Pratik Prabhanjan Brahma

Pretam Ray, Pratik Prabhanjan Brahma, Zicheng Liu et al.

Evolutionary agentic systems intensify the trade-off between computational efficiency and reasoning capability by repeatedly invoking large language models (LLMs) during inference. This setting raises a central question: how can an agent dynamically select an LLM that is sufficiently capable for the current generation step while remaining computationally efficient? While model cascades offer a practical mechanism for balancing this trade-off, existing routing strategies typically rely on static heuristics or external controllers and do not explicitly account for model uncertainty. We introduce AdaptEvolve: Adaptive LLM Selection for Multi-LLM Evolutionary Refinement within an evolutionary sequential refinement framework that leverages intrinsic generation confidence to estimate real-time solvability. Empirical results show that confidence-driven selection yields a favourable Pareto frontier, reducing total inference cost by an average of 37.9% across benchmarks while retaining 97.5% of the upper-bound accuracy of static large-model baselines. Our code is available at https://github.com/raypretam/adaptive_llm_selection.

Aditya Kumar Singh, Hitesh Kandala, Pratik Prabhanjan Brahma et al.

Vision-language models (VLMs) have achieved remarkable multimodal understanding and reasoning capabilities, yet remain computationally expensive due to dense visual tokenization. Existing efficiency approaches either merge redundant visual tokens or drop them progressively in language backbone, often trading accuracy for speed. In this work, we propose DUET-VLM, a versatile plug-and-play dual compression framework that consists of (a) vision-only redundancy aware compression of vision encoder's output into information-preserving tokens, followed by (b) layer-wise, salient text-guided dropping of visual tokens within the language backbone to progressively prune less informative tokens. This coordinated token management enables aggressive compression while retaining critical semantics. On LLaVA-1.5-7B, our approach maintains over 99% of baseline accuracy with 67% fewer tokens, and still retains >97% even at 89% reduction. With this dual-stage compression during training, it achieves 99.7% accuracy at 67% and 97.6% at 89%, surpassing prior SoTA visual token reduction methods across multiple benchmarks. When integrated into Video-LLaVA-7B, it even surpasses the baseline -- achieving >100% accuracy with a substantial 53.1% token reduction and retaining 97.6% accuracy under an extreme 93.4% setting. These results highlight end-to-end training with DUET-VLM, enabling robust adaptation to reduced visual (image/video) input without sacrificing accuracy, producing compact yet semantically rich representations within the same computational budget. Our code is available at https://github.com/AMD-AGI/DUET-VLM.

Prateek Agrawal, Pratik Prabhanjan Brahma

Detecting and predicting the behavior of pedestrians is extremely crucial for self-driving vehicles to plan and interact with them safely. Although there have been several research works in this area, it is important to have fast and memory efficient models such that it can operate in embedded hardware in these autonomous machines. In this work, we propose a novel architecture using spatial-temporal multi-tasking to do camera based pedestrian detection and intention prediction. Our approach significantly reduces the latency by being able to detect and predict all pedestrians' intention in a single shot manner while also being able to attain better accuracy by sharing features with relevant object level information and interactions.

Luca Anthony Thiede, Pratik Prabhanjan Brahma

Trajectory or behavior prediction of traffic agents is an important component of autonomous driving and robot planning in general. It can be framed as a probabilistic future sequence generation problem and recent literature has studied the applicability of generative models in this context. The variety or Minimum over N (MoN) loss, which tries to minimize the error between the ground truth and the closest of N output predictions, has been used in these recent learning models to improve the diversity of predictions. In this work, we present a proof to show that the MoN loss does not lead to the ground truth probability density function, but approximately to its square root instead. We validate this finding with extensive experiments on both simulated toy as well as real world datasets. We also propose multiple solutions to compensate for the dilation to show improvement of log likelihood of the ground truth samples in the corrected probability density function.

Pratik Prabhanjan Brahma, Qiuyuan Huang, Dapeng Wu

While deep learning has pushed the boundaries in various machine learning tasks, the current models are still far away from replicating many functions that a normal human brain can do. Explicit memorization based deep architecture have been recently proposed with the objective to understand and predict better. In this work, we design a system that involves a primary learner and an adjacent representational memory bank which is organized using a comparative learner. This spatially forked deep architecture with a structured memory can simultaneously predict and reason about the nature of an input, which may even belong to a category never seen in the training data, by relating it with the memorized past representations at the higher layers. Characterizing images of unseen object classes in both synthetic and real world datasets is used as an example to showcase the operational success of the proposed framework.