TraceSpec preprint reports faster decoding for AI agent workloads

Researchers presenting TraceSpec report a 12.5% to 22.6% improvement in replay throughput over its DSpark base across three target model sizes and two temperatures. The preprint studies speculative decoding, which uses draft predictions to accelerate language model output, during browsing and software engineering tasks. TraceSpec trains on intermediate assistant turns and allocates a fixed prediction budget using estimates conditioned on the agent's current state. Its evaluation corpus contains 10,000 training states and 2,500 test states. The analysis finds that aggregate acceptance metrics can conceal declining performance within matched late browsing actions. The reported throughput gains come from replay evaluation, rather than a demonstrated production deployment.