k3 report's reasoning-effort control: fine-grained juice-like values and an unusual GRPO setup
stochasticchasm · x · 2026-09-11
@stochasticchasm reviews the reasoning-effort control scheme in the k3 report:
- It resembles OpenAI's "juice" value, offering finer-grained control of reasoning effort.
- The GRPO setup is unusual: each group is sampled at one effort level, while each task is sampled at multiple effort levels.
- The reference length Lnorm looks similar to k3's calibrated per-problem reference length, but the report never explains how this reference length is obtained — a gap he flags, linking the relevant section.
A technically dense thread on training methodology details.
Related event: k3 report details: 50M sandboxes and reasoning-effort control(2 posts)→
More from Research
- Rachel Thomas points to classic paper on articulating the value of not building tech — math_rachel · 2026-10-28
- Meta's Omnilingual MT covers 1,600 languages, 1B–8B models rival 70B LLMs — kalyan_kpl · 2026-10-28
- Supercharging LLM classification with logprobs: soft classifiers you can threshold-tune — JnBrymn · 2026-10-28
- cryoET Data Portal releases three datasets benchmarking crossed laser phase plate — ZhongingAlong · 2026-10-28
- Blind Robot Hand Spins Baoding Balls 225 Times With Only Touch — Scobleizer · 2026-10-28
- Rollback: World Models Rewind Robot Decisions to Explore Counterfactual Actions — agibot-world · 2026-10-28