RATED
A reliability-aware tool retrieval framework for AI agents that combines a frozen dual-encoder for semantic matching with a lightweight contextual bandit for dynamic ranking, maintaining semantic retrieval quality while adapting to runtime tool failures.
Timeline (last 90 days)
The arXiv paper RATED (2026-10-09) layers a contextual bandit layer on top of semantic tool retrieval, enabling agents to adapt to real-time tool failures without retraining; in simulated interruption scenarios, the wrong-tool selection rate dropped from 86.2% to 2.9%; the accompanying EcoAgent benchmark contains 5172 tools and 210,000 queries.
RATED employs a frozen dual-encoder at the semantic matching layer, fine-tuned via contrastive learning.
RATED uses a lightweight contextual bandit module to re-rank candidate tools based on real-time execution feedback.
RATED's contextual bandit module can rapidly adapt to runtime tool failures without requiring model retraining.
RATED (Reliability-Aware Tool Retrieval via Execution Dynamics) is an AI agent tool retrieval framework that combines robust semantic matching with dynamic reliability adaptation.
In an API outage scenario with a 30% failure rate, RATED reduces the failure selection rate from 86.2% to 2.9%.
Compared to pure encoder inference, the additional latency introduced by RATED is less than 1 millisecond.
On real-world queries, RATED achieves a 52.2% relative MRR improvement over ToolLLM's API retriever.
All Facts (8)
The arXiv paper RATED (2026-10-09) layers a contextual bandit layer on top of semantic tool retrieval, enabling agents to adapt to real-time tool failures without retraining; in simulated interruption scenarios, the wrong-tool selection rate dropped from 86.2% to 2.9%; the accompanying EcoAgent benchmark contains 5172 tools and 210,000 queries.
50%UnverifiedRATED employs a frozen dual-encoder at the semantic matching layer, fine-tuned via contrastive learning.
50%UnverifiedRATED uses a lightweight contextual bandit module to re-rank candidate tools based on real-time execution feedback.
50%UnverifiedRATED's contextual bandit module can rapidly adapt to runtime tool failures without requiring model retraining.
50%UnverifiedRATED (Reliability-Aware Tool Retrieval via Execution Dynamics) is an AI agent tool retrieval framework that combines robust semantic matching with dynamic reliability adaptation.
50%UnverifiedIn an API outage scenario with a 30% failure rate, RATED reduces the failure selection rate from 86.2% to 2.9%.
50%UnverifiedCompared to pure encoder inference, the additional latency introduced by RATED is less than 1 millisecond.
50%UnverifiedOn real-world queries, RATED achieves a 52.2% relative MRR improvement over ToolLLM's API retriever.
50%