# Balanced Thinking: Improving Chain of Thought Training in Vision Language Models

arXiv: https://arxiv.org/abs/2603.18656 · HTML: https://mlsanity.com/paper/2603.18656/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Multimodal reasoning, one of 17 directions in LLM reasoning.
- Builds on 22 of its 34 references that are in the map.
- First posted on arXiv 19 Mar 2026.

## Placed in

- LLM reasoning / [Multimodal reasoning](https://mlsanity.com/lr/multimodal-reasoning/)
- Multimodal LLMs / vision-language models / [Embodied and 3D vision-language models](https://mlsanity.com/mllm/embodied-and-3d-vision-language-models/)

## Builds on (most cited first, 22 of 22)

- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://mlsanity.com/paper/2501.12948/) (2025)
- [Finetuned Language Models Are Zero-Shot Learners](https://mlsanity.com/paper/2109.01652/) (2021)
- [Gemma 3 Technical Report](https://mlsanity.com/paper/2503.19786/) (2025)
- [Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models](https://mlsanity.com/paper/2503.06749/) (2025)
- [VLM-R1: A Stable and Generalizable R1-style Large Vision-Language Model](https://mlsanity.com/paper/2504.07615/) (2025)
- [Video-R1: Reinforcing Video Reasoning in MLLMs](https://mlsanity.com/paper/2503.21776/) (2025)
- [Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models](https://mlsanity.com/paper/2503.16419/) (2025)
- [The Lessons of Developing Process Reward Models in Mathematical Reasoning](https://mlsanity.com/paper/2501.07301/) (2025)
- [Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs](https://mlsanity.com/paper/2506.14245/) (2025)
- [GUI-R1 : A Generalist R1-Style Vision-Language Action Model For GUI Agents](https://mlsanity.com/paper/2504.10458/) (2025)
- [When More is Less: Understanding Chain-of-Thought Length in LLMs](https://mlsanity.com/paper/2502.07266/) (2025)
- [Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree Search](https://mlsanity.com/paper/2412.18319/) (2024)
- [Crossing the Reward Bridge: Expanding RL with Verifiable Rewards Across Diverse Domains](https://mlsanity.com/paper/2503.23829/) (2025)
- [Med-R1: Reinforcement Learning for Generalizable Medical Reasoning in Vision-Language Models](https://mlsanity.com/paper/2503.13939/) (2025)
- [Multi-Step Reasoning with Large Language Models, a Survey](https://mlsanity.com/paper/2407.11511/) (2024)
- [GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents](https://mlsanity.com/paper/2505.15810/) (2025)
- [TinyLLaVA-Video-R1: Towards Smaller LMMs for Video Reasoning](https://mlsanity.com/paper/2504.09641/) (2025)
- [Post-hoc Reward Calibration: A Case Study on Length Bias](https://mlsanity.com/paper/2409.17407/) (2024)
- [Stable Reinforcement Learning for Efficient Reasoning](https://mlsanity.com/paper/2505.18086/) (2025)
- [Mind with Eyes: from Language Reasoning to Multimodal Reasoning](https://mlsanity.com/paper/2503.18071/) (2025)
- [Bias Fitting to Mitigate Length Bias of Reward Model in RLHF](https://mlsanity.com/paper/2505.12843/) (2025)

## Built on by (most cited first, 1 of 1)

- [Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation Enforcement](https://mlsanity.com/paper/2609.24243/) (2026)

## Abstract (arXiv)

Multimodal reasoning in vision-language models (VLMs) typically relies on a two-stage process: supervised fine-tuning (SFT) and reinforcement learning (RL). In standard SFT, all tokens contribute equally to the loss, even though reasoning data are inherently token-imbalanced. Long <think> traces overshadow short but task-critical <answer> segments, leading to verbose reasoning and inaccurate answers. We propose SCALe (Scheduled Curriculum Adaptive Loss), which explicitly separates supervision over reasoning and answer segments using dynamic, length-independent weighting. Unlike vanilla SFT, which overweights the <think> segment, SCALe-SFT gradually shifts the focus from <think> to <answer> throughout training via a cosine scheduling policy, encouraging concise and well-grounded reasoning. We evaluate SCALe across diverse benchmarks and architectures. Results show that SCALe consistently improves accuracy over vanilla SFT and matches the performance of the full two-phase SFT + GRPO pipeline while requiring only about one-seventh of the training time, making it a lightweight yet effective alternative. When combined with GRPO, SCALe achieves the best overall performance, highlighting its value both as a standalone method and as a strong foundation for reinforcement refinement.