New numbers on iterative LLM self-repair: where extra tries earn their keep, where they quietly waste tokens, and the retry loop I'll actually defend.
A new paper tested 8 reasoning models with 14 formatting perturbations. Open-weight models lost up to 55% accuracy. Here's what that means for production use.
A practical look at which new open-weight LLMs you can actually run on a single workstation with llama.cpp in 2026, and which ones are just…
New research scores 263 tasks across four LLMs. Math and coding top the automation list, but the jobs most exposed to AI demand the opposite…
New research shows LLM API testing misses how chatbots actually behave. Here's what the data says about the gap and how to fix your eval…
LLMs claim massive context windows but choke on long number sequences. A training-free trick called SepSeq fixes attention dispersion and boosts accuracy 35%.
A new benchmark shows LLMs drop 0.3 to 5.9% accuracy on grade-school math when names and places go non-Western. Same arithmetic, different answers.
A new Transformer variant called FBS tries to let LLM inference preview, skim, and skip, instead of grinding through every token. I read the paper…
A paper trains a tiny probe on a model's own hidden states to catch hallucinations at inference time, no judge model required. Here's why that…
Open models closed most of the quality and ops gaps with proprietary APIs. A pragmatic look at what changed and how to decide whether to…