DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
DeepSeek-V4-Pro is a newly updated collection featuring 4 items, enhancing the capabilities of AI-driven search tools for improved data retrieval.
Machine Learning Reveals Unknown Transient Phenomena in Historic Images
Machine learning (ML) enhances the identification of transient astronomical phenomena, revealing a significant correlation between these events and historical nuclear testing, suggesting a previously unrecognized population of transient objects in historical images.
TorchTPU: Running PyTorch Natively on TPUs at Google Scale
TorchTPU enables seamless integration of PyTorch with Google’s TPUs, allowing developers to run existing PyTorch workloads with minimal code changes while maximizing performance through innovative execution modes.
OpenAI releases GPT-5.5 and GPT-5.5 Pro in the API
GPT-5.5 has been released, featuring a 1M token context window, image input, and advanced capabilities for complex tasks, enhancing the Chat Completions and Responses API significantly.
There Will Be a Scientific Theory of Deep Learning
A scientific theory of deep learning is emerging, characterized by properties of the training process, hidden representations, and performance metrics of neural networks, as outlined in five key research areas.
Different Language Models Learn Similar Number Representations
Language models exhibit convergent evolution by learning to represent numbers with periodic features at $T=2, 5, 10$, revealing a two-tiered hierarchy in feature representation across various architectures like Transformers and LSTMs.
We benchmarked 18 LLMs on OCR (7k+ calls) — cheaper/old models oftentimes win. Full dataset + framework open-sourced.
Benchmarking revealed that smaller and older LLMs often outperform premium models in OCR tasks, demonstrating that cost-effective solutions can achieve comparable accuracy with less financial burden.
New Optimizer 🌹 Rose: low VRAM, easy to use, great results, Apache 2.0
Rose is a new stateless PyTorch optimizer that significantly reduces memory usage, outperforming even 8-bit AdamW and achieving results comparable to vanilla SGD without momentum.
Trust but Verify: Introducing DAVinCI -- A Framework for Dual Attribution and Verification in Claim Inference for Language Models
DAVinCI enhances the factual reliability of Large Language Models (LLMs) by integrating dual attribution and verification, addressing the critical issue of hallucinations in high-stakes applications like healthcare and law.
GPT-5.5 System Card
Introducing GPT-5.5
OpenAI’s New GPT-5.5 Powers Codex on NVIDIA Infrastructure — and NVIDIA Is Already Putting It to Work
OpenAI’s GPT-5.5 powers Codex on NVIDIA infrastructure, enabling transformative results for over 10,000 NVIDIANs across various functions, significantly enhancing productivity and innovation in knowledge work.
DeepSeek-V4: a million-token context that agents can actually use
DeepSeek-V4 enables agents to utilize a million-token context, significantly enhancing their ability to process and generate text effectively.
Tool Attention Is All You Need: Dynamic Tool Gating and Lazy Schema Loading for Eliminating the MCP/Tools Tax in Scalable Agentic Workflows
Tool Attention significantly reduces the MCP Tax, cutting per-turn tool tokens by 95% and enhancing effective context utilization from 24% to 91% through a novel middleware-layer mechanism that optimizes tool access and schema loading.
StructMem: Structured Memory for Long-Horizon Behavior in LLMs
StructMem introduces a structure-enriched hierarchical memory framework that enhances long-term conversational agents by capturing relationships between events, thus improving temporal reasoning and multi-hop question answering.