Every open-weights agent model that emits tool calls in the Qwen/LFM3 XML framing writes each parameter value between one newline on either side. We rendered 1,762 real files as write_file calls and parsed them back with five…
An independent research lab
Hybrid language models, measured carefully.
We study hybrid language models — how attention, recurrence and retrieval are put together — and the measurement methods that decide whether a change is real.
Recent work
For an odd-length list, the middle item can be requested two ways that name the same item at the same ordinal: “5th position from the beginning” and “5th position from the end”. Only the anchor word differs. On…
Programmes
When a model mis-indexes a list, what exactly did it get wrong?
What would make a small-model result worth believing?
How we work
One variable per experiment. Practical margins are written down before the run, and anything chosen afterwards is labelled as such. Negative results ship with the positive ones, and the analysis path runs on the Python standard library alone.