IIIT-H’s Research That’s Headed To NeurIPS 2026

NeurIPS features the latest research in deep learning, machine learning, computer vision, natural language processing, reinforcement learning, computational neuroscience, and statistics. Here’s a roundup of the research papers from IIIT-H that are slated for the prestigious event to be held in early December.

The 2026 edition of the multi-disciplinary NeurIPS (Conference on Neural Information Processing Systems) is scheduled to be held across 3 different locations – Sydney, Atlanta and Paris. As per the organizers, though spread across three continents, the lineups share common topics. “Each site has a slight accent of its own: science in Sydney, where three talks put machine learning to work in biology, astronomy and industry; mathematics in Paris, with the theory of learning and the future of theorem proving; and society in Atlanta, where two talks look at AI’s footprint on education, communities and the energy grid.” IIIT-H’s research is a bit of all.

Part Concepts
AI image generators like Stable Diffusion or FLUX turn text into pictures. They’re good at whole objects. For instance, ask for “a dog” and you get a dog. But they’re much worse at parts of objects. If you ask for “a dog with its left front leg raised” or “a person with a “left lower arm,” the model often ignores the detail or mixes it up.

To solve this problem, IIIT-H researchers Vaibhav Agrawala and Varghese Kuruvilla under the guidance of Prof. Ravi kiran Sarvadevabhatla looked at what happens inside such a model. When it reads “dog,” it focuses on the dog in the image. When it reads “left front leg,” its attention is scattered all over the place. While the models understand the words individually (“left,” “front,” “leg”) they can’t combine them into one idea tied to a specific spot on the animal. The researchers explored if could teach the model to treat a part phrase like “left ear” as a single concept, without breaking what it already does well.

What they did
Instead of three separate words, like left, front and leg, they merged the phrase into a single token, called a PartConcept. Then they retrained the text-understanding part of the model by tweaking selected layers, using a technique called LoRA. “It is a cost-efficient way to adjust a big model,” explains Prof. Ravi Kiran. With this, they aimed for two simultaneous goals, that of localization where the model’s attention for that phrase should light up on the correct part of the image, and preserving generation where the model ought to still make good images like before. 

What they found
The team found that their trained model not only achieves state-of-the-art results at part segmentation: finding and outlining specific parts from a text description but also that it’s better at following part-level instructions when generating images. So if the description is for purple stains on left lower arm, it lands on the exact spot of the correct arm without any special mechanism built in. 

The research which has been jointly undertaken with Adobe Research has been detailed in the paper titled, “Part Concepts: A Unified Mechanism For Fine-Grained Part Localization and Generation”. According to Harsh Rangwani, Adobe Research, “Part-level control is still an open problem for image generation models like Nano Banana2, Flux2.0 & co. Our paper tackles it head-on.” Prof. Ravi Kiran adds that the research is significant because it demonstrates finer control over AI image generation and will lead to better tools for finding parts of objects in images. 

One Photo, One Sentence, One Seamless Animation
In order to animate characters, create immersive VR and even for household robotics, one of the key requirements is to create realistic human motion inside 3D environments. But existing methods often require detailed 3D scans of the surroundings, and paired data making the process expensive and time-consuming. In their research paper titled, “ENACT: Single-Image Human-Scene Interaction Motion from Language via Foundation-Model Orchestration”, Prof. Charu Sharma and researchers Sreehari Rajan, Kunal Bhosikar, and Nikos Athanasiou have offered a simpler approach. “Since casual images of indoor environments exist aplenty, we leveraged those along with free-form text instruction to create animated motion,” says Prof. Charu. 

What They Did
What is impressive is that the researchers didn’t build a new model from scratch. Their argument is that every skill needed to create an animation already exists in some off-the-shelf AI tool, but what’s missing is a way to get them working together. The ENACT system decomposes the task into three stages – the first step where it reasons by reading the photo and the instruction which it then uses to prepare a structured plan describing what a person should do in the scene, then the plan is grounded in 3D geometry, and finally there is motion synthesis. “The biggest direct impact is the way it can scale embodied AI-training data,” remarks Prof. Charu. 

There are some limitations of the ENACT model though. For instance, right now, it only works on static scenes. It can’t handle carrying objects, opening drawers, or more than one person, and it inherits the mistakes of every tool it relies on. If one model misjudges a cluttered room or a shiny surface, the whole chain suffers. That said, if the approach holds up, it could help generate huge amounts of training data for household robots, and make it easier for animators and game designers to place characters in real spaces using plain language. Of course, making realistic 3D humans cheap to insert into real images could be misused for misleading or non-consensual content. For that, the team plans to use watermarking and a gated release. 

Safety and Fairness In Multi-Agent Behaviour
Researchers Ankita Kushwaha, Kiran Ravish and Preeti Lamba under the supervision of Prof. Pawann Kumar addressed a problem frequently encountered in multi-agent reinforcement learning – that of unfairness. They developed a novel learning method that has been detailed in the paper titled, “Safe-Fair MACPO: Burden-Fair Constrained Policy Optimization for Safe Multi-Agent Reinforcement Learning¹”.  

“Multi-agent reinforcement learning is a form of artificial intelligence in which several agents learn to work together. These agents could be robots, autonomous vehicles, delivery systems, or software agents sharing limited resources. Most existing methods train them to achieve two main goals: complete the task successfully and avoid unsafe behaviour. However, safety and performance alone do not guarantee fairness,” says Ankita. To explain further, she provides an example of a team of robots working together in a warehouse. The team may complete its task successfully and avoid accidents, but one robot might repeatedly carry the heaviest loads, take the longest routes, or work in the riskiest areas. Although the task was completed successfully from the team’s point of view, from the individual robot’s point of view, the work was not shared fairly.

What They Did
Oftentimes, a high overall score is awarded to a system which may be repeatedly placing more effort, waiting time, risk, or inconvenience on the same agents. Over time, this can create a serious imbalance, even if the team appears successful as a whole. The Safe-Fair MACPO learning method developed by the researchers considers performance, safety and fairness together. “Our method measures the burden experienced by each agent. Depending on the application, this burden could represent energy use, difficult movements, exposure to danger, waiting time, longer travel routes, or limited access to a shared resource. One of the main ideas in our method is called fairness debt. When an agent carries more than its fair  share of the burden, the system records this imbalance. It then considers that history while making future decisions. This way, it avoids repeatedly assigning difficult or risky work to the same agent,” says Ankita. 

The method also follows a safety-first approach. If a fairer decision could make the system unsafe, safety is given priority. Among the available safe choices, the method then tries to select actions that distribute burdens more fairly without causing a major loss in overall performance.

What They Found
Safe-Fair MACPO was evaluated on a range of simulated tasks involving robot movement, navigation, coordination, and resource sharing. The results show that their method generally distributes burdens more evenly while maintaining strong task performance and low or zero observed hard-safety violations. The implications are far-reaching especially as multi-agent AI systems become more common in transportation, robotics, logistics, and resource management. It will be increasingly important to ensure that they are not only effective and safe, but also fair to every participant.

Addressing Challenges of Federated Learning
Federated learning, also known as collaborative learning, relies on the collaboration of multiple entities to train a machine learning model without having to share their local data. But the entities or clients as they are more popularly known, can often act as a bottleneck in the process when they disconnect, or drop offline or stop participating. “The most typical example of an entity is a hospital or an institution and each of them wants the model to perform well on its own data. When the model performs poorly, the organizations drop out and when enough of them do so, the shared model deteriorates even further,” explains Prof. Sai Kiran Bulusu. To address this problem, the professor and his collaborators M Yashwanth, Arunabh Singh, Ashok Nayak, along with Prof. Anirban Chakraborty (IISc, Bangalore) proposed a framework called FedUCA in the paper titled,”Federated Learning by Utility-Constrained Stochastic Aggregation for Improving Rational Participation¹”.  

What They Did
Instead of assigning each organisation a fixed influence, the researchers’ framework FedUCA considers how much benefit each participant expects to receive and adjusts the aggregate weights accordingly. It also introduces a degree of randomness into this process, allowing the system to explore different combinations of contributions. “The idea is that a carefully chosen combination can help meet individual organisations’ performance expectations while still improving the overall model’s effectiveness without relying on monetary rewards to encourage participation,” says Prof. Sai Kiran. 

Based on the experiments conducted by the researchers, the results suggest that accounting for organisations’ individual interests can make collaborative AI training more stable and effective. However the tests were conducted in experimental settings and there is an assumption that participants report their preferences honestly.

Three Values. One GPU.
Large language models (LLMs) can perform a wide range of language-related tasks, but they require significant computing power and memory. This makes them difficult to run on smaller devices with limited resources. To address this challenge, Prof. Priyesh Shukla and researcher Nisharg Nargund developed TernaryLM, a 132-million-parameter language model designed to use less memory without compromising its language capabilities. Unlike conventional methods that reduce the size of an already-trained model, TernaryLM is trained from scratch using only three possible values for its parameters: -1, 0 and +1. This allows the model to use much less memory. 

The research that has been documented in the paper titled, “TernaryLM: Memory-Efficient Language Modeling via Native 1.5-Bit Quantization with Adaptive Layer-wise Scaling” will be presented at the GlobalSouthAI workshop at NeurIPS. Tests revealed that TernaryLM could perform well on language tasks. It achieved an F1 score of 82.47% on the MRPC benchmark, outperforming DistilBERT despite being trained on 55 times less data. It also used 2.4 times less memory than an equivalent model using standard 32-bit precision, without increasing processing time. Another interesting finding was that restricting the model to three possible parameter values helped prevent overfitting, allowing it to perform more consistently on new data. This offers useful insights into how memory and computational resources can be allocated more efficiently when designing smaller language models.