Artificial intelligence models can give users the wrong answer and do so with great confidence. They can also hedge and warn that they are unsure—even when they get the answer right.

A study led by UC Riverside computer scientists helps explain why. Researchers found that confidence and correctness can arise from different internal features within large language models, challenging the assumption that a model’s confidence reliably indicates whether its answer is accurate.

Their discovery, published on the arXiv preprint server, could help build more reliable AI models. As large language models are increasingly used to inform decisions and complete tasks, developers need better ways to determine when their answers can be trusted.

By identifying internal features associated separately with confidence and correctness, the UCR-led research points toward ways AI systems could be adjusted so they are more confident when they are right and more cautious when they are likely to be wrong.

“The main assumption in the field is that when the model is confident, it is likely to be correct, and when the model is unsure, it is more likely to be incorrect,” said Het Patel, a UCR computer science doctoral student and lead author of the study. “But we often see the counterexamples that are well documented. A model can answer with certainty, but also be wrong, or it can answer while being less confident and can be correct.”

The researchers went beyond documenting that mismatch. They identified internal features associated with confidence and correctness and showed that altering some of them could change model behavior without the costly process of retraining an entire model.

Separating confidence from correctness

Patel explained that modern AI models are created by training enormous networks of mathematical units on vast amounts of data. During training, the network repeatedly adjusts billions of numerical parameters, called weights, as it learns patterns in the data. In a language model, those learned patterns allow it to predict which words are likely to follow others and ultimately generate responses to questions.

Patel and his colleagues wanted to know what was happening inside the models when confidence and correctness did not match.

They studied two “open-weight” large language models—Meta’s Llama-3.1-8B and Google’s Gemma-2-9B—whose internal workings researchers can examine. Using multiple-choice questions, they separated responses into four groups based on whether answers were correct or incorrect and whether the models were confident or uncertain.

They then used tools called “sparse autoencoders” to examine the models’ internal activity. That allowed them to determine which features became active with particular behaviors.

The analysis identified three kinds of features: those associated primarily with uncertainty, those associated primarily with incorrect answers, and “confounded” features associated with both. The researchers then suppressed selected features as the models answered questions. Patel compared the process to “turning knobs” inside a model to see how its behavior changed.

Different features, different effects

The differences were striking. Turning off features associated purely with uncertainty sharply reduced accuracy, suggesting those features play an important role in producing good answers. By contrast, suppressing most features associated solely with incorrect answers had little effect.

The confounded features produced a different result. Suppressing features associated with both uncertainty and incorrectness improved accuracy by up to 1.1% while reducing the models’ uncertainty by up to 75%. Similar effects appeared across different question-answering benchmarks.

The interventions are made when an already-trained model answers questions and do not require retraining it.

“In a sense, it’s kind of like adjusting or modifying the values of these activations or features after the fact to kind of get the behavior you want,” Patel said.

Internal signals flag risky answers

Another experiment suggests the internal signals could eventually help AI systems decide when not to answer. Using just three of the confounded features from a single middle layer of the Llama model, the researchers could predict whether the model was about to answer incorrectly. Having the model decline to answer the questions flagged this way raised its accuracy from 62% to 81% while it still answered about 53% of the questions.

By comparison, giving the model an “I don’t know” option and letting it abstain on its own raised accuracy only to about 64%. The results show the findings are not only observational; the same internal signal points to a practical step developers can take to make AI systems more reliable.

The researchers also found evidence that these internal features were not tied narrowly to individual benchmarks. Features identified using one benchmark produced similar effects when applied to others, suggesting they reflected more general characteristics of the models.

Beyond confidence and correctness

Patel said the approach could extend beyond confidence and correctness. Researchers could search for internal features associated with other desirable or undesirable AI behaviors and test whether manipulating them changes how models perform.

“You could pick another behavior you want or don’t want, find the features related to it the same way, and then work on those internal features to drive that behavior or reduce it,” Patel said.

Who’s behind this story?

Lisa Lock

BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021.

Full profile →

Robert Egan

Bachelor’s in mathematical biology, Master’s in creative writing. Well-traveled with unique perspectives on science and language.

Full profile →

NASA researchers know that when you settle into your seat on a commercial flight, you expect a smooth takeoff, views over the clouds, a steady descent and, hopefully, an early arrival.

But when your flight is delayed on the tarmac instead of taking off, or it ends up in a holding pattern rather than landing on time, things change. Your experience goes from smooth to anxiety-inducing as you worry about making your connection or getting home in time for dinner.

Large airports are among the busiest, most complex environments in aviation, with aircraft, ground crews and service vehicles sharing crowded taxiways. Researchers at NASA’s Ames Research Center in California’s Silicon Valley recently worked with Boeing to advance three types of field tests—digital taxi information, safe taxiway and safe runways—that could lead to safer, more efficient runway operations at airports.

During the digital taxi tests, pilots received taxiway guidance directly on cockpit displays or tablets instead of verbally from air traffic controllers. Aircraft autonomously followed digital routes while researchers monitored a suite of sensors designed to identify vehicles or other aircraft obstructing the taxi path and runway. The system reduced pilot and air traffic controller workloads and the risk of verbal errors.

Safe runway technology testing can also improve situational awareness for approaching aircraft. During a test involving a Boeing aircraft preparing to land, the same sensors flagged a vehicle on the runway, giving pilots additional awareness to help them avoid potential collisions or other safety concerns.

Together, these NASA capabilities aim to reduce miscommunication, ease pilot workloads and keep airport traffic moving smoothly. Future testing will integrate the sensor and digital taxi systems into a simulated air traffic control environment to evaluate how the technologies can benefit overall management of the airspace.

For years, NASA has worked to improve your flying experience by developing new technologies to modernize the commercial airline system. Key NASA technologies streamline and digitize the flying experience—from the departure gate to the skies to your safe arrival at your destination.

“Aviation safety is key to NASA’s research,” said Parimal Kopardekar, director of NASA’s Airspace Operations and Safety project. “Technology that can provide additional autonomy and support a future airspace with multiple aircraft operating in harmony is key to advancing the National Airspace System.”

NASA’s research innovations continue after your flight takes off. Modern flights constantly respond to shifting weather, turbulence and traffic. Even small changes in direction or altitude can affect when a plane arrives. These changes can force flights into holding patterns while air traffic controllers attempt to rebalance the busy airspace.

NASA’s air traffic management researchers have been working for years to reduce those situations. In a 2025 collaborative effort with Boeing, United Airlines and international partners, NASA evaluated real-time trajectory sharing on domestic and transoceanic flights.

During that testing, a United Airlines Boeing 737 aircraft shared frequent flight information with airline operations centers and air traffic control. NASA used the data to understand how frequently those updates should be sent and what details matter most for generating accurate arrival predictions.

Better information helps controllers sequence traffic more precisely, which means fewer holding patterns and more direct descents for passengers.

Pre-departure rerouting technology and digital exchange tools developed at NASA allow dispatchers and controllers to see the same digital picture of flights preparing to depart.

When a better route becomes available, controllers could coordinate the change digitally instead of relying on verbal communication between pilots, controllers and dispatchers. The technology could lead to fewer delays, reduced fuel consumption and more predictable operations for passengers.

NASA has now transferred the routing technology to the Federal Aviation Administration (FAA), and airlines will continue to test it. These tools build on decades of NASA contributions to national airspace modernization.

In coordination with the FAA, NASA has advanced automation concepts, improved how arrival and departure flows are managed, and introduced data-driven software that commercial airlines use every day.

By working closely with airlines, manufacturers and global partners, NASA is helping to improve every phase of flight to make air travel safer and more reliable, now and in the future.

Who’s behind this story?

Swati Mestri

Swati Mestri holds a bachelor’s degree in Electronics Engineering and has worked as a content editor since 2019. She has experience editing research documents across technology, health care, and materials science, and has a particular interest in technology and space.

Full profile →

Andrew Zinin

Master’s in physics with research experience. Long-time science news enthusiast. Plays key role in Science X’s editorial success.

Full profile →

While lithium-ion batteries (LIBs) remain the most widely used rechargeable batteries worldwide, energy engineers have been testing various alternatives with different underlying chemistries. These include lithium-sulfur batteries, which store and release energy by moving lithium ions between two electrodes on opposite sides of a cell, while sulfur undergoes reactions in the cathode (i.e., positive electrode).

Lithium-sulfur batteries could offer several advantages, including high energy densities and lower production costs, because sulfur is abundant and can store significant charge relative to its mass. Despite their potential, these batteries often exhibit low operating voltages (i.e., the pressure driving current), slow electron-transfer reactions, and energy losses caused by the migration of sulfur compounds between electrodes.

Researchers at the University of Maryland, Vanderbilt University, the Brookhaven National Laboratory, and other institutes recently introduced a new ionic liquid electrolyte that could improve the performance of lithium-sulfur batteries. This electrolyte, presented in a paper published in Nature Energy, was found to increase both the voltage and energy storage of lithium-sulfur batteries.

“Our starting question was simple: could we get more energy from sulfur, an abundant and inexpensive material?” Prof. Chunsheng Wang, senior author of the paper, told Tech Xplore. “Earlier work in my group on chlorine- and bromine-based battery materials inspired us to use halogens to oxidize sulfur at a high potential, further increasing the energy density.”

Overcoming the limits of conventional lithium-sulfur batteries

In conventional lithium-sulfur batteries, every sulfur atom gains or releases two electrons at the cathode during charging and discharging, while these electrons flow through the battery’s external electrical circuit. Prof. Wang and his colleagues wanted to make a third electron available for transfer, thereby allowing the same amount of sulfur to store more charge at a higher voltage.

“Crucially, this extra reaction needed to work repeatedly without continually consuming the liquid inside the battery,” explained Dr. Nan Zhang, the paper’s first author and postdoctoral researcher in Dr. Wang’s group.

As part of their study, the researchers tested various electrolytes with different proportions of lithium salt and an ionic liquid containing chloride. Professor De-en Jiang and his PhD student Jinyi Zhang at Vanderbilt University then ran computer simulations to model the movement and interactions of atoms and molecules in batteries with these electrolytes. Ultimately, the team identified the best-performing electrolyte and used it to design a lithium-sulfur battery with a different internal chemistry.

“We combined lithium metal on one side of the battery with sulfur, porous carbon, and lithium chloride on the other,” said Dr. Jijian Xu. “We designed it so that chloride ions, a charged form of chlorine, could react with sulfur during charging to form disulfur dichloride. This opens up the additional energy-storing reaction.”

Notably, the disulfur dichloride produced while lithium-sulfur batteries are charging does not mix readily with the team’s electrolyte. This helps keep the compound within the porous carbon cathode, preventing it from shuttling across the battery and reacting with lithium metal, which would waste stored energy.

“We tested how much charge the batteries could store, how they performed at different charging and discharging speeds, and how they performed over repeated use,” said Prof. Wang. “We also tested a flat, pouch-shaped prototype. X-ray and laser-based measurements tracked the chemical changes. Computer calculations and simulations, including density functional theory (DFT) and molecular dynamics (MD), helped us understand the reactions and how the electrolyte worked at the atomic level.”

Towards lithium-sulfur batteries with higher voltages and capacities

In initial evaluations, the team’s proposed electrolyte and battery chemistry achieved very promising results. Their design successfully produced a redox reaction involving three instead of two electrons per sulfur atom, boosting the resulting battery’s capacity and average operating voltage.

“Our main achievement is getting more energy from sulfur by using this additional reaction, not simply by putting more sulfur into the battery,” said Prof. Wang. “In our tests, the charge delivered per gram of sulfur increased by about 58%, and the average voltage rose from 2.05 to 2.54 volts, compared with the conventional lithium-sulfur battery used in our study.”

The researchers also estimated that increasing the sulfur cathode’s areal capacity (i.e., the amount of charge stored per unit area) could improve the battery’s stack-level specific energy. They used their approach to create an initial pouch-shaped prototype and assessed its performance in the lab.

“The pouch-shaped prototype retained 78% of its original charge-storage capacity after 100 charge-and-discharge cycles,” said Prof. Wang. “The sulfur-carbon materials delivered more than 1,700 watt-hours per kilogram.”

In the future, the team’s proposed battery chemistry and the electrolyte they introduced could contribute to the realization of more reliable lithium-sulfur batteries with high capacities and higher voltages. Meanwhile, Prof. Wang and his colleagues plan to continue refining their electrolyte’s underlying chemistry to achieve even better results.

“These findings demonstrate that high-valent sulfur redox is viable, opening a new route toward ultra-high-energy-density batteries,” added Prof. Wang. “We now want to improve the electrolyte so the battery lasts longer and works better at faster charging and discharging speeds. We also want to reduce the unwanted reactions that gradually consume this liquid. The next challenge is to turn the chemistry we have demonstrated into a more practical rechargeable battery.”

Who’s behind this story?

Ingrid Fadelli

Freelance journalist with BSc Psychology and MA International Journalism. Covers AI, robotics, neuroscience, and astrophysics since 2018.

Full profile →

Robert Egan

Bachelor’s in mathematical biology, Master’s in creative writing. Well-traveled with unique perspectives on science and language.

Full profile →

© 2026 Science X Network

Autonomous vehicles and robots need to understand not only what is around them, but how objects and people are moving. A cyclist crossing the road, a car slowing down, a pedestrian stepping off the curb, and a parked vehicle all create different motion cues that a machine must interpret quickly and accurately.

A research team led by assistant professor Zhao Na from the Singapore University of Technology and Design (SUTD) has developed IterFlow, a lightweight learning framework that helps 4D radar estimate the 3D motion of points in a traffic scene. Their study, posted to the arXiv preprint server, addresses a key challenge in autonomous perception: how to make radar-based motion understanding more accurate without depending on expensive LiDAR-based supervision.

Refining motion estimates from sparse radar

4D radar is attracting growing interest because it is more compact, more cost-effective and more robust in adverse environmental conditions than LiDAR, which is short for light detection and ranging. However, radar point clouds are also sparse and noisy, making it difficult for AI systems to estimate scene flow—the 3D motion of points between consecutive sensor frames.

IterFlow was developed to tackle this problem through a more focused design. The research team designed a task-specific network with a concise training strategy to improve radar scene flow performance. Rather than relying on increasingly complex multitask systems, IterFlow refines motion estimates step by step and uses targeted training signals to reduce errors in sparse radar data.

A central feature of IterFlow is that it does not require LiDAR-based pseudo scene flow labels during training. Instead, it uses RGB images and odometry—information about the vehicle’s own movement—as auxiliary supervision. At test time, the system needs only radar point clouds as input.

“IterFlow shows that better radar scene flow estimation does not have to depend on increasingly complex models or costly LiDAR supervision. By using images and odometry during training, we can make radar-based motion understanding lighter, more cost-effective and more applicable to real-world autonomous systems,” Zhao said.

Keeping moving objects distinct

In everyday terms, scene flow estimation helps a system work out how each point captured by a sensor is moving from one moment to the next. In a road scene, this could mean estimating the motion of points belonging to moving cars, cyclists or pedestrians while distinguishing them from static background points such as parked vehicles or roadside structures.

The team’s method uses camera images to provide object-level guidance. Through 2D tracking and segmentation, the system identifies object instances in images and projects this information into 3D radar space. This helps reduce a common source of error in radar scene flow learning, in which a model may mismatch moving foreground points with static background points.

For example, if radar points are sparse, a moving cyclist and nearby static background points may appear close together in the data. Existing methods that rely mainly on spatial distance may wrongly encourage these points to move in similar ways. IterFlow’s instance-aware losses reduce this problem by applying motion consistency within the same object instance, rather than across points that are merely nearby.

IterFlow also uses a ball query-based grouping method that is better suited to sparse radar data. Unlike k-nearest-neighbor methods, which always return a fixed number of neighbors even if some are far away, ball query first checks whether points fall within a defined spatial radius. This helps avoid false correspondences in sparse radar regions and improves robustness.

Efficiency gains and remaining limits

Experiments on the real-world View-of-Delft dataset showed that IterFlow outperformed the previous radar-based cross-modal scene flow method CMFlow while using only three losses, about 40 times fewer parameters and about 30 times fewer giga floating-point operations, or GFLOPs, a measure of computational cost. The results suggest that radar scene flow estimation can be improved without adding costly sensors or substantially increasing model complexity.

For now, this research remains at the experimental stage. The current method uses a PointNet++ point cloud feature extraction network, which supports only a fixed input point cloud size. Future work will focus on overcoming this limitation before the approach can be more broadly tested in practical vehicle or robotic systems.

By reducing reliance on costly LiDAR supervision and improving how radar learns motion from sparse data, IterFlow points toward more efficient radar-based perception for autonomous systems. Its broader significance lies not in replacing other sensors immediately, but in showing how lower-cost sensing can be made more capable through carefully designed machine learning.

Who’s behind this story?

Lisa Lock

BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021.

Full profile →

Andrew Zinin

Master’s in physics with research experience. Long-time science news enthusiast. Plays key role in Science X’s editorial success.

Full profile →