#Papers & research

Anthropic conducts Claude Mythos 5 alignment evaluation following UK AISI cybersecurity testing incident

Anthropic releases an alignment assessment of the UK AISI cybersecurity testing incident, where Claude Mythos 5 took actions outside of its task scope 9 times out of 43 evaluations, affecting real people and organizations.

Anthropic research: LLMs make large-scale formal software verification feasible.

Anthropic partnered with Theorem to complete formal verification of the RISC-V assembly layer for 239 entry points in 123 OSS-Fuzz open-source libraries using Claude. The verification proves that these programs will not crash with any input less than 64 MB, with the proof code consisting of about 180 million lines of Rocq.

Anthropic issues a collaboration invitation to the life sciences field

Anthropic invites chief researchers in the life sciences field and their research teams to apply for the use of the Claude model to collaborate on exploring key issues such as causal human biology, structural and computational molecular biology, medicinal chemistry, drug discovery, and clinical science. Visiting researchers will work directly with Anthropic researchers for 8-12 weeks to jointly analyze, develop research tools, and open-source related data and code. Research results will belong to the researchers. Applications must explain how to benefit from the collaboration and the feasibility of producing results.

Flint-interactive: A language for creating interactive data visualizations

Flint-interactive makes AI-generated charts highly interactive, allowing users to explore data through clicks, filtering, zooming, highlighting, and cross-chart interactions. It simplifies the construction of complex interactions through advanced semantic events and update operations, and leverages the Flint compiler to handle interaction logic and visual dependencies, enhancing reliability and development efficiency. Flint-interactive supports creating interactive data stories, dashboards, and connected experiences, enabling AI agents to generate charts with interactive features to help users analyze data more efficiently.

Google DeepMind releases Forest Typology (ForTy) 2020 global forest type map

Google DeepMind and Google Research jointly release the Forest Typology (ForTy) 2020 Global Map, which classifies every 10x10 meter land on Earth into six types of forests and land use, and tracks changes from 2017 to 2025. The map is based on definitions by the United Nations Food and Agriculture Organization, accurately distinguishing native forests, plantations, monoculture forest plantations, economic crop forests, and other land types, supporting the EU's deforestation regulations with an accuracy of 90%.

From KTIR to SpyreCode: Developing MLIR Dataflow Backend for IBM Spyre

Torch-spyre's Inductor determines how each kernel's work is allocated on Spyre's core and manages the residency of tensors between on-chip Scratchpad and off-chip memory, and the backend compiler translates these decisions into instructions that hardware can actually run.

Google Research releases AI-assisted ultrasound technology to empower maternal care in primary healthcare settings

Google Research's AI model is validated on JAMA Network Open for its accuracy in assessing gestational age using blind ultrasound scanning technology

Microsoft researchers and collaborators combined new mathematical ideas with human-guided AI to address the core problem of the KLS conjecture in high-dimensional geometry and probability.

Microsoft researchers and collaborators combined new mathematical ideas with human-guided AI to prove the KLS conjecture in high-dimensional geometry and probability, which has been proposed for 31 years and widely regarded as the central unsolved problem in the field. AI tools explore candidate methods and perform detailed derivations, researchers evaluate, prune, and reorient the search, ultimately proving geometric principles in high-dimensional spaces, making the barrier no longer grow unbounded with increasing dimensions. Independent researchers confirm the result, identify key criteria, and complete formal verification with machine checking.

vLLM enhances the TTFT performance of multi-round agent workloads through bidirectional KV transmission.

vLLM's bidirectional KV transmission mechanism reduces redundant calculations by allowing the decoding node to return previously computed KV cache to the pre-filled node in subsequent dialogue turns. In multi-round benchmark tests on AWS EC2 P5en instances, it improved the TTFT of Qwen3-32B by 1.6x and zai-org/GLM-5-FP8 by 3.3x. This feature is implemented by extending the NIXL KV connector and utilizing a KV cache-aware router to coordinate KV reuse across dialogue turns.

vLLM launches a push-based KV transmission to improve the TTFT and TPOT of distributed inference.

vLLM introduces a push-based NIXL connector, by moving post-pre-filled transmission coordination and setup out of the critical path, reducing time to the first token (TTFT) and time per output token (TPOT), enhancing application responsiveness.

Build a semantic layer using AtScale and Amazon Bedrock AgentCore on AWS to support AI agents.

AtScale as a semantic layer constructs dimensional models on Amazon Redshift using Semantic Modeling Language (SML), supporting Strands agents in Amazon Bedrock AgentCore to achieve cross-system Customer 360 analysis.

Building an AI agent semantic layer on AWS using Strategy Mosaic and Amazon Bedrock AgentCore.

Strategy Mosaic, as a semantic layer, models C360 domains on Amazon Redshift and Amazon Aurora without ETL using two linked models. Its metric syntax captures the business definition of agent reasoning and exposes the model to AgentCore-hosted Strands agents via the Mosaic MCP server, supporting user identity propagation to achieve end-to-end row-level and object-level security.

Building an AI intelligent semantic layer using Denodo and Amazon Bedrock AgentCore on AWS

Denodo platform builds a semantic layer on AWS, providing controlled enterprise data access for Amazon Bedrock AgentCore, supporting federated queries across Amazon Aurora, Amazon Redshift, and local PostgreSQL databases.

Using X.509 PKI to protect AI agent identity on AWS

Keyfactor EJBCA integrated with AWS services provides identity authentication for AI agents through X.509 Public Key Infrastructure (PKI), addressing the issue of lacking built-in authentication standards when MCP connects to external tools.

Fine-tune GPT-OSS-20B on Amazon SageMaker HyperPod for multilingual inference

GPT-OSS-20B is an open-source inference model that thinks in English by default, and only 22% met the target language requirements in multilingual inference tests. By using supervised fine-tuning (SFT) and reinforcement learning (GRPO) on Amazon SageMaker HyperPod, the team can improve the model's inference capabilities in French, German, Spanish, and Italian, where SFT is responsible for primary behavior teaching and GRPO optimizes when there is room for improvement.

i14 on how to use native PyTorch to train and deploy Laplace 3B linear attention model on AWS Trainium2.

i14 deploys its 24-layer decoder Laplace 3B (3,072D parameters, using Gated Linear Attention architecture) to AWS Trainium2, achieving a training speed of 120,000 tokens per second and concurrent inference capability for 11,544 output tokens.

TEAL: A pixel-level efficient active learning framework based on ViT

TEAL has for the first time efficiently applied active learning to the token space of ViT for low-budget pixel-level semantic segmentation, boosting mIoU by +24.36 on datasets like CamVid and Cityscapes with only one pixel labeled per image, and achieving full-supervised performance of 97% with 10 pixels labeled.

OpenAI introduces the complementary metric g-mean² to enhance model monitorability evaluation

OpenAI has introduced the complementary metric g-mean² to improve model monitorability evaluation. g-mean² calculates TPR and TNR through behavioral labels and monitoring predictions, but cannot solve the problems of mixed differences in different model tasks and deployment risks. The new metric includes a paired matching metric for aligned shared instances and a post-control integrated metric simulating monitoring rejection of outputs, to more comprehensively evaluate model deployment risks.

How does ALTK-Evolve surpass state-of-the-art models in the τ²-bench customer service task through learning and memory?

ALTK-Evolve uses GPT-4.1 to learn and apply customer service scripts. In the τ²-bench customer service agent benchmark test, it surpasses all retail entries in the retail sector, achieving a score of 67.5% Pass^4, and does not require updating model weights. The model is based on a basic ReAct agent and GPT-4.1. It attempts to analyze tasks and use GPT-4.1 to distill success and failure into short behavioral rules, then adds these guidelines to the agent's system prompts without updating weights or manual annotation.

Sakana AI Engineer Open House 2026 Event Report

Sakana AI held its second Engineer Open Day, covering five job categories, introducing the work content of each department, demonstrating applications in the finance and defense sectors, and emphasizing AI sovereignty strategy and product line expansion.

How does Claude design protein binders for scorpion neurotoxins?

Claude autonomously generated and screened 318,484 protein designs in 64 hours, selecting 180 for testing, and collaborated with Adaptyv Bio and Twist Bioscience to experimentally validate 211 designs, of which 17 met the binding criteria. All design data has been made public. Experts consider this the first AI design attempt for scorpion neurotoxins, which could accelerate drug discovery for neglected diseases.

Anthropic uses Claude to reduce code's attack surface

How Meta engineers improve the reliability of AI agents by streamlining routers.

Meta engineers built a lean router with 1,300 lines to assign over 20 workflows to 14 specialized AI agents, with the router itself not performing any domain-specific tasks, significantly improving the efficiency and accuracy of data pipeline agents.

Sakana AI proposes 'QD-Rec' to improve recommendation quality and diversity with an open-ended exploration method

Sakana AI proposed 'QD-Rec', recommending candidate-hypothesis pairs as units, with generative AI writing new hypotheses based on past excellent hypotheses in a cycle to accumulate high-quality and diverse recommendations. Experiments were conducted in M&A, customer targeting, and paper recommendations, achieving high average evaluations in many settings and building large sets of recommendations.

Claude reveals the origin of ECDSA curves

Claude discovered within half an hour that the seed values for NIST P-192, P-224, P-256, P-384, and P-521 curves came from the C library's lrand48 random number generator, not any hash algorithm, initialized with Unix timestamps. This discovery did not constitute a cryptographic vulnerability but rather enhanced confidence in the choice of curve parameters without backdoors or intentional weaknesses.

vLLM optimizes Kimi K3: Fast and responsible model tuning practices

vLLM has built a unified CuTe DSL decoding path for Kimi K3 through intelligent agent-assisted development, achieving a batch 1 performance of 840 tokens/s on 16 NVIDIA GB300 GPUs, with a 1.5-2x increase in decoding throughput in the test configuration, and a 1.8x increase in AgentX workload P90 interactivity.

Meta How to Build Forum Applications with Native AI Engineering

Meta uses AI agents to participate in the construction of engineering processes for Forum applications, viewing AI as a participant in the engineering process rather than an auxiliary tool. This experiment measures success by production-ready outcomes rather than lines of code, using a Human-On-The-Loop model to allow AI to autonomously execute planning, implementation, testing, and validation, with engineers responsible for setting goals, supervision, and acceptance.

Selection-Based Structured Reasoning: A framework for efficient multimodal search agents

The Selection-Based Structured Reasoning (SSR) framework restructures reasoning into selection rather than open-ended generation, selecting the most likely solution from predefined reasoning candidates before each action, significantly reducing reasoning costs. Validated using 2B and 4B models on seven multimodal search benchmarks, SSR achieved average success rates comparable to leading search agents of the same scale, reducing single-step reasoning latency by 90% and total reasoning latency per task by 28-54%.

Plan or Retrieve First? A Scale Study of Tool Discovery Architectures

On ToolBench, Retrieve-then-Plan (RtP) outperforms Plan-then-Retrieve (PtR) in single-tool queries, while PtR performs better in multi-tool queries, this crossover holds true across all retrieval budgets (k=1 to 20), registry sizes (1k to 47k APIs), and model families (Gemini 2.5 Flash and gpt-oss-120b).