CoRT: Counterfactual Replay for Token-Level Rubric-Guided Policy Optimization
Open paper
Benchmark
Model
Rank
Results
instruction-following-on-ifbench
CoRT (Qwen3-14B)
–
Accuracy: 41.09
instruction-following-on-ifeval
Qwen3-14B + CoRT (AON)
–
Inst-level strict-accuracy: 93.48
Rank counts only results with a code link.