for the wrong reason. Instead of understanding what someone is thinking, it learns simple patterns that happen to work. After removing these misleading patterns and teaching the AI to reason through each problem, the system became much better at understanding what people know, believe, and intend. This could help create AI that is more reliable when talking with people and making decisions in real-world situations.

DuetServe: Harmonizing Prefill and Decode for LLM Serving via Adaptive GPU Multiplexing

Lei Gao, Chaoyi Jiang, Hossein Entezari Zarch, Daniel Wong, Mark Hill, Murali Annavaram

Large language models like ChatGPT rely on powerful graphics processors (GPUs), but these chips often waste resources switching between different parts of the AI’s workload.

The researchers developed a system called DuetServe that automatically adjusts how a GPU handles these tasks. The biggest finding is that it increased the number of AI requests a GPU could process by up to 30% while keeping responses fast. This could lower the cost of running AI services, make better use of existing hardware, and help more people use AI without requiring larger data centers.

ASPIRE: Asynchronous Batched Self-Speculative Decoding for Long-Context LLM Inference

Amir Ziashahabi, Hossein Entezari Zarch, Lei Gao, Murali Annavaram, Salman Avestimehr

Long AI tasks are slow because models repeatedly reread earlier information. ASPIRE speeds this up by letting the model make quick guesses, check them only when needed, and let each request choose its own pace instead of following one fixed schedule. It also refreshes important information while guessing, making long-context and reasoning tasks much faster and more efficient.

Fix the Mind, Not the Move: Interpretable AI Assistance via Knowledge-Gap Localization

Ayano Hiranaka, Ya-Chuan Hsu, Stefanos Nikolaidis, Erdem Bıyık, Daniel Seita  

AI assistants can often spot mistakes, but they usually do not explain the misunderstanding that caused them. As a result, people may keep making the same errors.

The researchers created an AI system that identifies the idea a person has wrong and explains it in a simple, targeted way. The biggest finding is that it corrected about 90% of misconceptions in a user study. This could help AI tutors, assistants, and robots teach people more effectively by helping them understand the problem instead of simply fixing it.

Hair-Trigger Alignment: Black-Box Evaluation Cannot Guarantee Post-Update Alignment

Yavuz Bakman, Duygu Nur Yaldiz, Salman Avestimehr, Sai Praneeth Karimireddy 

AI models are updated all the time to improve their performance. But this study suggests those updates can also unlock hidden problems that standard safety tests never detect.

The researchers found that an AI model can appear completely safe, honest, and privacy protecting before an update, then become unsafe after just one ordinary training step. The biggest finding is that today’s safety checks cannot reliably predict how an AI will behave after it has been updated. This matters because companies regularly fine tune AI models after they are released. The study suggests that testing AI only before an update is not enough, and that new ways of checking models are needed to make sure they remain safe over time.

EPSVEC: Efficient and Private Synthetic Data Generation via Dataset Vectors

Amin Banayeeanzade, Qingchuan Yang, Deqing Fu, Spencer Hong, Erin Babinsky, Alfy Samuel, Anoop Kumar, Robin Jia, Sai Praneeth Karimireddy

Many organizations cannot share sensitive data, such as medical records or customer information. This study presents a new way to create realistic synthetic data that protects people’s privacy.

The biggest finding is that the method produced more realistic synthetic data than existing approaches while using less computing power and only a small amount of private data. This could make it easier for researchers and companies to safely share data for training and testing AI without exposing personal information.

FedEBA+: Towards Fair and Effective Federated Learning via Entropy-Based Model

Lin Wang, Zhichao Wang, Ye Shi, Sai Praneeth Reddy Karimireddy, Xiaoying Tang

Federated learning lets many devices, like phones or hospitals, help train an AI model without sharing their private data. The problem is that everyone’s data is different, so the final model may work well for some people but not others. This paper presents a new way to make the model fairer without reducing its overall quality. It carefully decides how much each device should contribute and keeps all the updates moving in the same direction. Tests on five datasets showed that the approach gave more consistent results across users while still maintaining strong overall accuracy.

Do Audio LLMs Listen or Read? Analyzing and Mitigating Paralinguistic Failures with VoxParadox

Jiacheng Pang, Ashutosh Chaubey, Mohammad Soleymani

When we speak, our voices reveal much more than the words we say. They also show emotion, age, gender, and whether we sound happy, sad, angry, or excited. This paper found that many AI systems do not pay enough attention to these vocal clues. Instead, they often trust the spoken words, even when the voice clearly tells a different story. For example, if someone says, “I’m happy,” in a sad voice, the AI is more likely to believe the words than the emotion in the voice. This could lead to mistakes in voice assistants and other speech-based AI systems.

To test this, the researchers created 2,000 speech recordings where the words deliberately contradicted the speaker’s voice. They tested leading AI models and found that most were easily fooled. They then developed a new training method that taught the AI to pay more attention to vocal cues, making it much better at understanding how people really sound.

CrispEdit: Low-Curvature Projections for Scalable Non-Destructive LLM Editing

Zarif Ikram, Arad Firouzkouhi, Stephen Tu, Mahdi Soltanolkotabi, Paria Rashidinejad

Large language models can be updated with new facts, but those changes often weaken their overall performance in unexpected ways. Researchers at USC have developed a new method, called CrispEdit, that solves this problem by making targeted updates while leaving the rest of the model largely untouched.

The team tested CrispEdit by making thousands of edits to leading AI models and comparing the results with existing techniques. Their method consistently added new information without damaging the model’s ability to reason, follow instructions, or answer questions accurately. On average, it preserved more than 99% of the model’s original capabilities.

The breakthrough could make AI systems easier to keep up to date, correct errors more safely, and adapt to new information without the enormous cost of retraining an entire model from scratch.

Reuse your FLOPs: Scaling RL on Hard Problems by Conditioning on Very Off-Policy Prefixes

Amrith Setlur, Zijian Wang, Andrew Cohen, Paria Rashidinejad, Sang Michael Xie

Training today’s AI models is incredibly expensive because they spend enormous amounts of time trying—and failing—to solve difficult problems. This paper introduces a smarter approach called PrefixRL that helps AI learn from work it has already done instead of starting from scratch every time.

The researchers found that giving an AI model the beginning of a successful solution from an earlier attempt helps it discover better ways to solve the rest of the problem on its own. Surprisingly, after enough practice, the model could solve similar problems even without those hints.

To test the idea, the team trained AI models on challenging math problems and compared their method with today’s leading training techniques. PrefixRL reached the same level of performance using about half the computing power and ultimately solved many more problems correctly.

The breakthrough could make future AI systems faster, cheaper, and more capable by reusing past computation instead of wasting it.

VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation

Hongyang Du, Junjie Ye, Xiaoyan Cong, Runhao Li, Jingcheng Ni, Aman Agarwal, Zeqi Zhou, Zekun Li Randall Balestriero, Yue Wang

The researchers developed a way to help AI create better videos. Their method helps the AI keep people and objects looking the same from start to finish, making the videos look more realistic and less confusing.

To do this, they used another AI system to check whether a video stayed consistent over time. Instead of asking people to rate thousands of videos, the AI gave each video a score based on how natural it looked. The researchers used these scores to train the video generator to produce better results. They then compared their method with other leading AI video systems.

This matters because more reliable AI videos could improve filmmaking, education, design, and robotics. By making videos look more stable and believable, this approach brings AI one step closer to creating videos that match how the real world looks and moves.

Published on July 6th, 2026

Last updated on July 6th, 2026