Back to AI Research

AI Research

SkillZip: Evaluation-Free Skill Compression for Sel... | AI Research

Key Takeaways

  • SkillZip is a method designed to compress the accumulated "skills" of self-evolving AI agents without relying on external task evaluations or rollouts.
  • Self-evolving agents accumulate reusable skills by appending successful procedures and failure fixes.
  • Over time, the same requirement is often restated in several branches, examples, and warnings, while common action sequences are copied rather than reused.
  • The resulting skill becomes expensive to inject and difficult to maintain.
  • Evaluation-guided compression can test these behaviors, but it introduces rollouts, cost, and dependence on the compression-time evaluation set.
Paper AbstractExpand

Self-evolving agents accumulate reusable skills by appending successful procedures and failure fixes. Over time, the same requirement is often restated in several branches, examples, and warnings, while common action sequences are copied rather than reused. The resulting skill becomes expensive to inject and difficult to maintain. Generic prompt compression is ill-suited to this setting because a skill is not a flat passage: its name and description define when it applies, its workflow controls execution, its tool and output contracts constrain validity, and rare exceptions may remain essential even when no sampled task activates them. Evaluation-guided compression can test these behaviors, but it introduces rollouts, cost, and dependence on the compression-time evaluation set. We present SkillZip, an evaluation-free method that compresses a skill by finding its shortest faithful structural explanation. The intuition is explain once, reference many: state a repeated rule once at the scope where it applies, factor a repeated action sequence into a shared procedure, and keep only the differences as explicit exceptions. We formalize this intuition as a typed minimum description-length objective over a skill contract and a residual, subject to a hard coverage constraint for every extracted trigger, workflow edge, tool requirement, obligation, and output field. The formulation provides simple sharing thresholds, preserves unique rare rules by construction, and supports efficient local updates. SkillZip has a one-shot mode with one structured extraction call and deterministic optimization, and a continual Zip-on-Write mode that integrates each self-evolution patch without replaying tasks or reparsing the full history. Through comprehensive experimental evaluations, we demonstrate the effectiveness and superiority of SkillZip in compression performance, generalizability, and cost overhead.

SkillZip is a method designed to compress the accumulated "skills" of self-evolving AI agents without relying on external task evaluations or rollouts. As agents evolve, they often accumulate redundant instructions, such as repeated warnings, overlapping workflow branches, and copied action sequences. SkillZip addresses this by identifying and consolidating these repeated structures into a compact, human-readable contract, ensuring that the agent’s instructions remain efficient and maintainable as they grow.

The Problem with Evolving Skills

Self-evolving agents typically improve by appending new procedures and fixes to their existing skill sets. Over time, this results in an "append-only notebook" where the same requirements are restated across different branches or examples. This redundancy increases the cost of injecting the skill into the agent's context window and makes the instructions difficult to maintain. Existing prompt compression methods are often ill-suited for this because they treat skills as flat text, whereas a skill is actually a complex, typed structure involving specific tool contracts, workflow logic, and rare but essential exceptions that must be preserved regardless of how often they are triggered.

How SkillZip Works

SkillZip operates on the principle of "explain once, reference many." It treats a skill as a structured contract consisting of an interface, workflow, tool protocols, scoped rules, and output requirements. The method follows a minimum description length (MDL) objective, which seeks the shortest representation that still covers every normative requirement extracted from the original skill.
The process involves two primary modes:

  • One-shot mode: A deterministic scanner parses the skill into a typed contract, which is then optimized to remove redundancies while keeping unique, rare rules intact.

  • Zip-on-Write mode: This mode integrates new self-evolution patches into the existing compact contract incrementally. It avoids the need to re-parse the entire history or replay past tasks, performing local updates to the contract and only triggering a full repacking when enough reusable structure accumulates.

Preservation and Reliability

A core feature of SkillZip is its "hard coverage constraint." Unlike methods that use task-based feedback to decide what to keep, SkillZip ensures that every extracted requirement—including rare exceptions—is preserved by construction. Because it does not rely on sampled tasks or behavioral verifiers, the compressed skill is not dependent on the specific distribution of data used during the compression process. This approach ensures that the agent retains all necessary procedural knowledge even if that knowledge is rarely activated.

Practical Considerations

SkillZip is designed to produce an ordinary, human-readable text artifact, making it easy to inspect, version, and use across different agent backbones. By separating the interpretation of the skill from the optimization process, the authors ensure that the compression is deterministic and traceable to the original source text. While the method is effective at consolidating knowledge, the authors note that the structural parser is the primary source of semantic uncertainty; consequently, any source spans that cannot be interpreted with high confidence are kept as a "locked residual" to prevent the accidental loss of critical information.

Comments (0)

No comments yet

Be the first to share your thoughts!