💬 NLP

Learning Adaptive Reasoning Paths for Efficient Visual Reasoning

The paper proposes AVR, an adaptive visual reasoning framework that dynamically selects between full reasoning, perception-only, or direct answer formats to significantly reduce token usage and mitigate overthinking in visual reasoning models while maintaining accuracy.

Original authors: Yixu Huang, Tinghui Zhu, Muhao Chen

Published 2026-04-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yixu Huang, Tinghui Zhu, Muhao Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Over-Thinker" Robot

Imagine you have a brilliant but slightly anxious robot assistant. You ask it a simple question: "What color is the stop sign?"

A normal human would look, see red, and say, "Red."

But this robot? It starts a 10-page essay. It describes the texture of the sign, the angle of the sun, the history of traffic laws, and the chemical composition of red paint before finally concluding, "It is red."

This is what the paper calls "Overthinking."

Current AI models (Visual Reasoning Models) are great at solving hard problems, but they have a bad habit of using the same massive, complex "thinking chain" for every question, whether it's simple or hard. This wastes a huge amount of computer power (tokens) and money, and sometimes, by talking too much, the robot actually makes mistakes.

The Solution: AVR (The Smart Traffic Controller)

The authors propose a new framework called AVR (Adaptive Visual Reasoning).

Think of the AI's brain as a kitchen.

Old Way: Every time you order a sandwich, the chef chops vegetables, boils water, grills meat, and bakes bread, even if you just wanted a slice of cheese on toast.
AVR Way: The chef looks at the order and asks, "Do I need the whole kitchen, or just the toaster?"

AVR teaches the AI to recognize the difficulty of the task and choose one of three specific "uniforms" (formats) to wear:

The "Direct Answer" Uniform (The Express Lane):
- When to use: For easy questions like "What color is the sky?"
- Action: The AI skips the thinking entirely and just says, "Blue."
- Analogy: Walking straight to the fridge to grab a soda. No cooking required.
The "Perception-Only" Uniform (The Observer):
- When to use: For questions that need you to see but not think hard, like "How many balls are in the picture?"
- Action: The AI describes what it sees ("I see three red balls") and gives the answer. It skips the complex logic steps.
- Analogy: Counting apples in a basket. You just look and count; you don't need a math degree to do it.
The "Full Format" Uniform (The Detective):
- When to use: For hard questions like a complex math problem or a tricky logic puzzle.
- Action: The AI puts on its thinking cap. It looks, reasons step-by-step, checks its work, and then answers.
- Analogy: Solving a mystery. You need to gather clues, connect the dots, and deduce the solution.

How Did They Teach the Robot? (The Training)

The authors didn't just tell the robot to "be smart." They used a two-step training process:

Step 1: The Classroom (SFT): They showed the robot thousands of examples where the answer was given in one of the three formats. This taught the robot how to speak in these different styles.
Step 2: The Video Game (Reinforcement Learning): This is the magic part. They played a game with the robot where:
- If it gave the correct answer using the shortest possible method, it got a big bonus.
- If it gave the correct answer but used a long, unnecessary explanation, it got a smaller bonus.
- If it gave the wrong answer, it got no points.

Over time, the robot learned: "Hey, if I just say 'Red' for the stop sign, I get points faster and use less energy. If I try to write a novel, I might mess up and lose points."

The Results: Faster, Cheaper, Smarter

The paper tested this new robot on many different tasks. The results were impressive:

Token Savings: The robot used 50% to 90% fewer words (tokens) to solve problems. This means it's much cheaper to run and much faster.
Accuracy: It didn't get dumber. In fact, on simple tasks, it got more accurate because it stopped over-complicating things.
Adaptability: It learned to switch uniforms instantly.
- Simple task? "Direct Answer."
- Hard math problem? "Full Detective Mode."

The Big Takeaway

The paper argues that not every problem needs a PhD-level solution.

Just like a human doesn't need to use a supercomputer to decide what to wear in the morning, an AI shouldn't need a massive reasoning chain to answer a simple question. By teaching AI to be adaptive—knowing when to think deeply and when to just look and answer—we can make them faster, cheaper, and actually more reliable.

In short: AVR teaches AI to stop overthinking and start being efficient, saving time and money while keeping the answers correct.

1. Problem Statement: Reasoning Path Redundancy

The paper identifies a critical inefficiency in current Visual Reasoning Models (VRMs): Overthinking. While VRMs have demonstrated strong capabilities by integrating visual perception with language reasoning (Chain-of-Thought, or CoT), they frequently suffer from Reasoning Path Redundancy.

The Issue: Models often generate unnecessarily long reasoning chains for simple tasks that do not require complex logical deduction.
The Cause: Unlike purely textual reasoning, visual reasoning involves three distinct cognitive functions: Visual Perception, Logical Reasoning, and Answer Application. However, many visual questions can be solved correctly using only perception (or even direct knowledge) without invoking the full logical reasoning pipeline.
Consequences: This redundancy leads to excessive token consumption (increasing cost and latency) and can sometimes degrade accuracy due to error accumulation in intermediate reasoning steps.
Current Limitations: Existing solutions rely on heuristic strategies or binary "think or not" approaches that fail to account for the nuanced, multi-stage cognitive structure of visual reasoning.

2. Methodology: The AVR Framework

The authors propose AVR (Adaptive Visual Reasoning), a framework designed to dynamically allocate reasoning effort based on task complexity.

A. Cognitive Decomposition & Response Formats

AVR decomposes the reasoning process into three functional components and defines three corresponding response formats:

Full Format: <perception>...<reasoning>...<answer> (Used for complex tasks requiring multi-step deduction).
Perception-Only Format: <perception>...<answer> (Used for tasks requiring visual grounding but minimal logical inference).
Direct Answer: <answer> (Used for simple recognition or knowledge-based queries).

B. Two-Stage Training Pipeline

AVR employs a two-stage training approach to teach the model both format generation and adaptive selection:

Stage 1: Supervised Fine-Tuning (SFT) for Format Understanding
- The model is fine-tuned on a dataset of 11k diverse questions annotated with the three formats.
- Goal: Teach the model to generate responses in all three distinct formats correctly, establishing a "strategic initialization."
Stage 2: Reinforcement Learning (RL) for Efficient Format Switching
- The authors introduce FS-GRPO (Format-Selection Group Relative Policy Optimization), an adaptation of the GRPO algorithm.
- Reward Function ( $r_i$ ): The total reward combines three components:
  - Format Reward ( $r_{format}$ ): Incentivizes correct answers with shorter, more efficient formats (Direct Answer > Perception-Only > Full). It assigns higher bonuses to shorter formats ( $f_3 > f_2 > f_1$ ).
  - Diversity Reward ( $r_{div}$ ): Prevents "Format Collapse" (where the model converges to only one format) by rewarding less frequent formats within a sampled group. This uses a time-decaying factor to encourage exploration early in training.
  - Length Scaling Factor ( $s_L$ ): Penalizes verbosity if the token count exceeds a tolerance threshold ( $L$ ).
- Objective: Maximize the probability of selecting the most efficient format that still yields the correct answer.

3. Key Contributions

Identification of Redundancy: The paper provides empirical evidence (via an "Overthinking Score") that ~35% of visual questions exhibit a reasoning path redundancy score > 3, and ~32% can be solved via direct answering without explicit perception or reasoning.
Adaptive Framework (AVR): A novel framework that explicitly decouples perception, reasoning, and answering, allowing the model to dynamically switch between three response formats.
FS-GRPO Algorithm: A specialized RL objective that jointly optimizes for correctness, token efficiency, and format diversity, preventing the model from collapsing into a single reasoning style.
Generalizability: The approach is validated across multiple model scales (2B, 4B, 8B) and different model families (Qwen3-VL, InternVL3), demonstrating robustness.

4. Experimental Results

The authors evaluated AVR on seven benchmarks covering perception-intensive (OCRBench, TextVQA), reasoning-intensive (MathVista, MathVision), and general (MMMU, MMBench) tasks.

Token Efficiency: AVR reduces token usage by 50–90% compared to standard "Thinking" VRMs.
- On perception-intensive tasks (e.g., OCRBench), token reduction exceeded 80%.
- On reasoning-intensive tasks, the model still reduced tokens significantly while maintaining accuracy.
Accuracy:
- AVR maintains or improves accuracy compared to baseline thinking models.
- On OCRBench, AVR achieved a 3.1% accuracy gain alongside massive token reduction.
- On MathVista, it matched the performance of full-reasoning models while using significantly fewer tokens.
Comparison with Baselines: AVR outperforms existing selective reasoning methods like Think-or-Not (TON) and ARM2.
- Unlike TON's binary choice (Direct vs. Full CoT), AVR's three-format approach allows for a "Perception-Only" middle ground, which is crucial for visual tasks.
- On MMMU, AVR-4B reduced tokens from 285.0 to 128.7 while improving accuracy from 50.4% to 59.8% compared to ARM2.
Dynamic Selection: Analysis shows the model correctly adapts its strategy:
- Perception-heavy tasks: ~50% use Perception-Only, ~25% use Direct Answer.
- Reasoning-heavy tasks: ~78% use Full Format.

5. Significance

This work addresses a fundamental inefficiency in the current paradigm of "more reasoning is better." By recognizing that not all visual questions require full logical deduction, AVR introduces a task-aware mechanism that:

Reduces Computational Cost: Drastically lowers inference costs (tokens) without sacrificing performance.
Improves Reliability: Mitigates the risk of errors introduced by unnecessary intermediate reasoning steps.
Enhances Interpretability: The explicit decomposition of cognitive functions provides a clearer structure for how models process visual information.
Scalability: The method is model-agnostic and effective across different parameter sizes, suggesting that adaptive reasoning is a generalizable capability for future multimodal systems.

In conclusion, AVR demonstrates that adaptive, task-aware reasoning is superior to static, exhaustive reasoning chains, offering a path toward more efficient and robust Visual Reasoning Models.