Unverified50% confidenceOpinionExact time
EASTAR's two-stage design of "LLM prior + supervised fine-tuning" is essentially transferring the adapter or prompt tuning idea from parameter-efficient transfer learning to the urban representation learning setting.
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedEduBehaviors splits annotation into two stages: in the first stage an LLM measures observable behaviors, and in the second stage classifiers are trained based on these observable behaviors to predict the target construct.72% similarUnverifiedThere are two possible paths to implement a conditioned task: add an encoding of the instruction to the observation space so the policy network learns the conditional distribution, or adopt a hierarchical architecture (a high level parses the instruction to generate sub-goals, and a low-level policy executes the pick-and-place).71% similarUnverifiedPutting a planning model inside the simulation loop is similar to an AI Director in games or a high-level policy in hierarchical reinforcement learning: the high level outputs events or subgoals at a lower frequency, and the low-level world model renders frames at a high frequency.70% similarUnverifiedWhat existing automated model optimization works have in common is that the large model outputs suggestions or configurations, and it is still the training pipeline that actually changes the model parameters.70% similarUnverifiedExisting embodied instruction-following planners fall mainly into two categories: supervised learning based on large-scale annotated trajectories, and single-pass serial reasoning that leverages LLM few-shot/zero-shot capabilities.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1164989API
curl https://kongchang.com/api/v1/knowledge/claims/1164989MCP
get_claim(id=1164989)