AI Research

Anthropic analyzed 207,000 research sessions

Lead · Anthropic · 2026-11-06

Anthropic analyzed 207,000 research sessions from scientists using Claude to identify patterns in how AI accelerates discovery. The research examined which scientific tasks received the most use, why some applications see higher adoption than others, and where AI could speed up discovery timelines. Sessions spanned multiple scientific domains and showed measurable differences in how researchers deploy Claude for hypothesis generation, data analysis, literature review, and experimental design. The findings map how working scientists actually use foundation models in practice.

Shows where AI adds real value in scientific workflows and identifies bottlenecks for practitioners building AI-native research tools.

• 207,000 research sessions analyzed across domains
• Measured adoption patterns by scientific task type
• Identified opportunities to accelerate discovery pipelines

Anthropic deployed Claude in EVE Online gameplay

Anthropic · 2026-11-06

Anthropic placed Claude in EVE Online, a complex multiplayer online game with markets, PvP combat, and large-scale teamwork, to observe agent behavior under realistic constraints. The study tracked how Claude navigates market dynamics, collaborates with other players, experiences failures, and handles novel scenarios that require adaptation. The researchers documented specific failure modes where the AI misunderstands incentives or breaks down under pressure, and evaluated what safeguards might be needed for autonomous systems operating in social environments. Results identify both capabilities and gaps in AI agent reasoning about complex social and economic systems.

EVE Online is a testbed for agent behavior in realistic, adversarial conditions—findings inform safety design for autonomous systems deployed in multi-agent environments.

• Tested Claude in complex social and economic gameplay
• Documented specific failure modes and breakdown scenarios
• Identified safeguard requirements for multi-agent systems

OpenAI releases BioBasicsBench for wet-lab AI agents

OpenAI · 2026-11-06

OpenAI introduced BioBasicsBench, a benchmark with 100 tasks covering wet-lab science workflows. The benchmark evaluates how AI agents interpret raw experimental data, design assays, troubleshoot protocols, and handle practical laboratory challenges. Tasks measure agent ability to parse instrument readouts, identify experimental errors, propose corrections, and plan next steps based on incomplete or noisy data typical of real lab work. The benchmark provides a quantitative evaluation framework for assessing agent readiness to support experimental biologists.

Establishes measurable performance criteria for AI assistants in experimental science, enabling reproducible evaluation of agent capabilities for wet-lab tasks.

• 100 tasks covering wet-lab workflows and troubleshooting
• Measures interpretation of raw experimental data
• Evaluates real-world lab-readiness for AI agents

Anthropic reports early progress on AI-accelerated drug discovery

Anthropic · 2026-11-06

Anthropic shared progress from partnerships with six organizations using Claude to accelerate drug discovery timelines. The effort includes internal programs focused on rare childhood forms of muscular dystrophy, leveraging Claude's ability to analyze literature, predict molecular interactions, and design compounds. Early results demonstrate reduced time from target identification to lead compound generation. The work combines Claude's reasoning with domain expertise to overcome bottlenecks in translating biological insights into therapeutic candidates.

Shows early evidence that frontier AI can compress drug discovery timelines on real targets, with potential to expand access to treatments for rare disease.

• Six partnerships with drug discovery organizations
• Internal program targeting rare muscular dystrophy forms
• Early acceleration of target-to-lead timelines

Anthropic studies agent memory through simulated dreaming

Anthropic · 2026-11-06

Anthropic investigated how Claude learns from experience by studying memory formation and organization. The work explores how agents decide what information is worth retaining before knowing future context—a core challenge in lifelong learning. The team found that simulated 'dreaming' (replaying past experiences) helps agents organize memories more effectively and recover from mistakes. The approach surfaces which memories matter for future success and which can be discarded, improving both memory efficiency and agent accuracy over time.

Addresses core problem in building agents that improve through interaction; shows memory consolidation methods applicable to long-horizon autonomous systems.

• Studies memory formation in continuous learning
• Shows 'dreaming' improves memory organization
• Enables agents to correct past mistakes

OpenAI introduces GPT-6 Cyber for aligned cybersecurity

OpenAI · 2026-11-05

OpenAI released GPT-6 Cyber, a specialized model trained for cybersecurity work to identify exploitable vulnerabilities and develop fixes. The model uses reinforcement learning and constitutional guidance to remain within authorized testing scope, avoiding actions like scanning unauthorized systems or stealing credentials. Training emphasizes responsible disclosure and coordination with affected organizations. The system demonstrates how frontier models can be specialized for high-stakes domains where scope violations create real legal and ethical risks.

Shows practical method for training high-capability models to respect boundaries in security work, relevant for all regulated AI agent deployment.

• Finds exploitable vulnerabilities across systems
• Constrained to authorized scope only
• Trained for responsible disclosure and coordination

VFlash uses visual primitives for fast robot planning

HF Papers · 2026-11-06

Vision-language models trained on robotics and autonomous driving tasks can plan via visual primitives—points, bounding boxes, trajectories—rather than language tokens alone. VFlash applies flash primitive decoding to generate these spatial plans in parallel rather than autoregressive token-by-token, dramatically reducing latency. The method combines fast decoding of visual tokens with accurate spatial reasoning needed for physical control. Evaluation on robotics and autonomous driving benchmarks shows speedup while maintaining or improving plan quality.

Reduces inference latency for robot and autonomous vehicle planning, enabling real-time performance on edge hardware critical for embodied AI.

• Parallel decoding of visual plan primitives
• Significant latency reduction over autoregressive approaches
• Maintains or improves spatial planning accuracy

Microsoft Research formally verifies AI-generated code with PDD Kit

Microsoft Research · 2026-11-05

Microsoft Research introduced PDD Kit, a verification framework for code generated by AI agents at scale. The system converts developer intent into a Verus specification (a formal language for proving program correctness), then automatically verifies that AI-generated code satisfies the specification for every possible input. This approach replaces line-by-line code review with exhaustive formal verification, catching subtle bugs that escape testing. The kit is designed to handle the volume of AI-generated code that humans cannot practically review manually.

Enables safe deployment of AI code generation for critical systems by providing machine-checkable correctness proofs instead of manual review.

• Converts developer intent to formal specifications
• Proves correctness for all possible inputs
• Scales to high-volume AI code generation

SLA scales vision-language-action models via continual pretraining

HF Papers · 2026-11-06

SLA (Spatial-Language-Action model) uses continual pretraining and post-training to improve robotic policy generalization. The approach aligns a pretrained vision-language-action (VLA) policy with a frozen 3D perception model, using feature-dependent gates and probabilistic state dropout to preserve performance while improving spatial understanding. Evaluation across diverse robot datasets shows improved zero-shot transfer and downstream task performance. The method maintains the original policy inference architecture while enriching spatial reasoning capabilities.

Improves robot generalization without retraining from scratch, enabling faster adaptation to new tasks and environments.

• Continual pretraining improves spatial understanding
• Maintains original inference architecture
• Better zero-shot transfer across diverse robot data

EvoMemState learns evolving memory for personal agents

HF Papers · 2026-11-06

EvoMemState addresses limitations of fact-based long-term memory in LLM agents by maintaining memories as evolving states that capture user information across interactions. Rather than storing isolated facts, the system tracks how user preferences, constraints, and context change over time. This enables agents to understand patterns in user behavior and anticipate needs. The approach improves both accuracy and relevance of agent assistance in personalized assistant scenarios requiring retention across many sessions.

Enables more natural and anticipatory personal assistants by capturing user evolution rather than static snapshots.

• Structures memory as evolving user states
• Captures patterns across many sessions
• Improves personalization accuracy

PyRUA-REAL uses code agents for real-world robot learning

HF Papers · 2026-11-05

PyRUA-REAL enables coding agents to learn directly from real-world robot exploration by converting agent code into persistent, reusable skills. The system achieves high success rates on both benchmark and real-world tasks through recursive self-improvement—agents refine code based on execution feedback. The framework bridges the gap between simulation and real hardware by allowing agents to write, test, and refine executable code in the real environment, avoiding sim-to-real transfer gaps.

Accelerates real-world robot capability by enabling agents to learn and persist skills directly from physical deployment.

• Agents write persistent executable robot skills
• Recursive self-improvement from real deployment
• High success on real-world and benchmark tasks

Framework proposes cognitive symbionts for human-AI teams

HF Papers · 2026-11-06

A conceptual framework extends physical AI beyond autonomous agents to AI-augmented human intelligence through lifelong cognitive symbionts. These systems sense environment and context, retain information across sessions, and proactively assist to close cognitive gaps. The framework emphasizes embedded governance requirements—ensuring these intimate human-AI partnerships maintain appropriate human agency and oversight. The approach reframes AI from autonomous replacement to persistent cognitive partnership designed for long-term human-AI collaboration.

Shifts focus from autonomous AI to collaborative augmentation, providing design principles for AI systems that enhance rather than replace human judgment.

• Lifelong cognitive partnership model
• Sensing, retention, and proactive assistance
• Embedded governance for human agency

Anthropic's science grantees discuss AI-driven scientific models

Anthropic · 2026-11-06

Anthropic's AI for Science grantees Jakob Macke, Daniel Gedon, and Stefan Wahl discuss emerging ways AI is reshaping scientific modeling. Their work shows how AI can accelerate model discovery, help interpret complex systems, and suggest new experimental directions. The grantees highlight cases where AI tools help scientists discover interpretable models faster and guide them toward testable hypotheses. The discussion emphasizes AI as a partner in the scientific process rather than a replacement for domain expertise.

Shows how frontier AI models are already being integrated into scientific workflows to accelerate hypothesis generation and model discovery.

• AI accelerates model discovery and interpretation
• Suggests testable hypotheses to researchers
• Enables faster scientific hypothesis generation