TRACER: Trace-Based Adaptive Cost-Efficient Routing for LLM Classification
Paper • 2604.14531 • Published • 5 •
2
Adam Rida
adamrida
AI & ML interests
AI Research - Interpretability and Model Drift
Recent Activity
commentedon a paper about 3 hours ago
TRACER: Trace-Based Adaptive Cost-Efficient Routing for LLM Classification authored a paper about 12 hours ago
TRACER: Trace-Based Adaptive Cost-Efficient Routing for LLM Classification upvoted a paper about 14 hours ago
TRACER: Trace-Based Adaptive Cost-Efficient Routing for LLM ClassificationOrganizations
commented a
paper about 3 hours ago
authored a
paper about 12 hours ago
upvoted a paper about 14 hours ago
posted an update about 18 hours ago
Post
67
TRACER is an open-source routing system that learns from an LLM’s own production traces to replace a large share of future classification calls with lightweight ML surrogates. It deploys only when a parity gate shows the surrogate matches the teacher above a user-defined quality threshold, and it generates artifacts that make the routing boundary inspectable. In experiments, TRACER achieved 83–100% surrogate coverage on a 77-class intent benchmark, fully replaced the teacher on a 150-class benchmark, and correctly refused deployment on NLI when the representation was not reliable enough.
https://arxiv.org/abs/2604.14531
https://github.com/adrida/tracer
https://arxiv.org/abs/2604.14531
https://github.com/adrida/tracer
posted an update 13 days ago
Post
151
Open-sourced TRACER.
Many LLM classification calls in production are overkill.
For tasks like intent detection, moderation, tagging, or routing, TRACER learns which requests can be safely offloaded to a lightweight ML model trained on the LLM’s own outputs.
You keep the hard cases on the LLM, set a target quality bar, and offload the easy traffic.
On the right workloads, this can remove 90%+ of LLM calls.
GitHub:
https://github.com/adrida/tracer
Many LLM classification calls in production are overkill.
For tasks like intent detection, moderation, tagging, or routing, TRACER learns which requests can be safely offloaded to a lightweight ML model trained on the LLM’s own outputs.
You keep the hard cases on the LLM, set a target quality bar, and offload the easy traffic.
On the right workloads, this can remove 90%+ of LLM calls.
GitHub:
https://github.com/adrida/tracer
updated a
dataset 19 days ago
published a
dataset 19 days ago
updated 3
Spaces over 1 year ago
updated 7
Spaces almost 2 years ago
updated a
Space about 2 years ago
updated a
Space over 2 years ago