SchemaShield-Lite
An inference-time repair layer that addresses interface drift in LLM agents through a three-tier mechanism of schema validation, candidate-set retry and one-shot constrained repair, without retraining the base model.
Timeline (last 90 days)
SchemaShield-Lite is an inference-time repair layer that intercepts and repairs function-call errors through three mechanisms: schema validation, candidate-set retry, and one-shot constraint repair.
On the Qwen3.5-9B model, SchemaShield-Lite raised accuracy on the DICE-300 benchmark from 0.717 to 0.813, recovering 19 of the 22 drift-induced failures.
SchemaShield-Lite's repair layer introduced no new errors (no negative effects from repair were observed).
Ablation experiments show that a simple retry strategy cannot explain SchemaShield-Lite's gains; candidate-set retry must work in concert with schema validation and constraint repair.
SchemaShield-Lite requires no access to model internal states at deployment, no labeled data or additional training, and does not modify the base model's weights.
SchemaShield-Lite is positioned as a deployment-time robustness layer rather than a replacement for model capability, and can mitigate interface drift through lightweight middleware without interrupting service or retraining.
Future work may need to explore more complex drift scenarios, error propagation in multi-tool coordinated calls, and the compatibility of the repair layer with different base models.
All Facts (7)
SchemaShield-Lite is an inference-time repair layer that intercepts and repairs function-call errors through three mechanisms: schema validation, candidate-set retry, and one-shot constraint repair.
50%UnverifiedOn the Qwen3.5-9B model, SchemaShield-Lite raised accuracy on the DICE-300 benchmark from 0.717 to 0.813, recovering 19 of the 22 drift-induced failures.
50%UnverifiedSchemaShield-Lite's repair layer introduced no new errors (no negative effects from repair were observed).
50%UnverifiedAblation experiments show that a simple retry strategy cannot explain SchemaShield-Lite's gains; candidate-set retry must work in concert with schema validation and constraint repair.
50%UnverifiedSchemaShield-Lite requires no access to model internal states at deployment, no labeled data or additional training, and does not modify the base model's weights.
50%UnverifiedSchemaShield-Lite is positioned as a deployment-time robustness layer rather than a replacement for model capability, and can mitigate interface drift through lightweight middleware without interrupting service or retraining.
50%UnverifiedFuture work may need to explore more complex drift scenarios, error propagation in multi-tool coordinated calls, and the compatibility of the repair layer with different base models.
50%