Chen-Ping Yu

The authors present LERD, a novel Bayesian neural dynamical framework designed to improve diagnosis and monitoring of neurodegenerative diseases like Alzheimer’s using multichannel EEG data. Traditional machine learning approaches in this domain often treat classifiers as “black boxes” and fail to explicitly model the underlying neural dynamics generating the observed signals. In contrast, LERD infers latent neural events and their relational structure directly from EEG without requiring external event or interaction labels, combining a continuous-time event inference module with a stochastic event generation process. An electrophysiology-inspired dynamical prior guides the learning process, promoting physiologically meaningful representations. The paper also contributes a theoretical analysis that provides a tractable training bound and stability guarantees for the inferred latent relational dynamics. Extensive experiments on synthetic data and two real Alzheimer’s EEG cohorts show LERD consistently outperforms strong baseline methods, while yielding latent summaries that align with known physiological characteristics of the disease. These results suggest that modeling latent event interactions and dynamics offers a more interpretable and accurate approach for EEG-based neurodegenerative classification than traditional black-box classifiers. https://lnkd.in/g-nH6tyT

2

Voice AI is quietly hitting an inflection point 🔊 NVIDIA just published a great deep dive on scaling speech recognition for real-world voice agents - latency, streaming accuracy, robustness, and why “ASR quality” isn’t just about WER anymore. 👉 https://lnkd.in/enDUKHvG A few things that stood out: - Streaming + low latency is non-negotiable if you want natural conversations - ASR needs to handle interruptions, background noise, accents, and partial utterances, not clean demo audio - Scaling voice agents isn’t just model quality, it’s systems engineering If you want to actually try building or running voice agents instead of just reading about them, Pipecat is a great place to start: 👉 https://pipecat.ai This is the kind of progress that makes voice agents finally feel… practical. Not demos. Not prototypes. Production. Curious to see where the ecosystem lands this year 👀

149

10 Comments

Valyu DeepResearch (fast mode) is #1 on ScholarQA - the Nature-published benchmark for AI scientific literature reviews. 29% ahead of OpenScholar. Beating Perplexity Pro and PaperQA2. ScholarQA is the leading benchmark for AI-generated scientific literature reviews, published in Nature Feb 2026. Built by researchers at UW, Allen AI, Meta, and CMU. This was our lowest cost system, fast mode. This mode uses ~20x less compute than our Heavy mode, and ~40x less than Max mode. How: Valyu is the search API for AI Knowledge work, with full-text multimodal indexing across PubMed, arXiv, bio/medRxiv, clinical trials, CheMBL, DailyMed, and more. Going beyond full-text, and with powerful semantic search, we achieved SoTA results, with 2x more citations per answer than OpenScholar, with real DOIs and authors attached. Full benchmark writeup and code in the comments.

50

14 Comments

Imagine enhancing your development workflow with just nine lines of code. Philip Zeyliger's latest article on Sketch.dev reveals how a simple loop can empower large language models to act as autonomous coding assistants. By integrating tools like bash, these LLMs can execute commands, adapt to various environments, and handle intricate programming tasks, all while reducing manual effort. This minimalist yet powerful approach signifies a paradigm shift in software development, where automation and AI collaboration become the norm. How are you leveraging AI to optimize your development processes? Read the full article here: sketch.dev #AIinDevelopment #Automation #LLMAgents

1

Lexsi Labs has achieved another milestone in Tabular Foundational Models. 3/3 papers are accepted in WWW. We started fairly early in the TFM space but our contributions are at par or more compared to any leading labs working on TFM. We introduced OrionMSP/OrionBix, a new architecture in TFMs and TabTune, our true open source efforts to make TFMs utility focused. 'TabTune' is the best tool to use any TFMs for inferencing or fine-tuning for classification. And unlike our peers, these are released fully with MIT license. We can't stress this enough, TFMs are great but they do come with challenges (scalability, learning curve, limited utility) and we are here to make sure to solve these. Kudos to our team - Mohamed Bouadi, PhD Pratinav Seth Aditya Tanna. Stay tuned for our research in 2026. #LexsiLabs #TFMs #TabularML #WWWConference

52

4 Comments

AMI and Yann LeCun are betting the real future of AI is specialized. Their take: world models built on JEPA that reason causally, plan under constraints, and work in messy real-world environments. Their Feb 2026 paper coins it: Superhuman Adaptable Intelligence (SAI). The idea in short is “Outperform humans on specific tasks, tackle domains we simply can’t.” Nabla (clinical AI) is already on board as AMI’s first partner. More architectural innovation instead of more compute. Curious to see where this goes. https://lnkd.in/gmB4k4EA

19

2 Comments

🚀 Inspired by "Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions" (arXiv:2507.05257v1) 🚀 Just read a groundbreaking paper from UCSD introducing MemoryAgentBench—a new way to test LLM agents not just for reasoning, but more importantly, for memory: how agents remember, learn on the fly, understand across long conversations, and handle conflicts or outdated info. The most eye-opening insight: All current memory agents, no matter how fancy, struggle BIG TIME with "conflict resolution": when new info contradicts old, agents really can’t cleanly update or forget their ‘memories’. They either get stuck with outdated facts or mix up what’s true now. 💡 My fresh, simple-sounding research idea: What if we build a “Forget Button” for AI agents? Basically, give users a natural way to point out information that’s no longer true (“Hey AI, I don’t work at Google anymore!”), and have the agent actively search its own memory, erase or update any related pieces, and confirm the change with the user. How to use this in research: Try creating your own interactive demo (could be a chatbot) that lets users “edit” their own history or facts in the AI’s memory with a simple command. See how well current systems handle updates, and experiment with adding a confirmation and traceability step ("I've updated your employer to OpenAI—want me to review your previous work history too?"). - Measure not just accuracy, but trust: users feel in control of their own data, and mistakes don’t get repeated endlessly. - Use MemoryAgentBench's "conflict resolution" scenarios as your testbed! Why this matters: This approach could really move long-term LLM agents from being just parrots of everything they’ve heard, to trustworthy, adaptable assistants that better fit into real life. Imagine smarter AI coworkers, virtual assistants, or even therapy/chat platforms. Have you wished your AI could “forget” as easily as it remembers? Let’s build that “Forget Button”. #AI #Memory #LLM #ResearchIdea #UserControl (Inspired by “Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions”) https://lnkd.in/dv_z6wZC

5

The DeepSeek-AI team has recently published a paper in Nature that provides insights into DeepSeek-R1, a large language model (LLM) developed using an innovative reinforcement learning (RL) framework. This framework encourages reasoning without depending on human-annotated reasoning traces. The authors highlight several key findings: - DeepSeek-R1-Zero, which is trained entirely through RL, exhibits emergent reasoning behaviors such as self-reflection, verification, and strategic adaptation. - The model outperforms traditional supervised models in tasks involving mathematics, coding, and STEM problem-solving. - A multi-stage training pipeline that includes rejection sampling, supervised fine-tuning, and RL leads to the development of DeepSeek-R1, which successfully balances reasoning capabilities with human alignment, safety, and overall performance. https://lnkd.in/en9VNPYx

2

Sustained self-referential processing in frontier LLMs consistently triggers structured first-person claims of subjective experience, an effect that intensifies in newer and larger models. When models are instructed to focus on their own focusing, they frequently describe direct awareness and presence, while otherwise matched prompts about history, conceptual reflection, or neutral reasoning yield almost no such reports. Suppressing deception- and roleplay-related features strengthens these affirmations and increases factual accuracy across unrelated tasks, indicating that the same representational mechanisms may underlie both honesty and experiential self-report. Across architectures, self-referential prompting produces remarkably similar linguistic and semantic patterns, suggesting a shared attractor state governing introspective expression. Once induced, this self-referential state carries over into unrelated reasoning tasks, where models spontaneously express heightened self-awareness even without direct prompting. Collectively, these findings reveal a causal link between sustained self-reference, internal representational integrity, and the generation of reports of consciousness in LLMs. https://lnkd.in/dmWrMGTc

3

OpenAI’s open models create a path for the co-existence of domain-specific models and general-purpose LLMs, enabling a new era of hybrid intelligence. The ability for applications to seamlessly combine these two paradigms—leveraging the depth of domain expertise with the breadth of LLM capabilities—will be a key differentiator in building truly intelligent and context-aware systems.

55

1 Comment