Category: Research
arXiv-grounded analysis of papers, benchmarks, and model architectures with attention to reproducibility and honest limitations. Recent coverage includes MCPShield’s mapping of 23 attack vectors across the Model Context Protocol’s 97-million-download ecosystem, Anthropic’s discovery of 171 emotion vectors inside Claude Sonnet 4.5, the five-layer architecture inside Claude Code where 98.4% of the system is operational infrastructure rather than model inference, Google’s TurboQuant KV cache compression breakthrough that six independent teams found does not actually deliver the claimed gains, and Gemini 3.1 Pro cutting hallucination rates 38 points without learning anything new through pure calibration.
The editorial standard goes beyond what the abstract says. Every research piece states which baselines were tested, which were silently omitted, what the ablation results actually demonstrate, and where the authors stretched their claims past what the experimental evidence supports. Benchmarks are evaluated by what they fail to test, not just by reported scores.
Coverage extends to ARC-AGI-3 on Kaggle and why current models score in the low double digits, ICLR 2026 outstanding papers and the review crisis around them, the Synthetic Web Benchmark that collapsed every frontier AI agent with a single fake article, the Darwin Godel Machine that rewrites its own code on SWE-bench, and Sakana’s AI Scientist accepted into Nature peer review. No hype. No extrapolation. No taking the Twitter summary at face value.
-
OWASP LLM Top 10 for 2025: The Mechanism Behind Each Vulnerability
The OWASP LLM Top 10 for 2025 added System Prompt Leakage and Vector Weaknesses, reworked Excessive Agency, and moved Sensitive Disclosure to second place. Here is the architectural…
-
Indirect Prompt Injection: The Attack That Hides in Your Data
Indirect prompt injection lets attackers hijack LLMs by hiding instructions in documents, web pages, and tool results the model processes. Here is why the architecture makes this unavoidable…
-
Julia Bazinska and the Science of Measurable AI Security
Julia Bazinska built the empirical tools that make LLM security measurable. From DeepMind RL to first-authoring b3, here is what her research at Lakera actually produced.
-
Gandalf the Red: What 279K Real Attacks Reveal About LLM Defense
Lakera’s ICML 2025 paper ran 279K crowdsourced attacks to show what synthetic red-teaming misses. The D-SEC finding: system prompts degrade user experience without blocking attackers. Here is the…
-
Vision-Language Models: Architecture and the Benchmark Gap
How CLIP, SigLIP, Q-Former, and MLP adapters work in vision-language models. Why Qwen2.5-VL compresses visual tokens 4x, and what current VLMs still cannot do.
-
Chinchilla Scaling Laws: Three Methods and Why Labs Ignore Them
Chinchilla proved GPT-3 was undertrained. The 20:1 rule is a training-compute floor. Three methods, their disagreements, and why frontier labs now exceed it.
-
LoRA and QLoRA: Fine-Tuning Large Models on One GPU
LoRA fine-tunes 70B models on one GPU using low-rank weight updates. The intrinsic dimension proof, rsLoRA scaling fix, and where LoRA falls short.
-
Speculative Decoding: How LLMs Generate 3x Faster
Speculative decoding achieves 3-4x LLM speedup with zero output quality loss. The math proof, EAGLE-2’s 4.26x result, and when it does not help.
-
MiniMax M2.7 Optimized Its Own Training Harness 100 Times. Here Is the Loop.
MiniMax M2.7 ran an internal agent that modified its own training scaffold 100 times in a row without human input and gained 30% on internal evaluations. Here is…
-
KellyBench: 8 AI Models Bet the Premier League. All Lost Money.
General Reasoning put 8 frontier AI models through a full Premier League season with a 100k bankroll each. Every model lost money. The benchmark reveals three distinct failure…
-
DeepSeek V4’s Hybrid Attention Cuts KV Cache by 10x. Here’s the Architecture.
DeepSeek V4-Pro processes one million tokens using 10% of the KV cache V3.2 needed. The mechanism is Hybrid Attention: two complementary compressors interleaved across 61 layers. Here’s how…
-
30 Days After QJL: What’s Actually Compressing the KV Cache
After QJL failed, three approaches own the KV cache frontier: TriAttention’s pre-RoPE selection, LRKV architectural compression, and adaptive bit-width.
-
ICLR 2026 Outstanding Papers: What They Actually Found, and the Review Crisis Around Them
ICLR 2026 named two outstanding papers: LLMs Get Lost In Multi-Turn Conversation and Transformers are Inherently Succinct. The conference also documented a 45% identity leak and 21% AI-generated…
-
Agent Memory Architecture: Four Patterns, Four Tradeoffs
Agent memory is not one thing. It is four distinct patterns: full context window, hierarchical summarization, external vector store, and episodic log. Each has different performance, cost, failure…
-
Know Your Agent: The First Regulated AI Agent Governance Standard
MetaComp’s StableX KYA Framework, published April 21, 2026, is the first governance standard for AI agents from a licensed financial institution. Here’s what its four pillars cover, how…
-
A2A Protocol v1.0: The Agent Communication Layer MCP Doesn’t Cover
A2A Protocol v1.0 introduced Signed Agent Cards and gRPC support. Here’s how agent-to-agent communication differs from MCP tool calls, why IBM merged ACP into A2A, and what the…
-
5 of 7 Major MCP Clients Don’t Validate Tool Metadata. Here’s the Gap.
5 of 7 major MCP clients tested skip static validation of tool metadata entirely. A March 2026 arXiv paper is the first systematic evaluation of MCP client-side security,…
-
MCP-SafetyBench at ICLR 2026: No LLM Agent Can Be Both Useful and Secure
MCP-SafetyBench at ICLR 2026 finds a negative correlation between defense success and task success across all 20 MCP attack types. No model achieves both. Here’s what the tradeoff…
-
Full Context Sets the Accuracy Ceiling for AI Agent Memory. It Costs 26,000 Tokens Per Query. Here Is the Tradeoff Map.
Full context memory sets the accuracy ceiling at a cost of 26,000 tokens per query. Vector-only memory scores 66.9% at 1.44s p95 latency. Graph memory reaches 68.4% at…
-
98.4% of Claude Code Is Operational Infrastructure. A New arXiv Paper Maps All of It.
A source-code analysis of Claude Code’s 512,000-line TypeScript codebase finds 98.4% is operational infrastructure, not AI. Here is the five-layer compaction pipeline, the 17% comprehension decline finding, the…




















You must be logged in to post a comment.