LatentSkill: In-Context Textual Skills → In-Weight Latent Skills

TL;DR. Agent systems encode reusable procedures as textual skills injected into the prompt every step. That costs context tokens at every turn and exposes the skill as plaintext. LatentSkill converts a textual skill into a plug-and-play LoRA adapter via a pretrained hypernetwork, storing the skill in weight space rather than context space. It keeps the modularity of text skills (load, scale, compose) while removing the per-step token tax. On ALFWorld it improves success by 21.4 / 13.4 points (seen / unseen) with 64.1% fewer prefill tokens; on Search-QA it adds 3.0 EM with 72.2% lower skill-token overhead. Generated skill LoRAs form a structured semantic geometry, are controllable via the LoRA scaling coefficient, and compose through parameter-space arithmetic.

flowchart LR
  TXT[Textual skill<br/>procedure as prompt] --> HYP[Pretrained hypernetwork]
  HYP --> LORA[Skill LoRA adapter<br/>weight space]
  LORA --> AGENT[Agent: no per-step<br/>skill tokens]
  AGENT --> OUT[ALFWorld +21.4/+13.4 pts<br/>-64.1% prefill tokens]
  LORA --> COMP[Compose via<br/>parameter arithmetic;<br/>scale via LoRA coeff]
  CTX[Old way: skill in prompt<br/>every step] -.->|token tax + plaintext exposure| AGENT
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
  class TXT input
  class HYP decision
  class LORA,AGENT,OUT,COMP output
  class CTX warn

Key points

Weight-space skills remove the per-step token tax that in-context skill libraries pay, and stop exposing skill content as plaintext.
Hypernetwork generates the LoRA from the textual skill, so authoring stays text-level; deployment is weight-level.
Composable and controllable: skill LoRAs live in a structured geometry, scale with the LoRA coefficient, and combine by parameter arithmetic when components align.

How it relates to prior wiki knowledge

Latest entry in the parametric-internalization line. Code2LoRA (06-06, hypernetwork generates repo-specific adapters) and Video2LoRA (06-06) used hypernetworks to move context into weights; LatentSkill does the same for agent skills. Same engine (hypernetwork → LoRA), new payload. See parametric-context-internalization.
Efficiency answer to the skill-memory cluster. The self-evolving-agents line (Socratic-SWE, OpenSkill, 06-08) and today's Skill-RM / Experience-Makes-Skillful papers all grow skill libraries; LatentSkill makes storing and loading those skills cheap, the missing efficiency layer under the skill-memory wave.

Gaps

Tested on ALFWorld and Search-QA; whether hypernetwork-generated LoRAs preserve skill fidelity for long, compositional procedures is untested.
Parameter-space composition works "when skill components are aligned" — the alignment condition is not characterized, so arbitrary skill mixing may not compose.

Source: HuggingFace Daily Papers · Paper · raw: raw/huggingface/2026-06-09-latentskill-from-in-context-textual-skills-to-in-weight-late.md