PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Adaptive Reasoning Paths for Efficient Visual Reasoning

Yixu Huang, Tinghui Zhu, Muhao Chen

arXiv:2604.14568v1cs.CVcs.CL

TL;DR

VRMs often generate unnecessarily long reasoning paths because visual questions do not always require every cognitive function. AVR decomposes visual reasoning into three functions and trains adaptive selection among three formats, achieving 50–90% token reduction while maintaining or improving accuracy.

  • Problem

    VRMs suffer from overthinking because many visual questions do not require the full set of visual perception, logical reasoning, and answer application functions.

  • Method

    AVR combines multi-format supervised fine-tuning with FS-GRPO to select among Full, Perception-Only, and Direct Answer formats according to task requirements.

  • Results

    50–90% token reduction across seven benchmarks maintains or improves accuracy and generalizes across model scales and families.

  • Takeaways & Limitations

    Adaptive, task-aware reasoning can improve multimodal inference efficiency while preserving or improving accuracy.

Abstract

from arXiv · show

Visual reasoning models (VRMs) have recently shown strong cross-modal reasoning capabilities by integrating visual perception with language reasoning. However, they often suffer from overthinking, producing unnecessarily long reasoning chains for any tasks. We attribute this issue to \textbf{Reasoning Path Redundancy} in visual reasoning: many visual questions do not require the full reasoning process. To address this, we propose \textbf{AVR}, an adaptive visual reasoning framework that decomposes visual reasoning into three cognitive functions: visual perception, logical reasoning, and answer application. It further enables models to dynamically choose among three response formats: Full Format, Perception-Only Format, and Direct Answer. AVR is trained with FS-GRPO, an adaptation of Group Relative Policy Optimization that encourages the model to select the most efficient reasoning format while preserving correctness. Experiments on multiple vision-language benchmarks show that AVR reduces token usage by 50--90\% while maintaining overall accuracy, especially in perception-intensive tasks. These results demonstrate that adaptive visual reasoning can effectively mitigate overthinking in VRMs. Code and data are available at: https://github.com/RunRiotComeOn/AVR.

1 Introduction

VRMs can overthink visual questions by generating full reasoning paths even when simpler cognitive paths suffice. AVR addresses this redundancy by selecting task-appropriate formats, reducing tokens while preserving or improving accuracy.

  • VRMs integrate visual perception with language reasoning for compositional, multi-hop, and logic-intensive visual tasks.
  • Reasoning Path Redundancy arises because many visual questions do not require all three cognitive functions: perception, reasoning, and answer application.
  • AVR decomposes visual reasoning into three cognitive functions and adaptively selects Full, Perception-Only, or Direct Answer formats.
  • AVR combines format-annotated supervised fine-tuning with FS-GRPO reinforcement learning to encourage efficient format selection while preserving correctness.
  • Perception-intensive tasks use mostly Direct Answer and Perception-Only formats, reducing token usage by over 80% with improved accuracy, while MathVista uses more Full Format responses when needed.
  • 50–90% token reduction across seven benchmarks accompanies matching or improved accuracy, with generalization across model scales and families.

2 Related Work

Related work improves visual-language reasoning and language-model reasoning through multimodal chains, decomposition, reinforcement learning, and concise reasoning strategies. These approaches motivate AVR’s focus on adaptive visual reasoning efficiency.

  • Multimodal Chain-of-Thought reasoning and reasoning decomposition have improved performance on complex visual-language tasks.
  • Reinforcement learning, including RLVR and GRPO, has become a major approach for improving language-model reasoning using verifiable outputs.
  • Research on overthinking uses post-training and reasoning-trajectory distillation to produce concise yet effective reasoning steps.

3 Preliminary

The preliminary analysis identifies redundant reasoning paths in VRMs and quantifies their prevalence using compressed, correctness-preserving responses. These findings motivate adaptive selection of cognitive functions and reasoning formats.

  • Visual reasoning outputs follow visual perception, logical reasoning, and answer application as three structured cognitive functions.
  • Reasoning Path Redundancy occurs when a shorter subset of functional components produces the same correct answer.
  • 36.7% of Qwen3-VL-4B-Thinking questions can omit reasoning, while 32.1% can be solved through direct answering without explicit perception or reasoning.
  • The Overthinking Score compares original-response token length with a minimally sufficient compressed response that preserves correctness.
  • 35.4% of instances have an Overthinking Score greater than 3, and 41.8% fall between 1 and 3.
  • Selective invocation of cognitive functions can preserve correctness while reducing token consumption, motivating task-adaptive format selection.

4 Method

AVR decomposes visual reasoning into explicit cognitive functions and trains models to select among formats with different reasoning effort. Its reward design jointly favors correctness, concise responses, and format diversity during adaptive switching.

  • Framework Design: AVR decomposes visual reasoning into Perception, Reasoning, and Answer functions, replacing a monolithic reasoning structure with explicit stages.The design supports distinct response formats and provides an explicit mechanism for mitigating overthinking.
  • Training Procedure: AVR uses SFT on 11k annotated questions before FS-GRPO trains dynamic switching among formats according to task requirements.SFT initializes multi-format generation, while reinforcement learning targets efficient format selection while maintaining accuracy.
  • Framework Design: The framework defines Full, Perception-Only, and Direct Answer formats, progressively bypassing exhaustive reasoning for greater token efficiency.Full Format includes all components; Perception-Only combines perception with answer; Direct Answer targets immediate response efficiency.
  • Optimization: FS-GRPO computes group-relative advantages from sampled-output rewards and applies PPO-style clipping with a KL penalty for training stability.The reward incorporates format efficiency, verbosity penalties, and format diversity.
  • Reward Design: The format reward combines answer correctness with format-specific bonuses ordered as f3 > f2 > f1 ≥ 0, favoring shorter formats when confidence is high.Incorrect compressed answers are penalized, allowing longer formats when needed.
  • Reward Design: The diversity reward counters format collapse by rewarding less frequent formats within each sampled response group.Its influence decays over training so later optimization emphasizes optimal format selection over exploration.

5 Experiments

AVR is evaluated across diverse vision-language benchmarks, model scales, and task types using accuracy and token consumption. It reduces reasoning cost while preserving or improving performance, especially by adapting formats to task demands.

  • Evaluation setup: AVR is evaluated across perception-intensive, reasoning-intensive, and general benchmarks using accuracy and average token usage.The study uses Qwen3-VL-2B/4B/8B-Instruct models and compares against thinking variants, TON, and ARM2.
  • Baseline behavior: Thinking models degrade by approximately 5% on perception-intensive benchmarks while consuming nearly 10× more tokens than instruct models.On general and reasoning-intensive tasks, they consume 2–4× more tokens while gaining around 5% performance.
  • AVR versus thinking models: AVR achieves over 80% token reduction and a 2–4% accuracy gain on perception-intensive tasks compared with conventional thinking models.For reasoning-intensive tasks, AVR uses full format for approximately 70% of responses, preserving sufficient reasoning while reducing unnecessary computation.
  • AVR versus other methods: AVR reduces MMMU average token usage from 285.0 to 128.7 while improving accuracy from 50.4% to 59.8% against ARM2.The comparison reports both lower token consumption and higher accuracy on most benchmarks.
  • Adaptive format selection: AVR dynamically shifts toward lightweight formats for perception-intensive tasks and toward full format for reasoning-intensive tasks.On perception-heavy tasks, direct-answer and perception-only formats are favored; reasoning-intensive tasks receive more full-format responses.

6 Analysis

Ablation and sensitivity analyses show that AVR’s three-format design and balanced training rewards are central to efficient, adaptive format selection. The method performs best when length regularization and format preferences remain properly balanced.

  • Format ablation: The three-format design achieves the best balance between accuracy and efficiency among the tested format combinations.Full-format-only responses consume the most tokens; removing perception-only or direct-answer formats lowers accuracy in different task settings.
  • Diversity reward: Removing the diversity reward causes format collapse toward direct answers, while retaining it preserves all three formats during training.The diversity reward mitigates early format dominance and supports dynamically balanced reasoning strategies.
  • Length tolerance: A small length tolerance over-compresses complex reasoning, whereas a large tolerance increases token use without meaningful accuracy gains.The default L = 300 achieves the best efficiency–accuracy balance across OCRBench and MathVista.
  • Format bonus: Model performance is relatively insensitive to exact format-bonus values when their ordinal relationship is preserved.A hierarchical bonus structure is essential for reliably selecting the intermediate perception-only format.
  • Overall sensitivity: AVR’s effectiveness primarily depends on balancing length regularization with format preferences rather than on highly sensitive hyperparameter values.This conclusion summarizes the reported sensitivity results.

7 Conclusion

AVR addresses reasoning-path redundancy by decomposing visual reasoning into cognitive functions and training adaptive selection among three formats. Across diverse benchmarks, it reduces token use while maintaining or improving accuracy.

  • Conclusion: AVR identifies Reasoning Path Redundancy as a key inefficiency driver in visual reasoning and trains models to select among three reasoning formats.The framework uses SFT and FS-GRPO to support adaptive, task-aware reasoning.
  • Conclusion: AVR achieves a 50–90% token reduction while maintaining or improving accuracy across diverse benchmarks, model scales, and families.The paper presents adaptive, task-aware reasoning as a promising direction for efficient multimodal inference.

A.1.1 Statistics

The statistics analyze how often visual questions can be answered without invoking all reasoning stages. They also specify a perception-only prompting procedure that extracts relevant visual information without reasoning or answering.

  • Statistics: In a sample of 1,000 instances, 36.7% of questions were answerable without the reasoning stage and 32.1% through direct answering.The study tested perception-only responses by withholding the image after providing generated perception output, and direct answering by prompting the model directly.
  • Perception prompting: The perception extraction prompt instructs the model to report only question-relevant visual information, including objects, attributes, text, colors, and spatial relations.It explicitly prohibits hidden inference, reasoning explanations, and final answers.

A.1.2 Overthinking Score Implementation

The Overthinking Score measures reasoning redundancy by comparing a correct model response with a compressed, minimally sufficient version. Compression removes unnecessary reasoning while preserving correctness, but exact minimality is not guaranteed.

  • The procedure prompts Qwen3-VL-4B-Thinking to answer each visual question in its default reasoning mode, then filters responses for correctness before compression.
  • The resulting compressed response provides a practical upper bound on minimal sufficient length, so Overthinking Scores are likely conservative estimates of true redundancy.
  • The compression prompt instructs rewriting each response into the shortest correct answer by removing redundant reasoning, explanations, and intermediate thoughts.

A.1.3 Case Study

The case studies and experiments show that AVR adapts reasoning depth to task requirements, reducing unnecessary generation while preserving or improving accuracy. Training remains stable, format usage stays diverse, and results transfer across model families, although hidden reasoning within formats remains a concern.

  • Adaptive formats: AVR represents reasoning pathways with Direct Answer, Perception-Only, and Full Format, corresponding to none, partial, or complete use of the reasoning pipeline.The compact action space supports stable reinforcement learning while explicitly representing different reasoning depths.
  • Training dynamics: Training reward and accuracy increase and later stabilize, mean response length decreases, and diversity reward maintains a balanced mixture of formats.Validation trends similarly stabilize, suggesting efficient format control without overfitting.
  • Overall results: AVR reduces token usage across task categories while maintaining or improving accuracy, using shorter formats for simple tasks and Full Format for complex ones.Compared with unconditional CoT generation, AVR reduces token consumption while avoiding the reported overthinking behavior.
  • Cross-family generalization: AVR achieves consistent improvements across Qwen3-VL and InternVL3, suggesting adaptive reasoning is not tied to one architecture or training paradigm.The two model families differ in multimodal fusion and training strategy.
  • Perception-intensive tasks: OCRBench favors Perception-Only responses at 47.5%, which achieve 90.3% accuracy, while Full Format reaches 75.2% and Direct Answer 86.2%.The comparison indicates that unnecessary reasoning can coincide with lower accuracy on perception-intensive tasks.
  • Reasoning-intensive tasks: MathVista favors Full Format at 78.3%, which achieves the best accuracy at 43.7%, while lightweight formats are rarely selected and perform worse.This pattern supports allocating explicit reasoning to tasks requiring multi-step inference.
  • General tasks: MMMU uses all three formats more evenly, with Perception-Only at 59.2% and Direct Answer at 55.6%, reflecting heterogeneous reasoning requirements.
  • Limitation: A remaining concern is that models may hide reasoning inside earlier stages, superficially following predefined formats and weakening interpretability.Stage-mixing analyses report that most cases adhere to the intended separation, but the concern remains a stated limitation.
Loading 2604.14568v1…