Xiaorui Wang, Ph.D.

🔎 Dense vs Sparse Retrieval in RAG Dense and sparse retrieval are two fundamental approaches for finding relevant documents. Sparse retrieval uses keyword matching (like traditional search), while dense retrieval uses semantic embeddings to understand meaning and context. 🔍 Simple analogy: Sparse retrieval is like using Ctrl+F 🔤 to find exact words in a document. Dense retrieval is like asking a knowledgeable friend 🧠 who understands what you mean, even if you phrase it differently. 💡 How they differ: Sparse Retrieval (BM25, TF-IDF): Matches exact keywords or stems Fast and computationally cheap Works well for precise terminology, codes, or specific names Struggles with synonyms, paraphrasing, or conceptual queries Example: Searching "diabetes treatment" only finds documents with those exact words Dense Retrieval (Embeddings): Captures semantic meaning in vector space Handles synonyms, context, and intent naturally Better for conceptual or conversational queries Requires encoding overhead and more storage Example: "Managing blood sugar" retrieves diabetes-related content even without exact keyword matches 🚀 Key takeaway: Modern RAG systems often use hybrid retrieval—combining sparse methods for precision on specific terms with dense retrieval for semantic understanding, delivering the best of both worlds. #RAG #DenseRetrieval #SparseRetrieval #SemanticSearch #InformationRetrieval #TechExplained

5

Beyond ReconVLA: Annotation-Free Visual Grounding via Language-Attention Masked Reconstruction Last week I was reading ReconVLA (Song et al., 2025 arxiv.org/abs/2508.10333) and genuinely enjoyed the work. The idea is clever: instead of telling the model where to look via external detection modules, they train a diffusion transformer head to reconstruct the "gaze region" of the manipulation target. The reconstruction pressure forces the backbone to encode spatially precise representations. Clean concept. Strong benchmark results on LIBERO and CALVIN. But then I hit a wall. Before any training can begin, you need to annotate gaze regions across every trajectory in your dataset. That is eye-tracking data, or heuristic bounding boxes drawn around target objects, across 100k+ trajectories and 2 million samples. That is not a small ask. It is expensive, time-consuming, and hard to scale to new environments. So I started asking a different question: What if we kept the reconstruction concept but removed the annotation requirement entirely? The insight I kept coming back to: the backbone already processes the language instruction. Inside those transformer layers, cross-attention scores between instruction tokens and image patches exist right now, every forward pass. The word "bowl" already produces high attention weights on bowl-shaped patches. That is a gaze signal. It is just being thrown away. So I designed LA-ReconVLA. Instead of annotating gaze regions externally, the architecture derives reconstruction targets from the backbone's own cross-attention maps over the instruction text. Top-k attended patches get masked. A lightweight 4-layer MAE decoder reconstructs them in a single forward pass, replacing the diffusion transformer entirely. No eye-tracking. No annotation pipeline. No iterative denoising at inference. Theoretically the argument holds across four independent lines: - MAE research shows masking semantically meaningful regions produces stronger representations than random masking - The information bottleneck forces the backbone to retain spatial geometry in its latent space - Direct MAE gradients to the encoder are cleaner than multi-step diffusion gradients - Using attention maps as masking targets creates a self-reinforcing grounding loop during training I have written a full architecture breakdown with diagrams in a blog post. Now I am planning to validate this on LIBERO-Spatial with a small sample (3 tasks, 50 demos per task) on a single Colab T4. I will share the results openly, whether they support the hypothesis or not. But before I run the experiments, I genuinely want to hear from people in this space: Does this concept hold up, or does it just sound good on paper? #MachineLearning #Robotics #DeepLearning #ComputerVision #VLA #Research #OpenResearch #ReconVLA #Transformer https://lnkd.in/eQ25JsWS

2

Good material in this video - a few additional considerations. 1/ From a GenAI strategy perspective, prioritize building a GenAI evaluation function and the related LLMOps layer, both pre-deployment (e.g., leaderboards, model and dataset versioning) and online (observability). If needed, integrate with telemetry tools (e.g., Langfuse-like). It’s also important to support both technical and non-technical users who need to test and validate GenAI pipelines, while keeping the platform modular. 2/ [19:00]This is the approach I have found to work best in practice: create versioned evaluation datasets for company-specific internal tasks. Combine traditional metrics (e.g., BLEU, METEOR) with LLM-as-a-Judge, and allow customization of the judge prompt depending on what you want to evaluate. 3/ [28:00]LLM-as-a-Judge can be used with or without ground-truth labels. With labels (pre-deployment), it enables reuse of the same evaluation datasets across multiple model iterations and helps address limitations of traditional metrics. Online, when labels are not yet available (e.g., outputs not reviewed by auditors yet), LLMs or agents can be used to detect malicious behavior, sentiment issues, prompt injection, etc. In this case, LLM-as-a-Judge can be used either as a guardrail or to sample conversations that require deeper review. 4/ LLM-as-a-Judge — how can an LLM evaluate another LLM? The training set may be the same, but model architecture is not. In production, models are often constrained by latency and throughput, while evaluation/audit can be done offline using more capable (e.g., reasoning-enabled) models. Judges also operate with slightly different prompts, acting as a second pass on the same problem and potentially leading to different outcomes. This can be a practical way to balance automation and accuracy, compared to classical threshold-based approaches for supervised ML . 5/ [52:00]Humans evaluate the LLM-as-a-Judge, which in turn evaluates the upstream generative model. This is a scalability layer and can work in practice under the assumptions described above. 6/ [1:16:00] Tool usage. A good practice is to return meaningful outputs from tools, so tool usage can also be evaluated using the same approach mentioned above. The observability layer should expose tool usage and, where possible, tool accuracy. 7/ Evaluation benchmarks are not specific to a single company or business application, but the methodology can often be reused (e.g., language-model simulation, Tau-Bench-like approaches). Which methodology to adopt depends on the complexity of the agent being evaluated (e.g., SWE-Bench-style vs. Tau-style). It usually makes sense to start simple and add complexity only if needed. https://lnkd.in/ep2XcVWf

16

The blog post associated with the Github repo describes the design of the Neural ODE clearly enough that it can be re-implemented in PyTorch. https://lnkd.in/gfQHPjgd #Matlab #PhiML #NeuralODE

37

2 Comments

A predictive approach to enhance time-series forecasting Forecasting events from time-series data is central across science and engineering—anticipating seizures from EEG signals, predicting system dynamics, or managing financial and energy markets. But deep learning models still struggle: long-term dependencies fade, and sudden distribution shifts derail predictions. Skye Gunasekaran and coauthors present Future-Guided Learning (FGL), a predictive-coding inspired framework where two models work in tandem. A “teacher” looks slightly into the future, detecting critical events, while a “student” forecasts them from the present. When their predictions disagree, the student is updated more strongly, reducing “surprise” and dynamically adapting to new data. The payoff is substantial: up to +44.8% AUC-ROC in seizure prediction and −23.4% MSE in nonlinear dynamical forecasting. By distilling uncertainty from near-future signals, FGL stabilizes learning and improves accuracy—bringing predictive coding from neuroscience into practical machine learning. This work points to a new path for resilient forecasting models that can learn adaptively in the face of non-stationary data. Paper: https://lnkd.in/deRacQRq #ArtificialIntelligence #MachineLearning #DeepLearning #TimeSeries #Forecasting #PredictiveCoding #Neuroscience #SeizurePrediction #EEG #DynamicalSystems #DataScience #AIResearch #AdaptiveLearning #NonStationaryData #FutureGuidedLearning

19

📋Tired of copy-pasting between your AI chat and MATLAB? MathWorks is releasing an official MATLAB MCP Core Server to integrate MATLAB in agentic workflows. The Model Context Protocol (MCP) enables AI agents (in Visual Studio Code®, Claude Desktop®, Gemini CLI®) to use tools, like your file system, your Git repos ... and now MATLAB! 🔗 https://spr.ly/604173hR1

1