# Imbalanced Gradients in RL Post-Training of Multi-Task LLMs

arXiv: https://arxiv.org/abs/2510.19178 · HTML: https://mlsanity.com/paper/2510.19178/

## Key facts

- Cited by 20 papers in the map (data through 14 Oct 2026).
- Builds on 37 of its 52 references that are in the map.
- First posted on arXiv 22 Oct 2025.

## Builds on (most cited first, 30 of 37)

- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Measuring Mathematical Problem Solving With the MATH Dataset](https://mlsanity.com/paper/2103.03874/) (2021)
- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [DAPO: An Open-Source LLM Reinforcement Learning System at Scale](https://mlsanity.com/paper/2503.14476/) (2025)
- [HybridFlow: A Flexible and Efficient RLHF Framework](https://mlsanity.com/paper/2409.19256/) (2024)
- [Gradient Surgery for Multi-Task Learning](https://mlsanity.com/paper/2001.06782/) (2020)
- [GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks](https://mlsanity.com/paper/1711.02257/) (2017)
- [Multi-Task Learning as Multi-Objective Optimization](https://mlsanity.com/paper/1810.04650/) (2018)
- [Introduction to Online Convex Optimization](https://mlsanity.com/paper/1909.05207/) (2019)
- [MiniCPM: Unveiling the Potential of Small Language Models with Scalable Training Strategies](https://mlsanity.com/paper/2404.06395/) (2024)
- [Deep Batch Active Learning by Diverse, Uncertain Gradient Lower Bounds](https://mlsanity.com/paper/1906.03671/) (2019)
- [FinQA: A Dataset of Numerical Reasoning over Financial Data](https://mlsanity.com/paper/2109.00122/) (2021)
- [Conflict-Averse Gradient Descent for Multi-task Learning](https://mlsanity.com/paper/2110.14048/) (2021)
- [How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources](https://mlsanity.com/paper/2306.04751/) (2023)
- [Just Pick a Sign: Optimizing Deep Multitask Models with Gradient Sign Dropout](https://mlsanity.com/paper/2010.06808/) (2020)
- [Rewards-in-Context: Multi-objective Alignment of Foundation Models with Dynamic Preference Adjustment](https://mlsanity.com/paper/2402.10207/) (2024)
- [Self-Evolving Curriculum for LLM Reasoning](https://mlsanity.com/paper/2505.14970/) (2025)
- [Efficient Reinforcement Finetuning via Adaptive Curriculum Learning](https://mlsanity.com/paper/2504.05520/) (2025)
- [Curriculum Reinforcement Learning from Easy to Hard Tasks Improves LLM Reasoning](https://mlsanity.com/paper/2506.06632/) (2025)
- [FAMO: Fast Adaptive Multitask Optimization](https://mlsanity.com/paper/2306.03792/) (2023)
- [REBEL: Reinforcement Learning via Regressing Relative Rewards](https://mlsanity.com/paper/2404.16767/) (2024)
- [DUMP: Automated Distribution-Level Curriculum Learning for RL-based LLM Post-training](https://mlsanity.com/paper/2504.09710/) (2025)
- [Prompt Curriculum Learning for Efficient LLM Post-Training](https://mlsanity.com/paper/2510.01135/) (2025)
- [Conditional Language Policy: A General Framework for Steerable Multi-Objective Finetuning](https://mlsanity.com/paper/2407.15762/) (2024)
- [Mixture-of-LoRAs: An Efficient Multitask Tuning for Large Language Models](https://mlsanity.com/paper/2403.03432/) (2024)
- [MoDoMoDo: Multi-Domain Data Mixtures for Multimodal LLM Reinforcement Learning](https://mlsanity.com/paper/2505.24871/) (2025)
- [SPEED-RL: Faster Training of Reasoning Models via Online Curriculum Learning](https://mlsanity.com/paper/2506.09016/) (2025)
- [Dynamic Data Mixing Maximizes Instruction Tuning for Mixture-of-Experts](https://mlsanity.com/paper/2406.11256/) (2024)

## Built on by (most cited first, 20 of 20)

- [Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillation](https://mlsanity.com/paper/2608.19098/) (2026)
- [Multi-Task GRPO: Reliable LLM Reasoning Across Tasks](https://mlsanity.com/paper/2602.05547/) (2026)
- [To Mix or To Merge: Toward Multi-Domain Reinforcement Learning for Large Language Models](https://mlsanity.com/paper/2602.12566/) (2026)
- [Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention](https://mlsanity.com/paper/2605.29548/) (2026)
- [Advancing General-Purpose Reasoning Models with Modular Gradient Surgery](https://mlsanity.com/paper/2602.02301/) (2026)
- [Your Language Model is Its Own Critic: Reinforcement Learning with Value Estimation from Actor's Internal States](https://mlsanity.com/paper/2605.07579/) (2026)
- [Consolidating RLVR Capabilities Across Domains: A Deep Dive into Fusion Paradigms](https://mlsanity.com/paper/2608.27409/) (2026)
- [Training Data Selection with Gradient Orthogonality for Efficient Domain Adaptation](https://mlsanity.com/paper/2602.06359/) (2026)
- [Harmony in Diversity: Multi-domain Contrastive Policy Optimization for Large Reasoning Models](https://mlsanity.com/paper/2605.25443/) (2026)
- [REGEN: Replay-recycling for Expert-to-Generalist distillation with Offline Reinforcement Learning](https://mlsanity.com/paper/2607.19450/) (2026)
- [SFT Conflicts, RL Coexists: A Theoretical and Empirical Analysis of Multi-Task Learning for LLMs](https://mlsanity.com/paper/2608.03573/) (2026)
- [GIFT: Games as Informal Training for Generalizable LLMs](https://mlsanity.com/paper/2601.05633/) (2026)
- [Building Multi-Task Agentic LLMs via Two-Phase Distillation](https://mlsanity.com/paper/2606.30044/) (2026)
- [One Step, One Lead: Mitigating Higher-Order Interference in Multi-Domain Reinforcement Learning via Cross-Step Control](https://mlsanity.com/paper/2609.06469/) (2026)
- [A Better Spur Should Start From Each Objective](https://mlsanity.com/paper/2609.08211/) (2026)
- [AGGC: Adaptive Group Gradient Clipping for Stabilizing Large Language Model Training](https://mlsanity.com/paper/2601.11864/) (2026)
- [Schedule-and-Calibrate: Utility-Guided Multi-Task Reinforcement Learning for Code LLMs](https://mlsanity.com/paper/2605.06111/) (2026)
- [Manifold Bandits: Bayesian Curriculum Learning over the Latent Geometry of Large Language Models](https://mlsanity.com/paper/2606.19750/) (2026)
- [World Feedback for Clinical Agents: Diagnosing RL in FHIR Environments](https://mlsanity.com/paper/2607.01470/) (2026)
- [Mitigating Exploration Bias in RL for Multi-Instruction Following](https://mlsanity.com/paper/2608.23830/) (2026)

## Abstract (arXiv)

Multi-task post-training of large language models (LLMs) is typically performed by mixing datasets from different tasks and optimizing them jointly. This approach implicitly assumes that all tasks contribute gradients of similar magnitudes; when this assumption fails, optimization becomes biased toward large-gradient tasks. In this paper, however, we show that this assumption fails in RL post-training: certain tasks produce significantly larger gradients, thus biasing updates toward those tasks. Such gradient imbalance would be justified only if larger gradients implied larger learning gains on the tasks (i.e., larger performance improvements) -- but we find this is not true. Large-gradient tasks can achieve similar or even much lower learning gains than small-gradient ones. Further analyses reveal that these gradient imbalances cannot be explained by typical training statistics such as training rewards or advantages, suggesting that they arise from the inherent differences between tasks. This cautions against naive dataset mixing and calls for future work on principled gradient-level corrections for LLMs.