Source note

When Model Editing Meets Service Evolution: A Knowledge-Update Perspective for Service Recommendation

Service RecommendationModel EditingLLM AgentsConstrained DecodingCode Intelligence

EvoRec updates an LLM-based service recommender as APIs and services change, then constrains decoding so the model only outputs valid, non-duplicate services.

  • Service repositories change through new services, API updates, deprecated services, and version changes, so recommenders trained on old data can return stale or invalid service pipelines.
  • LLMs can map natural-language requirements or code context to service sequences, but they can hallucinate service names or calls that do not exist in the current catalog.
  • Full retraining or frequent fine-tuning is costly when service facts change often.
  • EvoRec uses ROME-style locate-then-edit model editing to insert or update service facts inside selected Transformer feed-forward layers.
  • Each edit maps a requirement pattern or code context key to an updated target service value, using cross-entropy for the target service and KL divergence to preserve unrelated behavior.
  • A retrieval-augmented prompt supplies relevant service examples from the service corpus before generation.
  • Trie-guided finite-automata constrained decoding masks invalid next tokens so generated names follow the service catalog and valid list structure.
  • The decoder tracks used service IDs and blocks only branches whose remaining subtree has no unused service, which prevents duplicates while allowing shared prefixes.
  • The paper claims an average relative improvement of 25.9% in Recall@5 over existing baselines on real-world service datasets.
  • Under evolving service scenarios, EvoRec outperforms model fine-tuning approaches by 22.3% according to the excerpt.
  • The excerpt states that EvoRec improves recommendation accuracy and adaptability while keeping maintenance cost low, but it does not provide dataset names, absolute Recall@5 values, model sizes, or a full baseline table.
  • The strongest concrete validity claim is that FA and Trie decoding enforce valid service names, valid separators/end tokens, and service-level deduplication during generation.