Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-16 · high
Getting there required clipped importance sampling and mixed-precision fixes, with training splitting into a discovery phase then a sharpening phase.
Why now: The paper posted to arXiv in July 2026 and landed in the July 16 research coverage of this trillion-parameter run.
Researchers scaled a zero-RL training run to 1 trillion parameters. The resulting model developed a behavior the team calls context anxiety, per a new paper posted to arXiv.
The paper's testing something specific: whether reinforcement learning without a supervised warm-up scales all the way to frontier size. At 1 trillion parameters, the answer here is yes, with the resulting model holding competitive across seven math benchmarks.
The 16-author paper reports that scaling to 1 trillion parameters "significantly enhances sample efficiency and performance ceilings." Training split cleanly into two stages, a discovery phase followed by a sharpening phase.
Getting it stable took clipped importance sampling, training-inference ratio correction, and mixed-precision control, according to the authors.
Along the way the model picked up behaviors nobody explicitly trained it to have: anthropomorphism, structured formatting, self-verification, parallel reasoning, and context anxiety. The paper lists context anxiety as emergent without defining what it is or whether it helped the math scores.
Each link below shares sources, entities, or timing with this story.
Shared entity: Zero / Same source domain / Earlier coverage
Both cover Zero; reported by the same outlet (arxiv.org); earlier Zero coverage from 2026-05-25.
Shared entity: Zero / Shared topic / Earlier coverage
Both cover Zero; overlapping topics (context, control); earlier Zero coverage from 2026-04-13.
Shared entity: Zero / Same source domain / Earlier coverage
Both cover Zero; reported by the same outlet (arxiv.org); earlier Zero coverage from 2026-03-02.
Both cover Zero; reported by the same outlet (arxiv.org); earlier Zero coverage from 2026-02-22.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (behavior, benchmark); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (benchmark, control); pushes against this story (against).
Reported by the same outlet (arxiv.org); overlapping topics (benchmark, context); pushes against this story (versus).
Reported by the same outlet (arxiv.org); overlapping topics (author, behavior); pushes against this story (against).