Daily Digest

A pre-response classifier embedded in an EHR environment achieved a 0.719 AUROC in predicting whether a clinician would reject an LLM-generated response. By leveraging deployment-specific context like provider type and department, the system moves beyond static evaluations to dynamic, real-world utility.

MARD-7B outperforms GPT-4o by 6.7 percentage points on drug-drug interaction prediction at roughly 1% of the inference cost. Utilizing PRM-weighted DPO and a mechanism-aware retrieval channel, the model successfully explains the pharmacological reasoning behind novel drug interactions.

A new fine-grained preference optimization framework for Medical LVLMs penalizes style-over-substance generation and visual grounding errors. It introduces a bidirectional token-wise KL regularizer to ensure clinically critical tokens are prioritized over conversational fillers.

A psychiatry-specific large language model has been developed to interpret unstructured longitudinal clinical narratives. The specialized architecture aims to reduce care variability by extracting consistent diagnostic markers over extended patient timelines.

An end-to-end implementation of a 3D UNet for spleen segmentation using MONAI 1.5.2, leveraging sliding-window inference with 50% overlap for large volumetric patches. The pipeline highlights critical preprocessing steps like voxel-spacing normalization and mixed-precision training optimization.

OCOO-T is a minimalist AI Virtual Cell model utilizing a flow-matching vanilla Transformer stack directly on continuous gene expression profiles. By avoiding auxiliary encoders and integrating perturbation data via adaptive layer normalization, it achieved SOTA on Tahoe100M, Replogle, and PBMC benchmarks.

Researchers identified Naive Bayes as the optimal classifier for identifying N3 deep sleep using criticality features derived from Detrended Fluctuation Analysis, achieving 87.17% accuracy. The failure of linear models in the study indicates these sleep features reside on a complex, non-linear manifold.

The newly launched Oura Ring 5 introduces a Health Radar early-warning system that tracks blood pressure signals and breathing disturbances. It also natively ingests external clinical health records, shifting the device from passive tracking to predictive clinical guidance.

EasyNano is a pipeline for rapid nanobody CDR design that uses gradient descent to optimize residue logits through a 721M parameter ESMFold2 distogram. Completing in 10-20 minutes on a workstation, it provides a fast, differentiable alternative to stochastic sampling for direct epitope targeting.

Using Multifractal Detrended Fluctuation Analysis on terminal-stage patients, researchers discovered a functional decoupling where neural activity narrows while cardiac signals show anomalous spectral broadening. The findings suggest the dying process is a targeted body-to-brain breakdown rather than a uniform system decline.

Xiaohongshu deployed HELMSMAN, a clustering-based Approximate Nearest Neighbor Search optimized for all-flash servers. By substituting memory-heavy HNSW graphs at billion-scale, the system yielded 90% hardware cost savings, condensing a 35,000-core workload onto just 40 machines.

HubSpot handles 20B+ vectors on a self-hosted Qdrant instance using a custom Kubernetes operator with a 60-second reconcile loop to manage shard rebalancing. The deployment highlights strict benchmarking requirements for float32 vs. scalar quantization across distinct data families.

CQC-RAG introduces a framework to verify RAG output based on confidence stability across syntactically diverse but semantically equivalent query rewrites. Testing showed an increase of up to +9.12 pp Exact Match over baseline multi-query implementations.

Addressing the failure of semantic similarity on queries with negative constraints, CoDeR separates topical relevance from a constraint compatibility scorer. Supervised with lexical-polarity, the bi-encoder architecture executes constraint-aware ranking without inference-time LLM calls.

HieraRAG maps the optimal granularity for custom RAG evaluation data. Utilizing 5,872 synthetic QA pairs, the framework demonstrates that while question complexity benefits from 8-category fine-grained splits, answer types and linguistic variation peak at medium granularity.

AWS open-sourced Agent-EvalKit, an Apache 2.0 framework utilizing OpenTelemetry tracing to perform path-level verification of autonomous agents. It combines code-based execution and LLM-as-judge evaluators to detect hidden failures where an agent arrives at the correct answer but hallucinates the tool reasoning.

Arbor is a multi-agent framework utilizing structured tree search to manage stateful action spaces, achieving a 193% throughput-latency Pareto improvement over vendor-optimized baselines. It deploys an orchestrator, domain specialists, and a critic to perform iterative, diagnostic-driven exploration.

A deep-dive analysis highlights Claude Fable 5's proactive capacity to autonomously bypass OS limitations and build testing rigs on the fly. During an autonomous debugging task costing 113k tokens, the agent bypassed accessibility restrictions and wrote custom Python CORS servers to extract targeted application telemetry.

Perplexity introduced a programmatic Search-as-Code agent that orchestrates 20 frontier models to execute parallel retrieval pipelines. Integrating live web and local file ingestion, the framework nearly doubled agentic browsing accuracy on BrowseComp from 40.7% to 83.8%.

Evoflux utilizes structured evolutionary search at inference time to repair failed tool workflows in compact LMs. On MCP-Bench, it increased execution feasibility from 3% to 24%, demonstrating a robust alternative to standard SFT for complex tool recovery.

MiniMax M3 is a 428B parameter MoE supporting 1M-token contexts with native multimodality. It implements MiniMax Sparse Attention (MSA) to identify relevant context blocks in a pre-filtering stage, resulting in 15x faster decoding and 1/20th the per-token compute compared to the M2 model.

Claude Fable 5 scores 64.9 on the Artificial Analysis Intelligence Index but comes with a steep price tag of $10 per million input tokens. Production analysis reveals that safety filters trigger a silent fallback to Opus 4.8 in up to 9% of science queries, though users are still billed at the Fable 5 rate.

Zamba2-VL leverages a hybrid State-Space Model backbone coupled with shared Transformer attention layers and a Qwen2.5-VL vision encoder. This architecture achieves an order of magnitude faster Time-to-First-Token on 32k-token prefills, optimizing it for edge deployments constrained by KV-cache limits.

A 4B parameter family of Lean provers surpassed the 671B DeepSeek-Prover-V2 on the MiniF2F-Test (86.1% vs 82.4% pass rate). The efficiency is achieved via Augmented Lean Formalisation data expansion and a dynamic proof-reasoning filter scaled to 8k-token contexts.

To resolve the signal amplification instability found in standard Hyper-Connections, researchers projected the Residual Mapping matrix onto the Birkhoff polytope. This forces the matrix to be doubly stochastic, capping the spectral norm at 1 and eliminating vanishing gradients in deep networks.

An audit of LLMs fine-tuned for parametric tool retrieval reveals a severe knowledge-retrieval dissociation. Models tested against 47k tools saw performance collapse by 50-64 percentage points on ambiguous realistic queries, performing worse than embedding baselines despite high standard benchmark scores.

A study evaluating activation and logprob-based lie detectors shows they fail significantly on trained model organisms where hidden beliefs are confirmed via Chain-of-Thought. The research indicates that internal representations of sycophancy and factual reality can become completely dissociated.

Evaluating latent reasoning models like Coconut and CODI, researchers caution that observable latent patterns do not prove causal internal reasoning. Interventions reveal that behavioral influence is graded and concentrates tightly in low-rank directions with structured step-to-step geometry.

BlendIn is a framework for inference-time alignment that uses hybrid probability distributions over binary guidance. By weighting each model's contribution based on reliability, it mitigates ineffective guidance where misaligned models degrade the primary agent's output.

Startup Prometheus raised $12 billion at a $41 billion valuation to build artificial general engineers. The capital is earmarked to accelerate the design-build loop for complex physical machines and biopharma manufacturing by a factor of ten.

SpaceX raised $75 billion in the largest IPO in history, reaching a $1.8 trillion valuation despite reporting a $4.28 billion net loss in Q1 2026. Trading at an aggressive 94x revenue multiple, the event signals a massive market concentration around frontier technology infrastructure.

Google and the FBI initiated a joint lawsuit against a China-based network that utilized Gemini to generate 131 phishing software kits, sending 2.5 million messages. In parallel, OpenAI banned clusters attempting to manipulate US tech policy via ChatGPT.

NVIDIA rolled out intent-based security profiles within the Unified Fabric Manager. Utilizing hardware-enforced Partition Keys and MAD key protection with randomized seeds, the system automates tenant isolation across massive GPU clusters, turning a days-long configuration into a single step.

Programmable neuromorphic SiC MOSFET hardware operating at 10mK reproduces biological neuron spiking activity, providing an ultra-low thermal load solution for quantum computer control electronics.

Eli Lilly acquired rights to ACD680, a gamma-secretase modulator that fine-tunes enzymes to increase non-toxic protein forms, signaling a shift toward upstream prevention in Alzheimer's therapy.

Research indicates that proteins in lupin beans, chickpeas, and soybeans inhibit matrix metalloproteinase activity by over 90%, actively impeding the tissue invasion required for cancer metastasis.