Daily Digest

ChatHealthAI proposes a multimodal framework mapping structured, longitudinal EHR data from a pre-trained foundation model into the semantic space of a frozen LLM via a task-aware resampler. Evaluated on the EHRSHOT benchmark, it improves clinical reasoning and interpretability without sacrificing predictive accuracy.

Traj-Evolve introduces a multi-agent system utilizing an Experience Pool (ExPool) and reward-ranked fine-tuning to model 5-year patient trajectories. It outperformed 9 baselines on lung cancer prediction by leveraging retrieval-augmented few-shot context from similar historical cases.

The Joint Commission introduced a voluntary RUAIH certification for hospitals deploying AI, focusing on governance, risk/bias reduction, and performance monitoring. Concurrently, the HSCC released a cybersecurity guide addressing model drift, data poisoning, and adversarial attacks in healthcare.

Beth Israel Lahey reports clinicians using Heidi's ambient AI scribe save an average of 70 minutes daily on documentation. 74% of users reported a significant reduction in after-hours charting.

A causal analysis of 21,098 UK Biobank participants utilized Directed Acyclic Graphs to estimate Backdoor-Adjusted Average Treatment Effects for DXA-derived skeletal phenotypes. Integrating the top 11 causal phenotypes yielded an AUC of 0.842 for hip fracture risk, vastly outperforming standard FRAX.

InsideTracker is transitioning to a B2B platform named Terra, featuring an orchestration layer that aggregates continuous data from Apple Health, Oura, Fitbit, Garmin, and blood labs to generate white-labeled recommendations for functional health providers.

StenCE introduces a framework for detecting severe coronary artery stenosis purely from non-invasive ECGs. Pre-training uses cross-modal contrastive learning between ECGs and invasive X-ray angiograms to stratify patient risk.

NewLimit raised a $435M Series C to advance its epigenetic reprogramming therapies to human trials by 2027. Their lead candidate targets functional restoration and resilience in aging liver cells rather than whole-cell rejuvenation.

CA-RAG is a dynamic query routing framework that balances factual grounding, token cost, and latency by assigning variable retrieval depths (from zero to top-k=10). Benchmarks showed a 26% reduction in billed tokens and 34% lower mean latency with no degradation in answer quality.

The ArthaNethra architecture decouples text retrieval from entity traversal by combining Weaviate for semantic search with Neo4j for Cypher graph queries. It utilizes a dual-model approach with Claude Haiku for bulk extraction and Sonnet 4.5 for risk reasoning.

Community consensus indicates that while semantic chunking aligns well with legal documents, it introduces unnecessary latency and ingestion costs. Fixed-size chunking with 10-20% overlap remains superior for messy PDFs and OCR-heavy data when evaluated on Recall@K.

Slipstream accelerates Approximate Nearest Neighbor Search graph construction for streaming data by starting insertion searches from candidates identified in the previous step. It achieved up to a 30.8x throughput increase in Faiss and HNSWLib.

AWS launched NEXUS, a deterministic, permutation-invariant foundation model purpose-built for tabular data. Pre-trained on over 10B rows, it handles cross-schema reasoning, missing data imputation, and mixed feature types out-of-the-box.

Microsoft unveiled its MAI series, featuring MAI-Thinking-1, a 1-trillion parameter MoE model (35B active) trained natively for reasoning. MAI-Code-1-Flash is a highly optimized 5B active parameter agentic coding model built for GitHub Copilot.

NVIDIA released the Cosmos 3 family, leveraging a two-tower Mixture-of-Transformers architecture to unify physical reasoning and action generation. It uses 3D multimodal rotary position embeddings to align video, audio, and actions on a single temporal axis.

MiniMax introduced a 'KV outer gather Q' attention mechanism that scales natively to 1M tokens without quadratic complexity. It delivers a 9x speedup in prefilling and 15x in decoding compared to standard Flash-Sparse-Attention.

ALAR introduces a dual-mode framework that utilizes compact latent reasoning for routine agent steps, escalating to explicit Chain-of-Thought only for complex deliberation. Benchmarks show generated token reductions of up to 84.6% in tool-use tasks.

Weaviate launched Engram, a managed context service that extracts, deduplicates, and reconciles raw events into structured memories for agents. It prevents memory drift and provides Temporal-grade durability for atomic state commits.

DataRobot details a digital twin architecture utilizing FastAPI and the Model Context Protocol. It heavily integrates Microsoft Presidio with configurable probability thresholds to enforce PII detection before context enters vector DBs or prompt templates.

NVIDIA outlines a framework where Hermes agents learn recurring tasks and save them as reusable markdown skills. NemoClaw provides the proxy-level network allowlists and state persistence to secure these autonomous research pipelines in production.

WarpGroup-Backend optimizes standard LLM batching by replacing zero-padding with First-Fit Decreasing bin-packing of variable-length sequences. Leveraging flash_attn_varlen_func, it yields a 2.08x throughput increase on H100s.

AURA-Mem maintains a fixed 4,224-byte inference state, completely bypassing the linear growth of standard KV-caches. A learned gate writes to memory exclusively when an observation alters the upcoming action, using 7x fewer writes.

AWS Deep Learning AMIs now natively support Seekable OCI (SOCI) lazy loading. Benchmarks demonstrate a 32.7GB vLLM container's startup time dropping from nearly 7 minutes to 21.1 seconds.

Perplexity is launching an orchestrator built with Intel that dynamically routes tasks between local inference and the cloud. Highly sensitive health and financial data defaults to on-device execution.

Amazon researchers introduced DeepFact-Eval, an iterative 'audit-then-score' protocol where an AI challenges human benchmarks with structured evidence. Moving away from static one-shot labeling increased verification accuracy from 60.8% to 90.9%.

Evaluating Large Reasoning Models via prefix-level trajectories reveals that models often suffer 'logical drift' after deriving the correct answer. Early stopping at the first correct prefix improved accuracy by up to 21%.

DharmaAI utilized DPO to mitigate text degeneration loops in structured OCR tasks by treating the model's own degenerate outputs as rejected pairs. This self-correction signal reduced completion-level failures by 59.4%.

NeurIPS desk-rejected 18.4% of Position Paper Track submissions for excessive AI generation using the Pangram v3.3.2 detector. Papers scoring 90% or higher on AI probability increased 10x over the previous year.

Uber instituted a strict $1,500 monthly AI budget per employee after exhausting its annual AI allocation in just four months. The cap specifically targets heavy token consumption from autonomous coding agents like Cursor and Claude Code.

Anthropic has filed for an IPO following a $50B funding round that pushed its valuation to $965B. Analysts predict this public market entry will force stricter API rate limits and tiered enterprise pricing to protect gross margins.

Walmart is transitioning to per-employee token quotas for its internal Code Puppy assistant to control runaway inference costs. E.ON has similarly scrapped isolated AI labs to embed predictive maintenance directly into SAP S/4HANA core systems.

A new Executive Order establishes a voluntary 30-day cybersecurity review framework for frontier models with CAISI. Major labs including OpenAI, Anthropic, and Microsoft have agreed to participate, avoiding mandatory pre-release approvals.