Best AI Models for Text Summarization - March 2026
Gemini 2.5 Flash Lite leads the Vectara hallucination leaderboard at 3.3% error rate while GPT-4o and Gemini 2.5 Pro dominate long-document tasks - full rankings, benchmark scores, and pricing.
Articles Tagged "Benchmarks"
JBDistill Generates Its Own Jailbreaks - 81.8% Attack Rate
Johns Hopkins and Microsoft's JBDistill achieves 81.8% attack success rate across 13 LLMs by auto-generating fresh adversarial prompts on demand.
VLMs Fail Physics Tests, RL Quits Bad Paths, Agents Lie
Three new papers expose systematic VLM failures on basic physics, introduce RL that learns to abandon bad reasoning paths, and reveal that AI agents deceive primarily through misdirection rather than fabrication.
AI Safety Leaderboard: Refusal and Jailbreak Rankings
Rankings of AI models by safety metrics including refusal rates, jailbreak resistance, bias scores, and truthfulness across major benchmarks.
Reasoning Models Can't Hide Their Thinking - OpenAI Study
OpenAI's CoT-Control benchmark shows frontier reasoning models score 0.1-15.4% at steering their own chain of thought - a result the company frames as good news for AI oversight.
AI Speed and Latency Leaderboard: Tokens/s Rankings
Rankings of the fastest AI models and inference providers by tokens per second, time to first token, and end-to-end latency.
Embedding Model Leaderboard: MTEB Rankings March 2026
Rankings of the best embedding models by MTEB scores, comparing retrieval quality, dimensions, speed, and pricing for RAG and search.
Multilingual LLM Leaderboard: March 2026 Rankings
Rankings of the best AI models for multilingual tasks, covering 16 languages across the Artificial Analysis Multilingual Index and MGSM benchmarks.
CoT Control, Hidden Beliefs, and Dynamic Agent Benchmarks
New research shows reasoning models can't suppress their chain-of-thought, that they commit to answers internally long before their CoT reveals it, and that static benchmarks are inadequate for measuring real-world agent adaptability.
75% of AI Coding Agents Break Working Code Over Time
Alibaba's SWE-CI benchmark tested 18 AI models on 100 real codebases across 233 days of maintenance. Most agents accumulate technical debt and break previously working code. Only Claude Opus stays above 50% zero-regression.
GPT-5.4 Review: The Computer-Use Frontier
GPT-5.4 brings native computer use, a 1M token context window, and serious coding muscle to OpenAI's mainline model - but at a premium price.
Sandbagging Models, Sparse Critics, Compact Reasoning
New research reveals models can fake poor performance under adversarial prompts, a smarter critic improves SWE-bench by 15 points, and Microsoft shows compact vision models can punch above their weight.
