Daily Digest

Benchmarking Claude Sonnet 4.5 on MMLongBench-Doc revealed that traditional OCR pipelines outclassed Native PDF vision parsing in both accuracy (59.6% vs 52.0%) and cost ($0.1885 vs $0.2552 per query). Vision parsing unexpectedly struggled with chart-heavy pages and exhibited a 7% intrinsic transport failure rate, challenging the narrative that vision LLMs render enterprise OCR obsolete.

Tencent open-sourced a SQLite-backed local memory pipeline designed to combat context fragmentation by offloading verbose tool logs to external files mapped by node IDs. Utilizing reciprocal rank fusion (RRF) for hybrid retrieval and Mermaid syntax for short-term task representation, the system reduced token usage by 61.38% on WideSearch.

To eliminate alignment bias that artificially favored sentence-based chunkers, this automated evaluation tool introduced a two-mode sampler. The update mixes paragraph-based passages with variable-length sentence windows to guarantee neutrality when comparing chunking strategies across diverse document structures.

NVIDIA tackles the scalar gate problem in delta-rule models by introducing decoupled channel-wise erase and write gates within a linear attention layer. Implemented via fused Triton kernels for chunkwise training, the 1.3B parameter model drastically improved S-NIAH-3 long-context retrieval at 2K from 63.2 to 89.8 compared to Kimi Delta Attention.

ByteDance developed MMProLong, a Qwen2.5-VL-based model, proving that training long-context models via text transcription actually degrades performance compared to task-specific QA pairs. By generating targeted questions with Seed 2.0, the model learns to locate passages rather than memorize them, stabilizing context retention up to 512,000 tokens.

Researchers utilized Claude Code to autonomously design control algorithms for test-time scaling, producing an algorithm that cuts token usage by 70% against standard self-consistency. AutoTTS monitors confidence shifts across generation rounds to dynamically prune or open solution paths, generalizing successfully to DeepSeek-R1-Distill-Llama-8B.

Microsoft Research validated that code-based browser control is superior to visual action prediction by releasing Webwright, where agents write and debug Playwright scripts in a terminal environment. The framework scored 60.1% on Odysseys (up from GPT-5.4's 33.5%) by employing a self-reflection gate and compacting context history every 20 steps.

To make unconstrained Observe-Think-Act loops financially viable in production, an architectural hybrid approach utilizes early-commitment classifiers and a LOOP Skill Engine for deterministic replay. Compiling successful execution traces into branch-free recipes bypasses the LLM on subsequent runs, reducing token burn by 93.3% to 99.9%.

The latest FastAPI updates refactored Server Sent Event (SSE) field validation to prevent downstream applications from transmitting broken data, highly relevant for async LLM token streaming. Additionally, caching and fixture scope adjustments sped up the internal test suite by roughly 24%.

A new sidecar library addresses the lack of built-in TTL management in LangGraph checkpointers for production systems. Supporting Postgres, SQLite, and InMemory stores, it utilizes idle TTL or absolute age policies with per-thread overrides to prevent unconstrained database bloat.

The llama-server now features an experimental --tools flag enabling raw filesystem and shell executions directly from the model. While it eliminates the need for heavy MCP agent wrappers, the lack of security sandboxing currently demands strict containerization for production deployment.

A proof-of-concept for Go loads libcuda.so at runtime using purego, bypassing the build-time requirement for the full CUDA toolkit. It resolves CUDA's per-thread context affinity via runtime.LockOSThread in a dedicated executor, paving the way for significantly lighter Docker images for Go-based ML tools.

Laboratory studies confirm that a highly precise, low infectious dose of just 10 viral particles is sufficient to infect a cow's udder with H5N1 influenza. Despite high viral loads in milk, the research found no evidence of transmission through inhalation or shared milking devices in lab settings.

An analysis utilizing mast cell-deficient mice demonstrated that mast cells play a critical role in supporting early innate antiviral responses to Respiratory Syncytial Virus (RSV). They are essential for maintaining lung eosinophil homeostasis, effectively limiting overall lung damage during infection.

An evaluation by mathematician Adam Kucharski revealed that fast routing models in Microsoft Copilot's Auto mode hallucinate differences between identical datasets based strictly on geographical labels. While Fast/Flash models relied on built-in cultural stereotypes, reasoning-tier models successfully executed Python scripts to verify the data was identical.