Natural Language Processing Innovations

Reasoning Agentic RAG: The Evolution from Static Pipelines to Intelligent Decision-Making Systems The AI research community has just released a comprehensive survey that could reshape how we think about Retrieval-Augmented Generation. Moving beyond traditional static RAG pipelines, researchers from leading institutions including Beijing University of Posts and Telecommunications, University of Georgia, and SenseTime Research have mapped out the emerging landscape of Reasoning Agentic RAG. The Core Innovation: System 1 vs System 2 Thinking Drawing from cognitive science, the survey categorizes reasoning workflows into two distinct paradigms: Predefined Reasoning (System 1): Fast, structured, and efficient approaches that follow fixed modular pipelines. These include route-based methods like RAGate that selectively trigger retrieval based on model confidence scores, loop-based systems like Self-RAG that enable iterative refinement through retrieval-feedback cycles, and tree-based architectures like RAPTOR that organize information hierarchically using recursive structures. Agentic Reasoning (System 2): Slow, deliberative, and adaptive systems where the LLM autonomously orchestrates tool interaction during inference. The model actively monitors its reasoning process, identifies knowledge gaps, and determines when and how to retrieve external information. Under the Hood: Technical Mechanisms The most fascinating aspect is how these systems work internally. In prompt-based agentic approaches, frameworks like ReAct interleave reasoning steps with tool use through Thought-Action-Observation sequences, while function calling mechanisms provide structured interfaces for LLMs to invoke search APIs based on natural language instructions. Training-based methods push even further. Systems like Search-R1 use reinforcement learning where the search engine becomes part of the RL environment, with the LLM learning policies to generate sequences including both internal reasoning steps and explicit search triggers. DeepResearcher takes this to the extreme by training agents directly in real-world web environments, fostering emergent behaviors like cross-validation of information sources and strategic plan adjustment. The Technical Architecture What sets these systems apart is their dynamic control logic. Unlike traditional RAG's static retrieve-then-generate pattern, agentic systems can rewrite failed queries, choose different retrieval methods, and integrate multiple tools-vector databases, SQL systems, and custom APIs-before finalizing responses. The distinguishing quality is the system's ability to own its reasoning process rather than executing predetermined scripts. The research indicates we're moving toward truly autonomous information-seeking systems that can adapt their strategies based on the quality of retrieved information, marking a significant step toward human-like research and problem-solving capabilities.

Exploring the Advancements in Large Language Models: A Comprehensive Survey ... Large Language Models (LLMs) have emerged as pivotal tools, revolutionizing natural language processing and beyond. A new research paper titled "Survey of Different Large Language Model Architectures: Trends, Benchmarks, and Challenges" sheds light on recent developments in LLMs and their multimodal counterparts (MLLMs). Here are some key insights worth sharing: 👉 The Evolution of LLMs The journey of LLMs began with foundational architectures like BERT and GPT, culminating in today’s sophisticated models. Central to this evolution is the **Transformer architecture**, introduced in 2017, which has fundamentally changed how we approach language tasks. 👉 Understanding Model Architectures The paper categorizes LLMs into three primary architectures: - "Auto-Encoding Models" (e.g., BERT): Focused on understanding context but limited in generating text. - "Auto-Regressive Models" (e.g., GPT): Excellent for generation tasks but may lack contextual awareness. - "Encoder-Decoder Models" (e.g., T5): Combine strengths of both types, applicable for complex input-output scenarios. This classification is crucial for selecting the appropriate model for specific tasks in real-world applications. 👉 Unleashing Multimodal Capabilities A significant focus of the paper is on "Multimodal Large Language Models (MLLMs)". These models can process and integrate multiple data formats—text, images, audio, and video—expanding the horizons for applications such as image captioning and video analysis. The ability to leverage diverse data sources represents a substantial shift in how we think about AI applications. 👉 Importance of Benchmarking Effective benchmarking is vital for assessing the performance of LLMs. The paper outlines several benchmarks used to measure various capabilities, ensuring that researchers and industry professionals can evaluate model efficiency and effectiveness reliably. This aspect is crucial for advancing LLM technology and aligning it with industry needs. 👉 Navigating Challenges and Future Directions While LLMs have made remarkable strides, the paper also highlights ongoing challenges, such as data limitations, model compression, and the complexities of prompt engineering. Addressing these issues will be essential for developing more robust and effective models in the future. The insights gathered from this research underscore the necessity of understanding not just the capabilities of LLMs but also the underlying architecture and challenges to fully leverage their potential in professional contexts. For those interested in delving deeper, I encourage you to read the full paper.

Sakana AI Introduces Doc-to-LoRA and Text-to-LoRA: Hypernetworks that Instantly Internalize Long Contexts and Adapt LLMs via Zero-Shot Natural Language Doc-to-LoRA (D2L) and Text-to-LoRA (T2L) are two innovative methods that utilize lightweight hypernetworks to instantly customize Large Language Models (LLMs) through a single forward pass. T2L enables zero-shot task adaptation based solely on natural language descriptions, matching the performance of specifically tuned adapters while significantly reducing adaptation costs compared to traditional in-context learning. D2L addresses the "long context" bottleneck by internalizing documents directly into model parameters through a Perceiver-based architecture and a chunking mechanism. This allows models to answer queries without re-consuming original context, maintaining near-perfect accuracy on information retrieval tasks at lengths exceeding the model's native window by more than four times while reducing KV-cache memory usage from gigabytes to less than 50 megabytes. Both systems operate with sub-second latency, effectively amortizing training costs and opening possibilities for rapid, on-device personalization. Remarkably, D2L also demonstrates cross-modal capability, transferring visual information from Vision-Language Models into text-only LLMs zero-shot to enable image classification purely through internalized weights..... Full analysis: https://lnkd.in/g9MHtyq3 Updates: https://lnkd.in/gtCTeEFM Doc-to-LoRA Paper: https://lnkd.in/g7BnBcBq Code: https://lnkd.in/gvWGSbM5 Text-to-LoRA Paper: https://lnkd.in/gbSZk948 Code: https://lnkd.in/g6dPx737 Sakana AI Rujikorn Charakorn Edoardo Cetin Shinnosuke Uesaka Yujin Tang Robert Tjarko Lange

🔥 𝗠𝗼𝗱𝗲𝗿𝗻𝗕𝗘𝗥𝗧: 𝗔 𝗙𝗮𝘀𝘁𝗲𝗿, 𝗦𝗺𝗮𝗿𝘁𝗲𝗿, 𝗮𝗻𝗱 𝗠𝗼𝗿𝗲 𝗦𝗰𝗮𝗹𝗮𝗯𝗹𝗲 𝗘𝗻𝗰𝗼𝗱𝗲𝗿 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗟𝗟𝗠 𝗘𝗿𝗮 Despite all the buzz around generative LLMs like GPT and Llama in recent years, it’s easy to overlook the true everyday heroes of NLP: encoder-only transformers like BERT. Since its introduction in 2018, BERT has been the go-to architecture for practical NLP tasks such as classification, entity extraction, and retrieval for its efficiency and low inference costs. In fact, just 𝗥𝗼𝗕𝗘𝗥𝗧𝗮 alone, a popular BERT variant proposed by Facebook AI in 2019, has more downloads on HuggingFace than the top 10 LLMs combined! That said, BERT architecture hasn’t seen many significant upgrades, aside from a few notable exceptions like RoBERTa mentioned above, 𝗗𝗲𝗕𝗘𝗥𝗧𝗮 (Microsoft, 2021), and more recently 𝗠𝗼𝘀𝗮𝗶𝗰𝗕𝗘𝗥𝗧 (Databricks, 2023), which further optimized pretraining efficiency. Meanwhile, the scientific community has poured immense effort into advancing generative transformers, introducing architectural innovations that pushed the boundaries of scalability, efficiency, and long-context processing. 𝗠𝗼𝗱𝗲𝗿𝗻𝗕𝗘𝗥𝗧 takes this progress full circle by absorbing these advancements and redefining what encoder-only models can achieve. As a result, 𝗠𝗼𝗱𝗲𝗿𝗻𝗕𝗘𝗥𝗧 outperforms mainstream models on standard academic benchmarks across information retrieval, natural language understanding, and code retrieval. Even against 𝗗𝗲𝗕𝗘𝗥𝗧𝗮𝗩𝟯, the go-to model for natural language understanding competitions on Kaggle, ModernBERT not only beats it on GLUE but also uses 𝟱𝘅 𝗹𝗲𝘀𝘀 𝗺𝗲𝗺𝗼𝗿𝘆 and is 𝘂𝗽 𝘁𝗼 𝟰𝘅 𝗳𝗮𝘀𝘁𝗲𝗿! 𝗠𝗼𝗱𝗲𝗿𝗻𝗕𝗘𝗥𝗧 𝗦𝗶𝘇𝗲𝘀: – 𝗠𝗼𝗱𝗲𝗿𝗻𝗕𝗘𝗥𝗧-𝗯𝗮𝘀𝗲: 22 layers, 149 million parameters. – 𝗠𝗼𝗱𝗲𝗿𝗻𝗕𝗘𝗥𝗧-𝗹𝗮𝗿𝗴𝗲: 28 layers, 395 million parameters. 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗮𝗹 𝗜𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁𝘀: – 𝗥𝗼𝘁𝗮𝗿𝘆 𝗣𝗼𝘀𝗶𝘁𝗶𝗼𝗻𝗮𝗹 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀 (𝗥𝗼𝗣𝗘) handle sequences up to 8K tokens. – 𝗕𝗣𝗘 𝗧𝗼𝗸𝗲𝗻𝗶𝘇𝗲𝗿 optimizes for diverse text, including code. – 𝗟𝗼𝗰𝗮𝗹-𝗚𝗹𝗼𝗯𝗮𝗹 𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 balances performance and efficiency. – 𝗚𝗲𝗚𝗟𝗨 𝗔𝗰𝘁𝗶𝘃𝗮𝘁𝗶𝗼𝗻 improves task performance over GeLU. – 𝗙𝘂𝗹𝗹 𝗨𝗻𝗽𝗮𝗱𝗱𝗶𝗻𝗴 reduces memory and computation costs. – 𝗙𝗹𝗮𝘀𝗵 𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 (𝟮 & 𝟯) boosts long-context inference speed by 2–3x. 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗗𝗲𝘁𝗮𝗶𝗹𝘀: – 𝗠𝗮𝘀𝘀𝗶𝘃𝗲 𝗣𝗿𝗲𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 on 2 trillion tokens (600x more than BERT), includes code. – 𝗪𝗮𝗿𝗺𝘂𝗽-𝗦𝘁𝗮𝗯𝗹𝗲-𝗗𝗲𝗰𝗮𝘆 (𝗪𝗦𝗗) ensures stable training and checkpoint reuse. – 𝗦𝘁𝗮𝗯𝗹𝗲𝗔𝗱𝗮𝗺𝗪 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗲𝗿 improves training stability with gradient clipping. – 𝗦𝗲𝗾𝘂𝗲𝗻𝗰𝗲 𝗣𝗮𝗰𝗸𝗶𝗻𝗴 efficiently handles variable-length batches. Find the paper and blog post details in the comments. #AI #MachineLearning #Transformers #NLP #DeepLearning #ArtificialIntelligence

Two new papers from Sakana AI caught my attention this week. Doc-to-LoRA and Text-to-LoRA offer two complementary ways to update models, together enabling continual learning systems. Here is the workflow of both methods: ➡️ Doc-to-LoRA (D2L): Turning documents into memory Doc-to-LoRA focuses on knowledge updates and internalizes documents as LoRA adapters. 1. A document (a policy, report, textbook, or even visual data) is provided as input. It acts as the knowledge source. 2. A text encoder or a vision-language model (VLM) converts the document into hidden activations representing its information. These activations capture the document’s semantic content. 3. A hypernetwork generates a LoRA adapter, outputting LoRA weight updates for the base LLM. The activations are fed into a hypernetwork. 4. The generated LoRA adapter encodes the document’s facts directly in the weights and is attached to the base model. No gradient training or optimization is needed at deployment time. 5. Once the adapter is active: • The model can answer questions about the document • The original document no longer needs to appear in the prompt ✅ Why Doc-to-LoRA is powerful • Adds long-term memory to LLMs • Reduces latency, VRAM usage, dependency on long context windows • Eliminates repeated document re-reading • Works beyond context limits • Can even internalize visual information into text models, for example, giving 75% accuracy on Imagenette in experiments --- ➡️ Text-to-LoRA (T2L): Turning instructions into skills Text-to-LoRA, in turn, focuses on model adaptation, instantly teaching the model new behaviors. Instead of running a fine-tuning pipeline, we simply describe the task in natural language and generate the adapter. ▪️ How Text-to-LoRA works step by step: 1. A task description, which specifies the desired behavior, is written. For example: • "Solve math problems step by step." • "Write legal summaries in formal language." • "Extract entities from medical reports." 2. The text description is processed into hidden representations capturing the task’s intent. 3. The encoded task description is passed to the hypernetwork, and the hypernetwork outputs LoRA weights that modify the base model’s behavior. Again, this happens in a single forward pass. 4. The generated LoRA adapter is attached to the base model. Now the model behaves according to the described task. 5. No dataset or training pipeline required ✅ Why Text-to-LoRA is powerful • Replaces expensive fine-tuning pipelines • Enables instant task specialization • Allows rapid experimentation with new behaviors • Makes adaptation as easy as writing instructions Overall, a model can learn new information with Doc-to-LoRA, learn new abilities with Text-to-LoRA, and use both simultaneously for continual learning. What I like is that it’s not one magic method – progress comes from symbiosis.

Summary: Artificial Intelligence (AI) has come a long way, especially in Natural Language Processing (NLP). At the heart of this progress are Large Language Models (LLMs), which can now generate human-like text, translate languages, and answer questions with ease. But LLMs are just the beginning. The next big leap? Large Reasoning Models (LRMs). These new models take things to the next level by handling complex reasoning tasks that were once thought to be uniquely human. One of the coolest innovations driving LRMs is the concept of "thought"—basically, a series of steps that mimic how humans think through problems. This approach allows LRMs to tackle tasks like tree search or reflective thinking. And with the help of Reinforcement Learning (RL), these models can now generate high-quality reasoning paths automatically. How It Works: Building LRMs involves a few key steps: Creating Data: Traditionally, training LLMs required humans to annotate data, which is slow and expensive. Now, researchers are using LLMs themselves to generate data through automated search, paired with external verification to ensure accuracy. Learning to Think: Early on, techniques like Supervised Fine-Tuning (SFT) were used to train these models. But RL and Direct Preference Optimization (DPO) have proven to be better—they help the model learn how to reason in ways that feel closer to how humans think. Test-Time Tricks: It’s not just about training; test-time techniques matter too. Methods like Chain-of-Thought (CoT) and Tree-of-Thoughts (ToT) guide the models through step-by-step reasoning during testing, making their answers more accurate and understandable. What’s Happening Now: OpenAI’s "ol" series models are a great example of what LRMs can do. They’ve nailed some of the toughest tasks in areas like math, coding, and scientific problem-solving. These models are great at breaking down big problems into smaller parts, connecting knowledge, and reasoning across a variety of fields. Why It Matters: LRMs could change the game for AI. Imagine what’s possible: helping students learn more effectively, pushing the boundaries of scientific discovery, or even simplifying software development. By handling complex tasks, LRMs could free up humans to focus on creativity, strategy, and other higher-level thinking. The future of AI is looking pretty exciting, don’t you think?

Pragmatic, unglamorous innovations are often the most useful. For example, consider NLP to label patient messages rather than Gen AI to answer them. In late 2022, Kaiser started applying its home-grown natural language processing (NLP) algorithms to label patient portal messages with categories such as admin question, medication issue, skin condition, and emergency. Over a five-month study period, the NLP labeled more than 3.6 million messages. Roughly 40% (1.5 million) of these messages were flagged and directed to a centralized “desktop medicine” team, which resolved them before they ever reached the patients’ personal PCP/nurse’s inbox. Pairing a (now) relatively unglamorous type of AI with a pragmatic team-based workflow meaningfully improved this vexing aspect of care. Compare this to the more headline-grabbing efforts to use GenAI to draft responses to patient messages, which has been disappointing so far. At Stanford, clinicians only used 20% of GPT-generated drafts. These drafts did not save physicians time nor reduce turnaround time [doi:10.1001/jamanetworkopen.2024.3201]. At UC San Diego, clinicians who used ChatGPT drafts paradoxically spent 22% more time reading messages/drafts and did not respond any faster [doi:10.1001/jamanetworkopen.2024.6565]. Though I believe GenAI drafts will be useful one day, physicians and nurses overloaded with patient messages need help now. (We must also recognize that editing GenAI drafts is sometimes harder than writing a response from scratch). All this to say, it’s often best to pick the lower-hanging fruit first. Also, tech alone is rarely the solution. What really made a difference at Kaiser was pairing NLP labels with a practical workflow and appropriately resourced centralized (“Desktop Medicine”) team that took work off their physician colleagues’ plates. #healthcareai #patientmessaging #healthcareonlinkedin https://lnkd.in/g5ycmxGb

The transformer architecture was initially celebrated as a breakthrough in NLP, but ultimately enabled breakthroughs across multiple modalities. Researchers are now addressing two of its main limitations—tokenization and quadratic scaling—opening up new multi-modal applications. At its core, the self-attention mechanism central to transformers is a simple and elegant way to extract patterns from input embeddings. The source modality of these tokens (text, images, sound) and their arrival order are irrelevant *. Self-attention enables effective comparison between all tokens in a set. This differs from architectures like CNNs or RNNs, which are tailored to specific modalities. While this makes them more data efficient (with stronger inductive biases), the remarkable scalability of transformers often compensates (see comments): we can increase dataset size until the advantage of more biased models diminishes. However, creating input embeddings remains highly modality-dependent. Text input relies on tokenization, which introduces issues like language bias and challenges in reading numbers. Additionally, the quadratic scaling of self-attention limits embedding granularity. Creating 10x more embeddings from the same input requires 100x more compute. In the last months, there was increasing focus on removing tokenization bottlenecks and reducing self-attention's quadratic cost by examining input data at different scales. This includes local attention mechanisms that combine local embeddings with global attention, and neural network-based approaches that generate embeddings dynamically (see comments). I’m really excited to soon see this enable byte-level, multi-modal models with unprecedented performance, speed and cost-effectiveness. As a bonus, 2025 might go down as the year we finally moved beyond tokenization and its quirks. #deeplearning #llms #genai * It might of course help (or even be needed) to add constraints (e.g. causal attention) or additional biases/information (e.g. positional encodings) depending on the modality to optimize. But the general idea of self-attention is really powerful irrespective of the modality, and enables us to mix modalities. * * Essentially implementing a form of fully-connected graph neural network layer.

This paper surveys the advancements and applications of pre-trained language models such as BERT, BioBERT, and ChatGPT in medical natural language processing (NLP) tasks, emphasizing their role in enhancing the efficiency and accuracy of medical data analysis. 1️⃣ Pre-trained language models have revolutionized various medical NLP tasks by leveraging large-scale text corpora for initial pre-training, followed by fine-tuning for specific applications. 2️⃣ The paper categorizes and discusses several medical NLP tasks, including text summarization, question-answering, machine translation, sentiment analysis, named entity recognition, information extraction, medical education, relation extraction, and text mining. 3️⃣ For each task, the survey outlines basic concepts, main methodologies, the benefits of using pre-trained language models, application steps, relevant datasets, and evaluation metrics. 4️⃣ The paper summarizes recent significant research findings, comparing their motivations, strengths, weaknesses, and the quality and impact of the research based on citation counts and the reputation of publishing venues. 5️⃣ It identifies future research directions, such as enhancing model reliability, explainability, and fairness, to foster broader clinical applications of pre-trained language models. ✍🏻 Luo X., Deng Z., Yang B., Luo M.Y. Pre-trained language models in medicine: A survey. Artificial Intelligence in Medicine. 2024. DOI: 10.1016/j.artmed.2024.102904

Sharing our latest research FINCH: Financial Intelligence using Natural language for Contextualized SQL Handling. In this research we introduce a novel dataset and evaluation metric designed to enhance the application of natural language processing (NLP) to financial Text-to-SQL tasks. The research addresses critical gaps in existing benchmarks and evaluation methods, providing a more robust framework for developing AI/ML solutions in finance. It is coauthored with Dr. Avinash Kumar Singh | Stefano Pasquali FINCH serves as a crucial resource for researchers and practitioners in Gen AI / Agentic AI, enabling more accurate evaluation and development of Text-to-SQL models tailored for the financial domain. Paper -> https://lnkd.in/dmSEeRTP Data -> https://lnkd.in/d3NdrZsd