ShanghaiTech University

Feeding an LLM's parameter-adjustment directions to Bayesian optimization as a soft prior raises hypervolume on a 3D CPU-GPU chiplet design space from 0.8210 to 0.8356, and to 0.8051 on a 2.5D space without retraining.

Feeding an LLM's parameter-adjustment directions to Bayesian optimization as a soft prior raises hypervolume on a 3D CPU-GPU chiplet design space from 0.8210 to 0.8356, and to 0.8051 on a 2.5D space without retraining.

The work proposes a hierarchical framework that relabels successful rollouts of a single-agent physics-based human-object interaction teacher as short-horizon object-proxy displacement supervision, distilling a frozen reusable Object-oriented Motion Skill, while downstream training only learns a high-level policy that emits region-wise proxy-motion commands; in Isaac Gym with 4096 parallel environments the skill achieves better tracking than InterPhys and SkillMimic on both teacher-rollout and random proxy-motion targets, and supports two-human sofa carrying, four-human H-box and cross-box carrying, and two- and four-human sofa pushing.

The work proposes a hierarchical framework that relabels successful rollouts of a single-agent physics-based human-object interaction teacher as short-horizon object-proxy displacement supervision, distilling a frozen reusable Object-oriented Motion Skill, while downstream training only learns a high-level policy that emits region-wise proxy-motion commands; in Isaac Gym with 4096 parallel environments the skill achieves better tracking than InterPhys and SkillMimic on both teacher-rollout and random proxy-motion targets, and supports two-human sofa carrying, four-human H-box and cross-box carrying, and two- and four-human sofa pushing.

The authors introduce Active Taskless Distillation (ATD), which selects prompts where the teacher and student's shared public ancestor is nearly indifferent between two ordinary words, and lets a student initialized from that ancestor learn solely from the resulting prompt-word pairs, without target-task examples, teacher logits, or teacher parameters; in the primary coding experiment with Qwen2.5-1.5B, 5,664 samples yield a 5.34 pp gain on HumanEval+ over an exact nuisance-matched control, with transfer also shown in scientific knowledge, commonsense reasoning, and reading comprehension across additional model generations, sizes, and families, and functional analyses indicating the learned signal is composable and tracks the teacher's update strength.

The authors introduce Active Taskless Distillation (ATD), which selects prompts where the teacher and student's shared public ancestor is nearly indifferent between two ordinary words, and lets a student initialized from that ancestor learn solely from the resulting prompt-word pairs, without target-task examples, teacher logits, or teacher parameters; in the primary coding experiment with Qwen2.5-1.5B, 5,664 samples yield a 5.34 pp gain on HumanEval+ over an exact nuisance-matched control, with transfer also shown in scientific knowledge, commonsense reasoning, and reading comprehension across additional model generations, sizes, and families, and functional analyses indicating the learned signal is composable and tracks the teacher's update strength.