DUET optimises agents and LLM graders together
The non-obvious takeaway is that agent failure modes shift as the agent improves, so a static grader becomes a bottleneck that joint co-optimisation removes. Developers building agent evaluation pipelines can apply the alternating solver-grader update loop to their own scoring setups. Against named alternatives, DUET's 96.5% tops GEPA's 94.4% and SkillOpt's 94.0% on the Finance Agent Benchmark.
A new paper introduces DUET, a method that alternately optimises an agent (solver) and its LLM grader instead of keeping the grader fixed. In each round, DUET selects training tasks, runs the solver, has the grader score and critique trajectories, then a tool-using updater revises either the grader or the solver while the other stays frozen. On the Finance Agent Benchmark, the solver reaches 96.5%, above GEPA (94.4%) and SkillOpt (94.0%), including when those baselines optimise against the benchmark's own reference evaluator. Grader agreement on GDPval rises from 68.4% to 78.5%, whereas training the grader alone moves agreement only from 67.9% to 68.8%.
Nice paper on optimizing agents and their graders together.
The key finding is that you want to update your LLM grader alongside the agent instead of keeping it fixed, because the agent's failure modes change as it improves.
DUET alternates between the two.
Each round picks training tasks, runs the solver, has the grader score and critique the trajectories, then a tool-using updater revises either the grader or the solver while the other stays fixed.
On the Finance Agent Benchmark, the solver reaches 96.5%, above GEPA (94.4%) and SkillOpt (94.0%). It also beats both when they optimize against the benchmark's own reference evaluator.
Grader agreement rises from 68.4% to 78.5% on GDPval. Training the grader with the solver frozen moves agreement only from 67.9% to 68.8%.
Paper: https://t.co/qmZBl4RHu5
