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

Source: [Transferability for General Reasoning: An Automated Curriculum for Multi-Domain RLVR](<https://www.sota2.com/research/paper/6a3cb63b206542251d7ca69b>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-07-06T06:00:00.878Z

Authors: Yongjin Yang, Jiarui Liu, Yinghui He, Lechen Zhang, Bernhard Sch\\"olkopf, Zhijing Jin

Published: 2026-06-23T21:10:29.000Z

## Abstract

Reinforcement learning with verifiable rewards (RLVR) has been extended from single-domain training to multi-domain reasoning suites spanning mathematics, programming, and science. However, the training curriculum (how often each domain is sampled) is typically fixed or hand-tuned, even though reasoning skills transfer unevenly across domains. Existing learnability-based curricula adapt to where the policy is currently improving, but are blind to whether a gradient step on the selected domain benefits the remaining domains. In this paper, we propose Transfer-Aware Curriculum (TAC), a bandit-style online curriculum that prioritizes domains whose updates broadly benefit the rest of the training suite. TAC repurposes signals already produced by RL training: per-domain advantages capture local learnability, and projected gradients, taken from the GRPO step being computed, estimate cross-domain transferability via gradient-geometry alignment, at negligible cost (\<1% wall-clock overhead). Across a six-domain reasoning suite, TAC achieves the best macro-averaged accuracy on both Qwen3-1.7B and Llama3.2-3B, outperforming proportional random sampling, a hand-designed schedule, and a learnability-only bandit, and improving over the last of these by up to 2.8 points (10% relative). Ablations show performance degrades sharply when the transferability term is removed, and TAC remains robust on imbalanced training mixtures where learnability-only curricula over-commit to dominant domains. Our findings establish cross-domain transferability as a key signal for curriculum design in multi-domain RLVR.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Code Simulation on CruxEval-I](<https://www.sota2.com/research/sota/code-simulation-on-cruxeval-i.md>): Code Simulation; CRUXEval-I; Pass@1: 61.4.

- [STEM Reasoning on SuperGPQA (Pass@1)](<https://www.sota2.com/research/sota/stem-reasoning-on-supergpqa-pass-at-1.md>): STEM Reasoning; SuperGPQA; Pass@1 Accuracy: 32.

- [Table Reasoning on FinQA (Pass@1)](<https://www.sota2.com/research/sota/table-reasoning-on-finqa-pass-at-1.md>): Table Reasoning; FinQA; Pass@1: 39.7.

- [Table Reasoning on HiTab (Pass@1)](<https://www.sota2.com/research/sota/table-reasoning-on-hitab-pass-at-1.md>): Table Reasoning; HiTab; Pass@1: 70.1.

- [STEM Knowledge on GPQA](<https://www.sota2.com/research/sota/stem-knowledge-on-gpqa.md>): STEM Knowledge; GPQA; Pass@1 Accuracy: 31.9.

- [Tabular Reasoning on MultiHierTT](<https://www.sota2.com/research/sota/tabular-reasoning-on-multihiertt.md>): Tabular Reasoning; MultiHiertt; Pass@1 Accuracy: 28.7.

- [General Reasoning on General Reasoning Aggregate](<https://www.sota2.com/research/sota/general-reasoning-on-general-reasoning-aggregate.md>): General Reasoning; General Reasoning Aggregate; Accuracy (Macro): 45.4.

- [Logical Reasoning on Zebra Puzzle](<https://www.sota2.com/research/sota/logical-reasoning-on-zebra-puzzle.md>): Logical reasoning; Zebra Puzzle; Pass@1: 36.9.

- [Mathematical Reasoning on CodeI/O](<https://www.sota2.com/research/sota/mathematical-reasoning-on-codei-o.md>): Mathematical Reasoning; CodeI/O; Pass@1: 5.6.

- [Table Reasoning on MH](<https://www.sota2.com/research/sota/table-reasoning-on-mh.md>): Table Reasoning; MH; Pass@1: 47.7.

- [Code Simulation on CruxEval-O](<https://www.sota2.com/research/sota/code-simulation-on-cruxeval-o.md>): Code Simulation; CRUXEval-O; Pass@1: 62.1.

- [Logical Reasoning on Zebra](<https://www.sota2.com/research/sota/logical-reasoning-on-zebra.md>): Logical reasoning; Zebra; Pass@1 Accuracy: 35.8.

- [Logical Reasoning on ARC-AGI](<https://www.sota2.com/research/sota/logical-reasoning-on-arc-agi.md>): Logical reasoning; ARC-AGI; Pass@1: 2.6.

- [Code Simulation on CodeI/O](<https://www.sota2.com/research/sota/code-simulation-on-codei-o.md>): Code Simulation; CodeI/O; Pass@1 Accuracy: 4.8.

- [Mathematical Reasoning on AIME (pass@1)](<https://www.sota2.com/research/sota/mathematical-reasoning-on-aime-pass-at-1.md>): Mathematical Reasoning; AIME; Pass@1: 12.3.

- [STEM Reasoning on GPQA (Pass@1)](<https://www.sota2.com/research/sota/stem-reasoning-on-gpqa-pass-at-1.md>): STEM Reasoning; GPQA; Pass@1 Accuracy: 42.5.

- [Code Generation on HumanEval (pass@1, Final Gap)](<https://www.sota2.com/research/sota/code-generation-on-humaneval-pass-at-1-final-gap.md>): Code Generation; HumanEval; pass@1: 66.7.

- [Code Generation on HumanEval (test)](<https://www.sota2.com/research/sota/code-generation-on-humaneval-test.md>): Code Generation; HumanEval (test); Pass@1: 83.5.

- [Code Generation on MBPP (test)](<https://www.sota2.com/research/sota/code-generation-on-mbpp-test.md>): Code Generation; MBPP (test); Pass@1: 67.4.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2606.25178>)

[PDF](<https://arxiv.org/pdf/2606.25178.pdf>)

[Implementation](<https://github.com/YangYongJin/transfer-aware-curriculum>)
