LLMs & Models
Large language models: GPT, Claude, Gemini, Mistral and open weights.
9 published articles
AI Research
Your AI agent passed the test by accident. Now there's a rubric for that.
SkillCoach is a self-evolving rubric framework that evaluates and improves agentic skill-use by analyzing skill selection, following, composition, and reflection processes, providing better supervision than outcome-only metrics.
2026-07-06
Model Evaluation
Ai2's olmo-eval gives LLM developers a microscope for every checkpoint
Ai2's olmo-eval brings per-question diffs and modular benchmarks to active LLM development, helping researchers tell real progress from statistical noise.
2026-07-06
Deep Learning
M3D and Real-Guidance Bring Dataset Distillation to High-Resolution Realms
Dataset distillation has long been stuck on low-res benchmarks, but a new approach called M3D changes that. By combining multi-scale matching, a data manifold prior, and a Real-guidance strategy, it scales to ImageNet-1K at 128×128 resolution, achieving 68.5% top-1 accuracy with just one image per class and cutting memory usage by ten times.
2026-07-05
AI Research
The verification horizon: why verifying coding agents is now harder than building them
A classical intuition holds that verifying a solution is easier than producing one. For today's coding agents, that intuition has inverted: generating complex solutions is now easy. The hard part is reliably verifying them.
2026-07-05
AI Research & Development
Ma Jiaqi taught MiniMax engineers a hard lesson about forgotten tokens
MiniMax's internal investigation into why its M2 model couldn't output the name 'Ma Jiaqi' revealed a structural mismatch between pre-training vocabulary and post-training data distribution. The root cause: low-frequency tokens' lm_head vectors drift during SFT, losing generation ability while retaining understanding. A full-vocabulary coverage fix resolved the issue and also mitigated language mixing in Japanese.
2026-07-04
Artificial Intelligence
MiniMax's new M2.5 coding model tops the benchmark at 5% of the price
MiniMax's M2.5 model tops the Multi-SWE-Bench coding benchmark, beats mainstream models on workspace tasks, and costs a tenth to a twentieth of competitors. Open-source weights are on HuggingFace.
2026-07-04
Artificial Intelligence
Microsoft's Phi-4 Model Redefines Efficiency in Breakthrough Research
Microsoft's Phi-4 model achieves state-of-the-art efficiency, matching larger models in reasoning tasks with significantly fewer parameters. Published on May 15, 2025, the research paper reexamines assumptions about scaling laws in AI.
2026-07-03
Tokenizer-Free Architecture
Aleph Alpha unveils T-Free: a tokenizer-free architecture for sovereign AI
Aleph Alpha unveils T-Free, a tokenizer-free LLM architecture that maps words directly to vectors. The approach delivers nearly seven characters per vector versus the typical four, cutting costs and energy use while improving performance on specialized domains and low-resource languages.
2026-07-03
effective context, output ceilings, and the hidden tax of long windows
Your AI model says it can read 1 million tokens. It's lying. Here's the real math.
All four frontier LLMs advertise 1M+ token contexts, but effective recall, output limits, and real-world cost differ sharply. DeepSeek V4 Pro leads in output ceiling and cost, Gemini excels under 200K tokens, and Claude Opus wins on caching for interactive code review. This analysis breaks down the numbers from April 2026.
2026-06-30