LLMs show transferable goal encodings: a Python plan replanted into C++ gets re-implemented
paul_cal · x · 2026-11-20
Arnab and Paul Calvano report signs of a 'goal encoding' in LLMs: the model plans chunks of code to be written later, abstractly enough that it can be injected into a stub in another language. They extracted this encoding from Python code, planted it into a C++ stub, and the model implemented the same algorithm in C++ with different libraries, syntax and variables — evidence of plasticity, where the LM realizes the goal through environment-appropriate pathways.
More from Research
- Persimmon user model to be presented at NeurIPS usersim workshop — ericzelikman · 2026-11-20
- Custom MLE-Agent hits 100% gold on MLE-Bench Tiny on a 2-vCPU EC2 box — InfamousTrouble7993 · 2026-11-20
- Matt Green muses: are n^25 galactic algorithms normal, or a glimpse of provable Algorithmica? — matthew_d_green · 2026-11-20
- MIT and NVIDIA's QDA uses agents to auto-design LLM quantization recipes, beating top public ones by up to 2.2 points — omarsar0 · 2026-11-20
- AI-powered forensic metascience: auditing 7 million papers, 500K files scanned, 100+ issues found — gleech · 2026-11-20
- Harmonic's Aristotle writes proven-correct fusion plasma controllers in C and Verilog in 4 hours — satnam6502 · 2026-11-20