New Algorithms 2026: Breakthroughs in AI, Crypto, Graph Learning & Optimization

Multi-agent systems based on large language models often rely on sequential, multi-round dialogue, which incurs high latency and token costs. TopoDIM (Topology generation with Diverse Interaction Modes) is a framework that enables agents to autonomously construct heterogeneous communication topologies in a single shot, without iterative coordination [citation:1].

Designed for decentralized execution to enhance adaptability and privacy, TopoDIM achieves significant efficiency gains:

46.41% token reduction 1.50% performance improvement

The framework demonstrates strong adaptability in organizing communication among heterogeneous agents. Code is available at: https://github.com/Sundiasy/TopoDIM [citation:1].

Signed graphs — networks with both positive and negative edges — appear in recommendation systems, trust-distrust networks, and computational biology. GraphC is a hierarchical clustering algorithm that operates without requiring a predefined number of communities, making it particularly valuable when the "true" number of groups is unknown or changes over time [citation:2].

Built on a balanced-spectral duality result connecting Harary cuts to Laplacian eigenvectors, GraphC scales to real-world graphs with nearly ten million vertices and over twenty million edges within practical time and memory budgets [citation:2].

The algorithm employs a loss function that symmetrically accounts for both positive and negative edges, mitigating bias from highly skewed sign distributions. Applications include isolating malicious actors in Bitcoin OTC trust networks and discovering functional protein modules [citation:2].

Kolmogorov-Arnold Networks (KAN) offer symbolic regression and interpretability but suffer from slow convergence and training difficulties. DATCNN (Dense Affine Transformation Convolutional Neural Network) is a novel architecture that inherits KAN's core strengths while offering significantly enhanced flexibility [citation:3].

Each DATCNN layer employs a dual-path architecture — a backbone of stacked convolutional layers with batch normalization, plus a residual connection. By applying affine transformations to activation functions, DATCNN achieves the effect of learnable activation functions in KAN while leveraging established CNN theories and design principles [citation:3].

Extensive experiments demonstrate that DATCNN outperforms KAN and its variants in both accuracy and training efficiency across function fitting, image classification, and natural language processing tasks [citation:3].

Octopus Camouflage Optimization is a metaheuristic inspired by four adaptive behaviors of Octopus vulgaris: distributed intelligence, chromatophore-based camouflage, ink-cloud escape, and context-sensitive behavioral switching [citation:4].

The algorithm introduces the Octopus Camouflage Factor (OCF), a diversity-conditioned, dimension-adaptive energy signal that governs the search process. This creates a closed-loop feedback mechanism that modulates exploration intensity based on real-time population dynamics [citation:4].

Unlike hybrid methods, OCO maintains full algorithmic purity, deriving all search operators from the unified OCF architecture. Extensive validation on 32 benchmark functions against 20 state-of-the-art algorithms demonstrated a Friedman rank of 2.98 and statistically significant superiority in 82.8% of Wilcoxon pairwise comparisons [citation:4].

Lattice-based cryptography underpins post-quantum security, but state-of-the-art schemes require power-of-two cyclotomic fields that limit attainable security levels. Radical Ring-LWR (RR-LWR) develops the Learning With Rounding counterpart to recently proposed radical ring LWE [citation:5].

The work provides a proved closed-form bound on error sampling distortion, placing RR-LWR inside the regime identified as safe for instantiating Ring-LWE/LWR. Two schemes are instantiated: Mithril (an IND-CCA KEM) and Octarine (an EF-CMA signature scheme) [citation:5].

Optimized AVX2 and Arm Cortex-M4 implementations show both are competitive with — and in some cases significantly faster than — the MLKEM and MLDSA standards [citation:5].

Blockchain consensus and distributed systems require efficient multi-signature schemes. Lemur+ is a compact post-quantum synchronized multi-signature scheme based on Module LWE and Module SIS [citation:17].

Building on Lemur (CCS 2026), it achieves an almost constant signature size of approximately 56 KB even for one million signers and a 42-year key lifetime, compared to 491 KB in Lemur — a 8.8× improvement [citation:17].

The scheme supports non-interactive aggregation and targets large-scale, long-term distributed applications [citation:17].

Federated Foundation Models (FedFMs) aim to leverage data from small client models to enhance a large server-side foundation model, but existing methods require expensive local training or incur high communication costs with privacy risks. FedGRPO reformulates the problem as a reinforcement learning-style evaluation process [citation:7].

The framework performs competence-based expert selection using a lightweight confidence graph, then leverages the "Group Relative" concept from Group Relative Policy Optimization by packaging each question with its solution rationale into candidate policies [citation:7].

By exchanging reward values instead of data or model updates, FedGRPO reduces privacy risk and communication overhead while enabling parallel evaluation across heterogeneous devices. Empirical results demonstrate superior downstream accuracy and communication efficiency compared to conventional FedFMs baselines [citation:7].

Graph transformers have attracted considerable attention for graph representation learning, but existing methods often neglect hyper-order structures arising from implicit node-groups. DOGT (Double-Order Graph Transformers) dynamically incorporates hyper-order features through Adaptive Node-Group (ANG) learning [citation:8].

Instead of using predefined heuristics or clustering algorithms, DOGT uses a learnable discriminative mask matrix to infer hyper-order edges, then applies Double-Order Attention (DoA) to compute weights along the transformation from raw-order to hyper-order graphs [citation:8].

Extensive experiments on real-world datasets show DOGT achieves rapid convergence and outperforms state-of-the-art methods on graph-level classification and regression tasks [citation:8].

Converting pretrained dense models into Mixture-of-Experts (MoE) architectures typically requires expensive training from scratch. Cluster-aware Upcycling, developed by Seoul National University and LG AI Research, leverages the semantic structure already present in pretrained models to promote expert specialization from the outset [citation:9].

The method collects activations at each layer, groups tokens with similar features into clusters, computes whitening matrices for each cluster, and applies truncated SVD to give each expert a distinct initialization tailored to its assigned cluster [citation:9].

Applied to CLIP, the method consistently outperformed conventional Sparse Upcycling: VTAB-Natural accuracy increased from 62.0% to 63.3%, ObjectNet from 42.3% to 43.5%, and ImageNet-R from 71.2% to 72.1% [citation:9].

Vision Transformers achieve high performance in video analysis but require substantial computation and memory, making real-time deployment on edge devices difficult. Ouroboros, developed by Seoul National University, reduces redundant computations by incorporating motion information of objects and backgrounds within videos [citation:18].

The system leverages motion vectors from hardware video encoders to identify identical visual information across consecutive frames. A toroidal input space treats left-right and top-bottom boundaries as seamlessly connected, enabling accurate recognition of objects crossing frame boundaries [citation:18].

On NVIDIA Jetson Orin edge devices, Ouroboros achieved up to 87.0% reduction in computational load, 2.61× faster inference, and 64.5% reduction in energy consumption, while maintaining accuracy loss below 1% in object detection and instance segmentation tasks [citation:18].