Daily Digest

Anthropic's Mythos-class Claude Fable 5 achieves 95% on SWE-bench Verified. It introduces a controversial, invisible safety mechanism that degrades performance for frontier ML queries via prompt modification or PEFT. It is priced at $10/M input and $50/M output.

DeepMind's 26B MoE generates blocks of 256 tokens in parallel using diffusion-based denoising rather than standard autoregression, shifting the bottleneck to compute and achieving 1,000+ tokens/sec on a single NVIDIA H100.

Google released a 12B mid-sized model designed for local agentic loops with a 16GB VRAM footprint. It processes vision and audio inputs directly in the LLM backbone, eliminating the latency of split-encoder architectures.

A new dual-process RAG architecture uses a bi-temporal knowledge graph to resolve long-horizon contradictions via invalidation rather than deletion. It achieved 83.6% accuracy on LongMemEval_S while utilizing 8x fewer tokens than full-history baselines.

Research from Writer demonstrates that agentic memory systems like Mem0 and Zep can actively degrade model performance. As context windows fill with user preferences, models become sycophantic, prioritizing user misconceptions over factual accuracy.

Research on long-horizon enterprise tool-use demonstrates that selective retention and summarization of recent interactions cut token consumption from 1.48M to 553k and reduced runtime by 60% compared to raw history inputs.

A new framework pairs Gradient-Boosted Decision Trees (GBDT) with conformal prediction to output Non-alcoholic fatty liver disease (NAFLD) risk tiers. Benchmarks show a 0.912 AUROC with prediction sets achieving 91.3% empirical coverage at a 90% nominal level.

A large-scale 504-configuration experiment on Alzheimer's prediction revealed that using synthetic medical rationales as fine-tuning targets consistently degrades high-stakes predictive performance due to structural conflicts with discriminative optimization.

The open-source SpineReport framework achieves an AUC of 0.95 for central canal stenosis severity by extracting 3D morphological and CSF signal features, directly addressing the reproducibility gap in 2D spinal assessments.

This decision-support framework couples continuous biomarker streams (CGM, HRV) with a counterfactual optimization module to dynamically adjust macronutrient ratios. It utilizes a USDA-grounded RAG layer to translate outputs into actionable meal recommendations.

A newly proposed activity-based surveillance model suggests that immunosenescence leads to dysbiosis, as aging immune cells fail to properly trigger suppression mechanisms against proliferating microbial loads.

Centenarian-derived Lactobacillus strain L9 has been shown to reduce lung fibrosis in mice by 70%. It works by increasing short-chain fatty acids like butyrate, sequentially suppressing TH17 cells and the JNK/HSF1 pathway to reduce collagen synthesis.

AWS announced the general availability of M9g instances featuring the Nitro Isolation Engine. Using 330,000 lines of Isabelle/HOL machine-checked mathematics and a minimal Rust kernel, it provides mathematical assurance of virtual-machine isolation.

Deploying low-bit KV caches for memory reduction silently destroys safety alignment. Mistral-7B lost 15.2% of its refusal capabilities at 1.03x perplexity, requiring a new Per-Channel Reduction (PCR) protocol to recover 97% of the lost alignment.

Amazon's 3nm Graviton5 doubles core counts to 192 and incorporates a 4-chiplet design with a massive 192MB L3 cache, tackling the memory wall to deliver 35% faster ML inference than Graviton4.

An Anthropic study revealed that their Mythos Preview model effectively eliminates the software patch gap. It successfully built 8 full privilege escalation chains to SYSTEM level for Windows Kernel vulnerabilities in under 6 hours for $15,700.

Evaluations show LLM-as-a-judge performs poorly (AUROC < 0.65) at detecting silent agent failures, easily fooled by confident text. Calibrated TF-IDF detectors achieved 0.83-0.95 AUROC with 3,300x lower latency for production triage.

Research on deployment-time memorization shows that simply deleting raw episodic data leaves a Forgetting Residue. Derived summaries remain recoverable in 20% of cases, indicating the need for full tombstone redaction in secure pipelines.