hierarchical

FreeFlow is a hierarchical transformer for optical flow estimation that eliminates task-specific inductive biases and achieves state-of-the-art accuracy on benchmarks like Sintel and KITTI-2015.

HiMA-MDD introduces a hierarchical multi-agent system for interpreting multimodal clinical interviews to detect depression, achieving state-of-the-art performance on the E-DAIC dataset.

This paper introduces HC-RAG, a hierarchical cross-modal retrieval-augmented generation framework for evidence-centric financial question answering over 10-K filings, along with a new benchmark Multi-Doc-2025. It outperforms RAPTOR and GraphRAG on financial QA benchmarks, especially for long-document and table-related queries.

Introduces FreSH, a frequency-segmented hierarchical multi-expert framework for multivariate time series classification, achieving state-of-the-art accuracy on UEA benchmarks with reduced model size and computational cost.

The paper introduces DocTrace, a hierarchical framework for long document visual question answering that casts the task as explicit evidence graph reasoning. It achieves state-of-the-art results on three benchmarks while enabling traceable evidence provenance, outperforming Qwen3-VL-8B-Instruct by 11-14 points.

Setoka is a benchmark for evaluating memory-augmented personalized agents' ability to understand users hierarchically (semantic memory, episodic memory, behavior patterns, personality traits) from heterogeneous data, revealing that current memory systems struggle with tasks requiring cross-source integration and abstraction.

Introduces Hierarchical Experimentalist Agents (HExA), an in-context, experiment-centric self-improvement framework that enables LLM agents to design experiments, learn reusable skills, and answer queries in novel domains, achieving significant improvements over baselines on the Interphyre physics simulation benchmark.

HiComm is a plug-in communication module for cooperative multi-agent reinforcement learning that grounds messages in the sender's hierarchical observation structure, using a receiver-driven query and three-stage decoding to reduce communication volume by up to 23x.

FoggyTrust is a hierarchical extension of FLTrust that localizes trust computation to fog nodes, improving robustness against Byzantine attacks in heterogeneous federated learning settings, achieving over 50% improvement on challenging attacks like Krum and Trim on CIFAR-10.

This paper introduces Simplified Sparse Attention (SSA), a method that uses gist tokens during continued pretraining to enable efficient chunk selection at inference without architectural changes, achieving high compression ratios and outperforming baselines on long-context tasks like LongBench and retrieval-augmented generation.

This paper introduces SCOPE-FL, a hierarchical federated learning framework that uses the Top Trading Cycle algorithm to ensure strategy-proofness and Pareto efficiency in client selection, with reward distribution via Shapley value approximation and blockchain-based execution.

This paper proposes HyGRAG, a hierarchical graph RAG framework that integrates contextual and relational information for multi-hop reasoning, achieving a 9.7% average accuracy improvement over existing methods.

Introduces ProHiFlo, a hierarchical flow matching framework for de novo protein generation with coarse-to-fine generation, functional guidance, and SE(3)-equivariant architecture, achieving state-of-the-art performance with 4x fewer sampling steps.

Introduces ArabiGEE, the first comprehensive Arabic grammatical error explanation taxonomy with a hierarchical structure spanning orthographic, morphological, syntactic, and lexical dimensions, comprising 27 error types, 140 correction types, and 324 explanations.

Proposes a Multi-Granularity Reasoning Network (MGRN) that explicitly leverages hierarchical semantic features for natural language inference, outperforming strong baselines on multiple benchmarks.

We propose hierarchical RBF-KAN and RBF-SKAN architectures for multidimensional function approximation and random field learning. The frameworks offer universal approximation properties and partially alleviate the curse of dimensionality, with empirical results showing improved accuracy over existing methods.

CHAM-net introduces a contrastive hierarchical adaptive meta-network that captures site-specific and cross-year dynamics for robust global methane flux prediction, outperforming baseline methods on simulation and observational datasets.

This paper presents CosmicFish-HRM, a compact 82.77M parameter language model with a hierarchical reasoning module that dynamically allocates reasoning compute during inference, learning when to halt based on input complexity.

Introduces BOHM, a zero-cost hierarchical attribution method for compound AI systems that extracts attribution from routing weights, outperforming Shapley-based methods in many real-world deployments.

Maestro is a reinforcement learning-driven framework that dynamically composes ensembles of frozen expert models and skills for multimodal tasks, achieving 70.1% average accuracy with a 4B orchestrator, surpassing GPT-5 and Gemini-2.5-Pro.