Goal Encoding in LLMs
arXiv cs.AI · 13 Nov 2026, 04:55 AEST
Research investigates how large language models (LLMs) encode goals and guide actions. The study reveals that specific attention heads localize sub-goal encoding and can adapt across different environments.
Key facts
- We investigate whether and how LLMs encode goals: internal neural representations that precede and guide action, and that are flexible enough to be realized through many different execution paths.
- In this setting, we localize an encoding of the next sub-goal in the outputs of a sparse set of attention heads.
- We verify that this encoding is causal for next action and is plastic: transplanting it into a different environment, e.g., another programming language, triggers the same operation realized with a different sequence of tokens.
- We propose that it does not need to. Under a non-recurrent goal-seeking (NRGS) account, the LM frequently re-reads the goal from the task specification, compares it with the progress visible in the growing context, and pursues what remains.
Read aloud
Research investigates how large language models L L M s encode goals and guide actions. The study reveals that specific attention heads localize sub-goal encoding and can adapt across different environments. We investigate whether and how L L M s encode goals: internal neural representations that precede and guide action, and that are flexible enough to be realized through many different execution paths. In this setting, we localize an encoding of the next sub-goal in the outputs of a sparse set of attention heads. We verify that this encoding is causal for next action and is plastic: transplanting it into a different environment, e.g., another programming language, triggers the same operation realized with a different sequence of tokens. We propose that it does not need to. Under a non-recurrent goal-seeking NRGS account, the LM frequently re-reads the goal from the task specification, compares it with the progress visible in the growing context, and pursues what remains.
Sources in this cluster
- Goal Encoding in LLMs
arXiv cs.AI · 13 Nov 2026, 04:55 AEST