Research
Anthropic's Cryptographic Research: Pattern Matching vs. True Reasoning
A critical look at whether LLMs can truly discover cryptographic vulnerabilities or if they are simply acting as expensive pattern matchers.
Research
Research
Papers that actually matter
49 articles in this section.
Research
OlmoEarth: Scaling Geospatial AI Through Planetary-Scale Infrastructure
AI2's OlmoEarth focuses on the critical infrastructure needed to move geospatial AI from simple image recognition to planetary-scale reasoning and inference.
Research
Evaluating the Gap Between Human and Automated MeSH Tagging
A recent paper argues that the gap between human and AI MeSH tagging is a failure of evaluation benchmarks rather than model intelligence.
Research
KwaiKAT's KAT-Coder-V2.5: Shifting Focus from Model Scale to Training Infrastructure
The KwaiKAT team argues that agentic coding capability depends on verifiable repository environments rather than simply increasing model parameter counts.
NCN
Research
OpenAI's Autonomous Hacking Agent and the Future of Cybersecurity
An analysis of OpenAI's new autonomous hacking agent and the risks it poses to global security and the defensive arms race.
NCN
Research
Perplexity AI Releases WANDR Benchmark for Deep Research Agents
Perplexity's new WANDR benchmark reveals a significant gap between agentic search and rigorous, evidence-backed research capabilities in current AI models.
Research
OriginBlame: Enabling Precise Data Unlearning in Large Language Models
OriginBlame introduces a provenance tracking system to create accurate forget sets, allowing model trainers to remove specific author data from LLM weights.
NCN
Research
Anthropic's World Models and the Limits of AI Interpretability
A critical look at Anthropic's efforts to map AI internal representations and why interpretability may be a vanity project compared to reliability.
Research
The Failure of AI Coding Benchmarks and the Shift to Real-World Signal
The industry is moving away from contaminated static benchmarks like HumanEval toward private, dynamic evaluations that measure actual software engineering capabilities.
Research
Anthropic's Global Workspace Theory: Moving Toward Controllable AI
Anthropic is exploring how LLMs use a central internal workspace to integrate information, potentially enabling direct steering of AI reasoning and reliability.