What's this about?
Two papers. LightNav-0 elicits spatial intelligence already present in vision-language models for generalist embodied navigation (24 upvotes). SHAPE examines the shape of chain-of-thought in mathematical reasoning (26 upvotes).
What does "eliciting" mean as distinct from training?
That the capability is already in the model and the contribution is getting it out. Vision-language models trained on internet images have absorbed a great deal about spatial layout without being trained to navigate; eliciting means finding the prompting, structure or lightweight adaptation that surfaces it. It is a much cheaper claim than training a navigation model and a more interesting one, because it says the bottleneck was access rather than knowledge.
Why is it worth noticing?
- Elicitation results are practical — they need no new training run.
- SHAPE's angle is structural: not whether the reasoning is right, but what form it takes, which connects to the faithfulness work elsewhere in this digest.
- Both are modest-upvote papers doing careful work rather than making large claims.