Research map / Vision-language and multimodal / Multimodal LLMs
MARS: Multimodal Adaptive Reasoning Model for Avoiding Overthinking
Tan Yue, Qiong Wu, Dongyan Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MARS, a multimodal LLM reasoning model that adaptively adjusts its reasoning strategy to question difficulty to avoid overthinking and reduce computational cost.
Problem
Multimodal large language models often produce excessively long reasoning chains, causing high computational cost and inefficiency on vision-language tasks.
Approach
Enables adaptive adjustment of the reasoning strategy based on question difficulty so that easy questions receive shorter reasoning while harder ones get fuller reasoning; the specific mechanism is not stated in the abstract provided.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CoT Masking Learning (stage 1) (ablation) | MARS (Ins.+RL) · SciChart (out-of-domain) | R-Acc: 38.23 -> 41.68 | Improved |
| add IB-GRPO in place of raw GRPO (ablation) | MARS three-stage · SciChart (out-of-domain) | average tokens: 256.68 -> 95.81 | Improved |
| add IB-GRPO in place of raw GRPO (ablation) | MARS three-stage · SciChart (out-of-domain) | R-Acc: 41.68 -> 42.31 | Improved |
| add RL stage (with IB-GRPO) after Mask+Ins. (ablation) | MARS · SciChart (out-of-domain) | average tokens: 218.65 -> 95.81 | Improved |
| add RL stage (with IB-GRPO) after Mask+Ins. (ablation) | MARS · SciChart (out-of-domain) | R-Acc: 40.68 -> 42.31 | Improved |
| add Adaptive Reasoning Instruction Learning (ablation) | MARS (Mask+RL) · SciChart (out-of-domain) | average tokens: 686.17 -> 95.81 | Improved |
| add Adaptive Reasoning Instruction Learning (ablation) | MARS (Mask+RL) · SciChart (out-of-domain) | R-Acc: 41.15 -> 42.31 | Improved |
| apply MARS training to Mimo-VL-7B-RL | Mimo-VL-7B-RL · SciChart (out-of-domain) | average tokens: 980.09 -> 95.81 | Improved |
| apply MARS training to Mimo-VL-7B-RL | Mimo-VL-7B-RL · SciChart (out-of-domain) | R-Acc: 41.77 -> 42.31 | Improved |
| apply MARS on FWHM questions | Mimo-VL-7B-RL · SciChart (out-of-domain) | R-Acc: 16.95 -> 15.39 | Hurt |
| apply MARS on Peak Position questions | Mimo-VL-7B-RL · SciChart (out-of-domain) | R-Acc: 48.99 -> 48.56 | No clear effect |
| thinking mode vs non-thinking | Kimi-VL-A3B · SciChart (out-of-domain) | R-Acc: 25.10 -> 27.78 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CogVLM 17B — CogVLM: Visual Expert for Pretrained Language Models (NeurIPS 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping (ICLR 2026 · similarity 0.90)
- SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models (ICLR 2026 · similarity 0.89)
- Learn to Think: Improving Multimodal Reasoning through Vision-Aware Self-Improvement Training (ICML 2026 · similarity 0.88)
- Revisiting the Data Sampling in Multimodal Post-training from a Difficulty-Distinguish View (AAAI 2026 · similarity 0.88)
- R1-ShareVL: Incentivizing Reasoning Capabilities of Multimodal Large Language Models via Share-GRPO (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
