Causal abstraction · Scaling laws · Neuro-symbolic reasoning — We investigate the mathematical underpinnings of intelligence. Our team recently derived new scaling exponents for mixture-of-experts architectures and connected active inference to free-energy principles. We found that performance (e.g., cross-entropy loss) improves smoothly as a power law of model size, dataset size, and compute. Collaborative work with KSG Research Tokyo explores compositionality in vision transformers.

See Kundu (2025) on Price Equation and information geometry for details.

Diffusion · Autoregressive · Flow matching — Our StreamDiffusion architecture achieves real-time synthesis on edge. We also have studied semantic guidance and alignment using human preference models. Our GenEval benchmark is used by 20+ labs.

Flow matching with the continuous Euclidean setting is a simple yet effective generative modeling paradigm that has found widespread adoption in diverse domains and large-scale applications. It is inspired by the efficient training of diffusion models, but offers a simpler perspective and enables easy implementation and generalization. At its core, flow matching follows a simple blueprint: regress onto conditional velocities that generate single data examples, and the result is a model that generates the full distribution.

See Lipman (2024) on Flow Matching for Generative Modeling for details.

Meta-learning · Robust optimization · Graph neural networks — We have derived a generalised Bayesian regret bound for online learning under shift. AutoML tools for predictive maintenance deployed inside KSG.

Recent advancements in Graph Neural Networks (GNNs) are characterized by a move beyond traditional message-passing to overcome fundamental limitations in expressivity and scalability. One key trend is the development of architectures that transcend the 1-Weisfeiler-Lehman (1-WL) expressivity barrier, with frameworks like SymGraph using discrete symbolic hashing for superior interpretability and efficiency , and LogiX-GIN embedding intrinsically explainable logic rules directly into the AI model. Our research focus is on capturing long-range interactions and higher-order structures, as seen in RANGE, a model-agnostic framework that mitigates over-squashing to accurately model molecular systems, and HL-HGAT, which leverages Hodge-Laplacian operators on simplicial complexes to learn from complex, multi-level relationships in data from neuroscience to logistics.

See Peng (2024) on Towards Few-shot Self-explaining Graph Neural Networks for details.

See Mengfan Xu (2020) on Regret Bounds and Reinforcement Learning Exploration of EXP-based Algorithms for details.

NEAT · Quality-diversity · Multi-objective evolution · Price equation — Our research on the fundamentals of Price Equations and propose a novel tensorial reformulation of the Price equation for multi‑criteria evolutionary optimization. By extending the classical Price equation to tensor form, we develop a unified mathematical framework that simultaneously captures selection pressures across multiple objectives. The proposed model naturally evolves Pareto-optimal solutions by balancing the covariance between decision variables and multiple fitness objectives through selection tensors. We provide rigorous mathematical proof that this approach generates Pareto-nondominated solutions superior to existing methods including NSGA-II and other similar scalarization techniques. Numerical experiments on benchmark problems demonstrate faster convergence to higher quality Pareto fronts with improved distribution characteristics. The integration with information geometry offers theoretical guarantees on solution optimality, while Kalman filtering extensions handle uncertainty in fitness evaluations. This work bridges theoretical evolutionary biology with computational optimization, opening new interdisciplinary directions.

See Kundu (2026) on Tensorial Price Equation for Multiobjective Optimization for details.

Knowledge Graphs · Multiomics · Signaling Pathways — KSG Research has designed a multilevel, multi-omics knowledge graph that integrates various biological ontological layers to build a Human digital twin for disease prediction and health monitoring. KSG-Karute’s goal is to create a dynamic, virtual representation of a person's health that can be used for early disease detection (like diabetes or cancer) and to interact with this model via a Large Language Models (LLMs). Recent work combines LLMs with multiomics ontological layers and ICD11, RX-Norm for graphical tool use.

See Karute → Kundu (2023) on Human Digital Twin Knowledge Graph for details.

See Karute Hong Kong, Kundu et.al. (2020) for details.

sim-to-real · world models · dexterous manipulation — We research tactile learning with ManiFeel, which demonstrates how tactile sensing enhances policy performance across diverse manipulation scenarios, benchmark tests policy learning under partial observability. Our recent work combines LLMs with visuomotor policies for tool use. KSG proposes that an important milestone in learning to abstract and generalize, is compositionality: the ability to compose and decompose a whole into reusable parts, like the redness of an object. How we get this ability is a key question in developmental neuroscience – and in AI research. The earliest neural networks, which have later evolved into the large language models (LLMs) revolutionizing our society, were developed to study how information is processed in our brains. Ironically, as these models became more sophisticated, the information processing pathways within also became increasingly opaque, with some models today having trillions of tunable parameters.

See Lei Zhang et.al. (2025) on "ResponsibleRobotBench": Benchmarking Responsible Robot Manipulation using Multi-modal Large Language Models for Details.

interpretability · concept-based explanations · counterfactuals · feature attribution — KSG Research has developed faithful and stable explanation methods. Our XAI audit toolkit is used internally at KSG for regulatory compliance.

We collaborate with RIKEN XAI Team, Japan.

ReAct+ · tree-of-thoughts · multi-agent collaboration · human laughter — KSG Research foundation has recently proposed a novel concept of Agent swarms with memory, reflection, and tool use. Our AgentBench tracks real-world API performance. Current focus on recursive self-improvement. Agentic AI refers to artificial intelligence systems designed to act with a degree of autonomy, pursuing complex, multi-step goals with minimal human intervention. When applied to modeling human creativity, this agency becomes a powerful tool for simulating the iterative and exploratory nature of the artistic process. Instead of simply generating a single output, an agentic creative AI can be tasked with a high-level objective, such as "create a novel piece of music." The system would then operate like a digital artist, generating initial ideas, critically evaluating them against its training, making autonomous decisions to refine its work, and even seeking out new data or feedback to overcome creative blocks. This ability to set its own sub-goals, iterate on its creations, and persist toward a final objective allows Agentic AI to move beyond simple mimicry and into a process that more closely mirrors the deliberate, self-guided journey of human creativity.

See Kundu (2024) on The FBSS Model: A Unified Formal Theory of Human Laughter and Creativity Under Pricean Selection for details.

See Karute → Kundu (2023) on Human Digital Twin Knowledge Graph for details.

quantisation (INT8/FP4) · speculative decoding · early exiting — We are studying FastServe engine that can speeds up LLMs 3×. Co-design with FPGA and neuromorphic hardware for edge AI. KSG is also implementing Mix-QViT, an explainability-driven MPQ framework that systematically allocates bit-widths to each layer based on two criteria: layer importance, assessed via Layer-wise Relevance Propagation (LRP), which identifies how much each layer contributes to the final classification, and quantization sensitivity, determined by evaluating the performance impact of quantizing each layer at various precision levels while keeping others layers at a baseline. Additionally, for post-training quantization (PTQ), we introduce a clipped channel-wise quantization method designed to reduce the effects of extreme outliers in post-LayerNorm activations by removing severe inter-channel variations. We validate our approach by applying Mix-QViT to ViT, DeiT, and Swin Transformer models across multiple datasets. Our experimental results for PTQ demonstrate that both fixed-bit and mixed-bit methods outperform existing techniques, particularly at 3-bit, 4-bit, and 6-bit precision. Furthermore, in quantization-aware training, Mix-QViT achieves superior performance with 2-bit mixed-precision.

See Ranjan et. al. (2025) on Mix-QViT: Mixed-Precision Vision Transformer Quantization for details.

We apply the Price equation from evolutionary biology to understand selection in ML (gradient descent, genetic algorithms). Δz = Cov(ω,z) + E(ωΔz). Our current work links natural gradients and multilevel selection in agent populations.

See Kundu (2024) on The FBSS Model: A Unified Formal Theory of Human Laughter and Creativity Under Pricean Selection for details.

• 2023, Karute → Kundu (2023) on Human Digital Twin Knowledge Graph
• 2024, Kundu (2024) on The FBSS Model: A Unified Formal Theory of Human Laughter and Creativity Under Pricean Selection
• 2025, Kundu (2025) on Price Equation and information geometry
• 2026, Kundu (2026) on Tensorial Price Equation for Multiobjective Optimization