Daily Digest

A large-scale evaluation of open-weight models (7B-72B) across ten biomedical QA datasets found that retrieval provides only a marginal 1-2 point improvement over no-retrieval baselines. The study concluded that the primary bottleneck in medical QA is the underlying model's inability to effectively utilize retrieved expert evidence, not the retrieval quality itself.

Built on Qwen3.5-VL, Argus introduces query-conditioned late-interaction retrieval using a region-aware Mixture-of-Experts (MoE) architecture. It achieves SOTA on ViDoRe V1 with highly efficient 1024-dimensional retrieval heads, allowing the retriever to dynamically focus on specific charts or tables based on the query.

Researchers propose replacing standard Recall@k with "1/Ratio@k" (Inverse Approximation Ratio) for evaluating Approximate Nearest Neighbor search in RAG applications. They demonstrate that downstream utility remains stable even when Recall@k drops, provided the approximation ratio is high, enabling cheaper and faster vector search configurations.

Testing against MMLU-Pro, this study reveals a massive 56-84 percentage point shift in how models adopt retrieved context based solely on wrapper labels. Using "Instruction:" or "Reference:" maximizes context adoption, while "Example:" labels consistently suppress it.

Amazon AGI introduced the DeepFact-Eval protocol to address the failure of static benchmarks in highly complex reasoning tasks. They found that unassisted PhD-level experts were only 60.8% accurate at verifying deep research claims, but accuracy jumped to 90.9% when humans audited rationales actively generated by an AI "challenger."

CrossDNA is a new parameter-efficient language model explicitly designed for cross-strand DNA modeling. By capturing the biophysical realities of the double helix rather than relying on single-strand paradigms, it significantly improves predictive performance for non-coding variants.

An XGBoost pipeline optimized with SMOTE and Optuna achieved 0.943 accuracy in classifying Normal, MCI, and Alzheimer's Disease using only eight routine clinical features. SHAP analysis validated the clinical relevance of the features, proving that high-accuracy early detection is possible without expensive imaging or CSF tests.

This GNN framework predicts isoform-aware druggability, outperforming purely geometric methods by capturing affinity-based resistance via learned embeddings. It enables the computational triage of splice variants that cause kinase inhibitor resistance.

Researchers demonstrated that pretraining Brain Foundation Models (BFMs) destroys higher-order structure like third-order co-skewness in fMRI data. A simple linear pipeline projecting fMRI into a co-skewness-preserving subspace outperformed 650M-parameter BFMs in predicting cognitive performance, requiring zero pretraining or GPUs.

NEXUS is a foundation model purpose-built for structured tabular data, pre-trained on over 10 billion rows to handle numbers, categories, and dates deterministically without LLM-style tokenization. The permutation-invariant architecture natively handles missing values and eliminates traditional feature engineering for billion-row datasets.

NVIDIA's 550B LatentMoE model (55B active parameters) is optimized for long-context agentic reasoning and supports million-token context lengths. Using the new NVFP4 format, it claims 5x faster inference and a 30% cost reduction for autonomous workflows compared to equivalent dense models.

Gemma 4 12B features an encoder-free architecture that processes text, image, audio, and video natively without separate vision or audio encoders. Designed to run on 16GB consumer hardware, it projects patches and audio frames directly into the LLM's hidden dimension, simplifying multimodal fine-tuning.

The PostgreSQL 19 Beta introduces auto-scaling asynchronous I/O and SQL/PGQ for property graph queries using standard SQL. Additional improvements include 2x faster inserts with foreign key checks, query stability advice extensions, and a new REPACK command for non-blocking table rebuilds.

LiftQuant introduces continuous bit-width control by projecting a 1-bit lattice from a higher-dimensional space, overcoming rigid integer-based quantization limits. The framework successfully compressed a 70B parameter LLM to 2.4 bits, allowing it to fit into a 24GB GPU while outperforming standard 2-bit models.

A systematic investigation into Attention projections found that sharing Key and Value matrices (Q-K=V) achieves a 50% KV cache reduction with minimal perplexity degradation. Combining this with Multi-Query Attention yields up to a 96.9% reduction in cache size, offering massive memory optimization for long-context inference.

AWS now supports Seekable OCI (SOCI) on Deep Learning Containers, enabling lazy loading of massive image files. Benchmarks show a 32.7GB container image pulling and starting in just 21 seconds, down from nearly 7 minutes, vastly improving GPU auto-scaling responsiveness.

Researchers proposed an ontology-grounded simulation framework to verify LLM agents against rigorous regulatory requirements in healthcare and finance. By generating scenarios mapped directly to legal ontologies rather than broad personas, the system achieved significantly higher regulatory coverage and generates machine-verifiable trust certificates.

OpenJarvis is an on-device agent framework that leverages cloud models as "teachers" at search-time to correct failure clusters and update local tool specifications. By running a refined spec locally, it achieves accuracy within 3.2 percentage points of Claude Opus while reducing API costs by 800x.

DataRobot showcased an architecture where agents use a secure Deno-based JavaScript VM to write their own tools dynamically from raw OpenAPI specs. This eliminates the need for static tool registries and API wrapper maintenance, allowing the agent to persist successful functions into a stable toolkit.