SpecAgent
A model collaboration protocol for tool-using agents that draws on the idea of speculative decoding: a weaker model drafts a complete tool-call trajectory, while a stronger model performs semantic-level verification in a single forward pass and repairs the first rejected step, thereby reducing inference latency and memory traffic.
Timeline (last 90 days)
SpecAgent proposes a collaborative protocol that shifts the strong model from the role of author of each step's decision to the role of reviewer of drafts, with the weak model drafting the trajectory and the strong model performing a single forward pass to verify and repair the first rejected step.
SpecAgent draws on the idea of speculative decoding but makes three modifications: the drafting granularity is raised from tokens to complete tool-call actions; verification is raised from the string level to semantic-level judgment of trajectory legality; and a repair stage is added to give an alternative action for the first rejected step.
On τ²-Bench, SpecAgent's success rate is within less than one percentage point of the strong model, while memory traffic and wall-clock time are roughly halved.
On AutomationBench and VitaBench, SpecAgent recovers 44% to 75% of the quality gap between the two models, while reducing latency and memory overhead.
In the edge-cloud deployment scenario on MCP-Atlas, where the test involves three cloud models, SpecAgent closes 53% to 75% of the quality gap between the weak and strong models, while the number of tokens generated by the strong model is at most half of the original.
The applicable boundary of SpecAgent is: it relies on the weak model being able to draft a trajectory that is basically semantically reasonable, otherwise the strong model needs to repair frequently and the benefit is discounted.
Results from multiple independent benchmarks point to the same consensus: a strong model's reviewing ability is more cost-effective than its generating ability, and having the strong model review a draft that has already been executed is far more efficient than having it generate step by step from scratch.
All Facts (7)
SpecAgent proposes a collaborative protocol that shifts the strong model from the role of author of each step's decision to the role of reviewer of drafts, with the weak model drafting the trajectory and the strong model performing a single forward pass to verify and repair the first rejected step.
50%UnverifiedSpecAgent draws on the idea of speculative decoding but makes three modifications: the drafting granularity is raised from tokens to complete tool-call actions; verification is raised from the string level to semantic-level judgment of trajectory legality; and a repair stage is added to give an alternative action for the first rejected step.
50%UnverifiedOn τ²-Bench, SpecAgent's success rate is within less than one percentage point of the strong model, while memory traffic and wall-clock time are roughly halved.
50%UnverifiedOn AutomationBench and VitaBench, SpecAgent recovers 44% to 75% of the quality gap between the two models, while reducing latency and memory overhead.
50%UnverifiedIn the edge-cloud deployment scenario on MCP-Atlas, where the test involves three cloud models, SpecAgent closes 53% to 75% of the quality gap between the weak and strong models, while the number of tokens generated by the strong model is at most half of the original.
50%UnverifiedThe applicable boundary of SpecAgent is: it relies on the weak model being able to draft a trajectory that is basically semantically reasonable, otherwise the strong model needs to repair frequently and the benefit is discounted.
50%UnverifiedResults from multiple independent benchmarks point to the same consensus: a strong model's reviewing ability is more cost-effective than its generating ability, and having the strong model review a draft that has already been executed is far more efficient than having it generate step by step from scratch.
50%