Transferability for General Reasoning: An Automated Curriculum for Multi-Domain RLVR

Short summary

Transfer-Aware Curriculum (TAC) optimizes multi-domain reasoning model training by prioritizing domains whose updates benefit the broader training suite. Using gradient geometry to estimate cross-domain transferability, TAC achieved 2.8-point improvements on reasoning benchmarks across Qwen and Llama models with negligible computational overhead.

• •Novel curriculum algorithm uses gradient geometry to measure cross-domain transfer benefits
• •Achieved 2.8-point accuracy gains (10% relative) vs. learnability-only baselines
• •Tested across six domains (math, programming, science) on Qwen and Llama models

Generated with AI, which can make mistakes.