The whole research project in eight hand-drawn pages: from alerts to root cause to resolution, with the architecture, datasets, agents and tech stack laid out visually.

RAG answers depend on what the retriever returns, and a few poisoned documents can mislead it. TriShieldRAG stacks three independent defenses: ingest, retrieval and generation. It stops non-adaptive poisoning almost completely, but an adaptive attacker who changes only document formatting bypasses the first ring and the whole stack. It is an honest negative result for layered defenses that share t…

AIOps platforms generate huge volumes of logs, making timely root cause analysis hard. Cloud LLMs raise privacy, cost and latency concerns, and aggregate metrics can hide prediction bias. TriCalRAG is a reproducible benchmark for privacy-aware, on-premise LLM log-anomaly detection with generated root-cause explanations, across four datasets, two open models and three prompting strategies, on a sin…

We introduce LLaMA 4, a suite of open language models ranging from 8B to 405B parameters, demonstrating state-of-the-art performance on multiple benchmarks with improved efficiency and reasoning capabilities.

We propose a unified architecture combining vision transformers with diffusion models for high-quality image generation, achieving superior visual fidelity and semantic control.

We demonstrate how language models can autonomously improve through iterative tool use and feedback loops, showing significant gains in reasoning and task completion.

We present a comprehensive framework for multimodal fusion across vision, language, and audio modalities, enabling richer scene understanding and cross-modal reasoning.

We propose novel optimization techniques for reducing memory footprint during LLM inference through advanced KV-cache management and dynamic quantization strategies.

We analyze emerging vulnerabilities in LLM applications to prompt injection attacks and propose defense mechanisms with formal guarantees.

We leverage graph neural networks for efficient and accurate protein 3D structure prediction from amino acid sequences, achieving performance comparable to AlphaFold.

We introduce RAG 2.0, an improved retrieval-augmented generation framework with adaptive chunking, hybrid search, and efficient re-ranking for enterprise applications.

We scale constitutional AI methods to larger models, demonstrating how feedback from principles can align language models without human annotation overhead.

We propose flow matching as an alternative to diffusion for generative modeling, achieving faster sampling and better quality in text-to-image synthesis.

We address temporal consistency in video object detection through novel temporal attention mechanisms, reducing computational cost while improving accuracy.

We present a framework for orchestrating complex workflows with multiple autonomous agents, enabling collaborative problem solving and task decomposition.

We study scaling laws for vision-language models and propose efficient training strategies for models reaching 100B parameters.

We propose sparsity-aware optimization for transformers, achieving 4x speedup with minimal accuracy loss through attention sparsification and weight pruning.

We develop certified defense mechanisms against adversarial attacks in vision models, providing formal guarantees on robustness.

We apply deep learning techniques to climate modeling, achieving significant speedups in weather prediction and climate projection.

We demonstrate how to efficiently expand context windows in language models to 1M tokens without retraining, enabling new applications.

We scale mixture of experts architectures to trillion-parameter models, achieving better training efficiency and inference performance.

We improve SETR with better positional encoding and multi-scale features, achieving state-of-the-art semantic segmentation on ADE20K.

We develop agents capable of autonomous planning and tool use, demonstrating the ability to solve complex multi-step tasks without human intervention.

We propose self-supervised learning from audio-visual data, achieving strong performance on downstream speech and vision tasks.

We reduce communication overhead in distributed training through advanced gradient compression techniques, enabling efficient large-scale model training.

We demonstrate practical approaches to training privacy-preserving language models with differential privacy guarantees.

We apply generative models for molecular design, accelerating drug discovery with AI-guided compound generation.

We improve LoRA with dynamic rank adjustment and hierarchical composition, enabling more efficient fine-tuning of large models.

We develop methods for automatically optimizing prompts through gradient-based learning, improving task-specific performance.

We advance neural radiance fields with better geometry priors and faster training, enabling high-quality 3D reconstruction from sparse images.

We explore memory mechanisms and reasoning capabilities in language models, showing improvements beyond simple scaling.

We develop efficient cross-modal retrieval through joint embedding spaces, enabling fast image-text and audio-text search.

We solve challenges in serving billion-parameter models with batching, caching, and dynamic acceleration strategies.

We analyze poisoning attacks on ML models and propose defense mechanisms with formal robustness guarantees.

We develop hybrid classical-quantum algorithms for machine learning, demonstrating speedups on specific problem classes.

We advance code generation models with better architecture design and training techniques, improving code quality and correctness.