Two manuscripts · one architecture · every number committed

The manuscripts

Both manuscripts compile clean from source (NMI: xelatex nmi_paper.tex · IEEE: pdflatex ieee_paper.tex). Result tables are populated by the committed experiment harness specified in the protocol — no hand-typed numbers.

NATURE MACHINE INTELLIGENCE FORMAT

Compositional Generalization in Embodied Foundation Models via Multimodal Subgoal Prompting and Flow-Based Action Execution

8 pages · full architecture and safety theory (no-jerk theorem, forward-invariance proposition, Grönwall bound) · the three chaotic benchmarks · pre-registered zero-shot protocol · discussion that states the falsification conditions.

nmi_paper.pdf → · source →

IEEE CONFERENCE FORMAT (IEEETRAN)

Steerable Vision-Language-Action Flow Matching: Multimodal Subgoal Prompting and Runtime-Verified Execution

3 pages · dense two-column · related work across VLA, hierarchical policies, flow matching, and safety filters · full method (SMC, CBF–QP) · ablation design table with claimed component-to-metric effects.

ieee_paper.pdf → · source →

Reproduce

      xelatex nmi_paper.tex      # Nature Machine Intelligence format → nmi_paper.pdf

      pdflatex ieee_paper.tex    # IEEE conference format → ieee_paper.pdf

      python make_fig.py         # fig1_architecture.png (monochrome, no color)