LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents

Hugging Face Daily Papers Papers

Summary

LatentSkill converts textual skills into LoRA adapters stored in weight space, reducing context overhead while maintaining modularity and composability for LLM agents, achieving significant improvements on ALFWorld and Search-QA benchmarks.

Agent systems increasingly use textual skills to encode reusable task procedures, but injecting these skills into the prompt at every step incurs substantial context overhead and exposes skill content as plaintext. We present LatentSkill, a framework that converts textual skills into plug-and-play LoRA adapters through a pretrained hypernetwork. LatentSkill stores skill knowledge in weight space rather than context space, removing per-step skill tokens while preserving modular loading, scaling, and composition. On ALFWorld and Search-QA, LatentSkill outperforms the corresponding in-context skill baseline while using substantially fewer prefill tokens: it improves ALFWorld success by 21.4 and 13.4 points on the seen and unseen splits with 64.1% fewer prefill tokens, and improves Search-QA exact match by 3.0 points with 72.2% lower skill-token overhead. Further analysis shows that generated skill LoRAs form a structured semantic geometry, can be precisely controlled via the LoRA scaling coefficient, and can be composed through parameter-space arithmetic when skill components are aligned. These findings suggest that weight-space skills provide an efficient, modular, and less exposed substrate for extending LLM agents.
Original Article
View Cached Full Text

Cached at: 06/09/26, 08:40 AM

Paper page - LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents

Source: https://huggingface.co/papers/2606.06087 Authors:

,

,

,

,

,

,

,

,

,

Abstract

LatentSkill enables efficient deployment of textual skills in agent systems by converting them into LoRA adapters stored in weight space, reducing context overhead while maintaining modularity and composability.

Agent systems increasingly use textual skills to encode reusable task procedures, but injecting these skills into the prompt at every step incurs substantial context overhead and exposes skill content as plaintext. We present LatentSkill, a framework that converts textual skills into plug-and-playLoRA adaptersthrough a pretrainedhypernetwork. LatentSkill stores skill knowledge inweight spacerather thancontext space, removing per-step skill tokens while preserving modular loading, scaling, and composition. On ALFWorld and Search-QA, LatentSkill outperforms the corresponding in-context skill baseline while using substantially fewer prefill tokens: it improves ALFWorld success by 21.4 and 13.4 points on the seen and unseen splits with 64.1% fewer prefill tokens, and improves Search-QA exact match by 3.0 points with 72.2% lower skill-token overhead. Further analysis shows that generated skill LoRAs form a structuredsemantic geometry, can be precisely controlled via theLoRA scaling coefficient, and can be composed throughparameter-space arithmeticwhen skill components are aligned. These findings suggest that weight-space skills provide an efficient, modular, and less exposed substrate for extending LLM agents.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2606\.06087

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.06087 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.06087 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.06087 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

From History to State: Constant-Context Skill Learning for LLM Agents

arXiv cs.AI

This paper introduces 'constant-context skill learning,' a framework that moves procedural knowledge from prompts into model weights to reduce token usage and improve privacy for LLM agents. The method achieves strong performance on benchmarks like ALFWorld and WebShop while significantly reducing inference costs.