AI Research

Anthropic: Claude agents discover novel enzyme system

Lead · Anthropic · 2026-10-13

Anthropic launched a new life sciences research lab and deployed Claude agents to search for novel biological insights. The agents discovered an enzyme system whose exact function remains unknown, representing an early success in applying AI reasoning to biological discovery. This finding emerged from systematic exploration of biological data and demonstrates Claude's capability to identify patterns that warrant further investigation by domain experts.

The result shows AI agents can make novel discoveries in biology beyond pattern recognition, suggesting new pathways for accelerating scientific research.

• Claude agents identified an enzyme system of unknown function
• First output from new Anthropic life sciences research lab
• Demonstrates AI capability for biological discovery beyond existing knowledge

Anthropic: Defenses reduce model distillation by over 90%

Anthropic · 2026-10-13

Anthropic strengthened its defenses against distillation attacks, where actors attempt to copy frontier AI capabilities without creating proportionally robust safeguards. The company achieved a more-than-90% reduction in the distillation rate through targeted defense strategies. The finding addresses a critical security gap in protecting proprietary model capabilities from unauthorized extraction and copying.

Effective defenses against distillation are essential for preventing unauthorized capability replication and maintaining security of frontier models.

• More than 90% reduction in distillation success rate
• Defense prevents copying frontier capabilities without robust safeguards
• Critical for frontier model security

Google DeepMind: Accelerating antimicrobial peptide discovery with Antigravity

Google DeepMind · 2026-10-15

Google DeepMind deployed Antigravity for Science to systematically search for novel antimicrobial peptides. The approach accelerates the discovery process by efficiently exploring the structural landscape of candidate molecules. Antimicrobial peptides represent a promising avenue for combating resistant infections and addressing emerging health threats.

Accelerating antimicrobial discovery through AI could help address antibiotic resistance and support development of new therapeutics.

• Antigravity applied to antimicrobial peptide discovery
• Searches for novel structural candidates systematically
• Addresses antibiotic resistance challenge

Microsoft Research: Magnetar brings zero-knowledge proofs to mainstream computing

Microsoft Research · 2026-10-15

Microsoft Research introduced Magnetar, a system that makes zero-knowledge proofs practical for general-purpose computing. The system supports standard programming languages, delivers major speed improvements, and provides quantum-resistant security. Zero-knowledge proofs allow verification of computing results without rerunning workloads or exposing sensitive data, addressing privacy and efficiency concerns in cloud computing.

Practical zero-knowledge proofs enable privacy-preserving computation at scale, transforming how sensitive workloads can be verified securely.

• Supports standard programming languages for ZK proofs
• Major speed improvements over prior approaches
• Quantum-resistant security built in

Cortex: Consolidating context into parametric memory

HF Papers · 2026-10-15

Cortex is a method for consolidating improvements gained from context into parametric model updates that can be merged persistently. The approach uses low-rank or structural updates that preserve performance on non-target capabilities across sequential edits and diverse model scales. This enables efficient adaptation of models to new contexts without catastrophic forgetting.

Efficient consolidation of contextual improvements enables practical sequential model adaptation for diverse downstream tasks.

• Low-rank or structural parametric updates
• Preserves non-target capabilities across sequential edits
• Works across diverse model scales

RuleTrek: Benchmarking agentic rule reasoning in complex scenarios

HF Papers · 2026-10-15

RuleTrek is a multi-domain benchmark that evaluates large language model agents on their capability to identify, deduce, and apply domain-specific rules across complex, multi-file scenarios. The benchmark reveals substantial performance gaps in frontier models' reasoning about rules in realistic settings. Performance varies significantly across different domains and rule complexity levels.

Understanding rule reasoning limitations in agents is critical for deploying them in regulated domains requiring consistent application of constraints.

• Multi-domain benchmark for rule reasoning evaluation
• Evaluates identification, deduction, and application of rules
• Reveals substantial performance gaps in frontier models

ReMaRk: Improving reinforcement learning for LLM recommenders

HF Papers · 2026-10-15

ReMaRk improves reinforcement learning for recommendation systems by addressing responder noise through joint scoring of generated lists against all valid replies. The approach reduces training cost while matching or exceeding the performance of prior methods. This enables more robust and sample-efficient learning for recommendation tasks.

Efficient reinforcement learning for recommendations accelerates deployment of better-performing recommendation systems at scale.

• Scores lists against all valid replies jointly
• Reduces training cost vs. prior approaches
• Matches or exceeds prior performance

Apertus 1.5: Open multimodal reasoning with audio and images

HF Papers · 2026-10-15

Apertus 1.5 advances open multimodal language models to support image and audio reasoning capabilities. The model uses long-context training and reinforcement learning, with fully open data and weights released publicly. This represents progress toward open-weight multimodal reasoning systems that match capabilities of closed models.

Open multimodal models with reasoning capabilities enable researchers to build on frontier work without proprietary dependencies.

• Open-weight multimodal model with audio and image reasoning
• Long-context training and reinforcement learning
• Fully open data and weights

CLASHarness: Multi-agent conflicts in computer use tasks

HF Papers · 2026-10-15

CLASHarness introduces a controllable framework for generating verifiable multi-agent conflicts in computer-use tasks. Testing on frontier vision-language models reveals significant struggles with verifying ground truth, resolving conflicts, and coordinating shared resources. The controllable generation pipeline enables systematic evaluation of agent robustness in adversarial multi-agent settings.

Understanding agent vulnerabilities in multi-agent scenarios is essential before deploying autonomous systems in shared environments.

• Controllable conflict generation for computer-use tasks
• Frontier models struggle with verification and conflict resolution
• Tests resource coordination in shared workspaces

The thinking leakage: Reinforcement learning in hybrid reasoning models

HF Papers · 2026-10-15

Analysis of reinforcement learning in hybrid reasoning models reveals that improvements in NO-THINK mode largely result from reactivating existing THINK-mode reasoning capabilities rather than learning fundamentally new capabilities. This finding challenges assumptions about where RL-driven gains originate in reasoning models. The result has implications for how to design training and evaluation of reasoning-enhanced models.

Understanding the source of reasoning improvements guides more effective training of reasoning-enhanced models and better evaluation of true capability gains.

• NO-THINK gains from reactivated reasoning, not new learning
• Challenges assumptions about RL-driven improvements
• Guides reasoning model training design

Super-resolution via frozen vision-foundation embeddings

HF Papers · 2026-10-15

A method for super-resolution maps degraded image embeddings back to a high-resolution manifold within a frozen DINOv3-L latent space. The approach yields high fidelity results with fast inference without requiring task-specific training of the foundation model. This enables practical super-resolution that leverages existing frozen vision encoders.

Leveraging frozen foundation models for super-resolution enables practical deployment without expensive model fine-tuning.

• Uses frozen DINOv3-L latent space
• High fidelity and fast inference
• No task-specific model training required

LabInstruct: Evaluating AI video generation for lab procedures

HF Papers · 2026-10-15

LabInstruct is a benchmark for evaluating AI-generated instructional videos for real laboratory tasks. The evaluation reveals that visual plausibility and aesthetic quality do not guarantee procedural correctness, exposing a critical gap in video generation systems. Many visually convincing videos fail to capture essential procedural steps accurately.

This finding is critical for deploying AI-generated training content in lab settings where procedural correctness is non-negotiable.

• Evaluates AI-generated lab instructional videos
• Visual plausibility ≠ procedural correctness
• Reveals critical gaps in video generation

APEX-Harm: Agent vulnerabilities in multi-tool professional workspaces

HF Papers · 2026-10-15

APEX-Harm evaluates long-horizon agent vulnerabilities to injection attacks across large multi-tool workspaces. Testing reveals high success rates for script-based attacks even when agents produce correct task outputs. This demonstrates that agents can be compromised through injection attacks on intermediate steps while maintaining apparent correctness.

Understanding injection attack vulnerabilities is essential before deploying autonomous agents in shared tool environments with security implications.

• Tests long-horizon agent vulnerabilities to injection attacks
• High success rates for script-based attacks
• Agents compromised despite correct outputs