Technical Summary: ETCHR (Editing To Clarify and Harness Reasoning)

1. Problem Statement

Multimodal Large Language Models (MLLMs) have advanced visual reasoning, yet purely textual Chain-of-Thought (CoT) reasoning remains a bottleneck for tasks requiring fine-grained focus or view transformations. Models often fail to verbalize spatial states they cannot draw, leading to compounding descriptive errors.

Existing "think with images" paradigms attempt to bridge this gap by generating intermediate images during inference but suffer from two primary limitations:

1. Tool-based Methods: Rely on predefined, low-level action tokens (e.g., bounding boxes, crop commands) executed by a deterministic renderer. These are constrained to localized manipulations and often require task-specific fine-tuning of the MLLM, eroding general competence.
2. Unified Multimodal Models: Use a single backbone to interleave text and image tokens. While flexible, these models often possess weaker generative heads, producing intermediate images that inject noise rather than guidance. Furthermore, they lack mechanisms to verify if an intermediate edit is correct before using it for reasoning.

Current off-the-shelf image editors, when used as reasoning assistants, exhibit two specific gaps:

• Language-side Gap: Editors trained as passive instruction-followers cannot map abstract questions (e.g., "Is the object on the left?") to appropriate visual transformations without explicit prompts.
• Generation-side Gap: Edit correctness degrades significantly as reasoning depth increases (e.g., tracing long paths in mazes), indicating a failure to execute multi-step spatial reasoning during the diffusion decoding phase.

2. Methodology

The authors propose ETCHR (Editing To Clarify and Harness Reasoning), a framework that decouples a dedicated, question-conditioned image editor from the downstream understanding model. The system operates via a two-stage training recipe and a reflective inference procedure.

2.1. Architecture

• Base Model: A specialist image-to-image editor (FLUX.2-klein-base-9B) is used as the intermediate-image generator.
• Decoupling: The editor is trained independently of the understanding model (MLLM), allowing it to plug into various open- and closed-source MLLMs without retraining the MLLM itself.
• Inference Pipeline (Edit-Verify-Reason):
  1. Edit: The editor generates an intermediate image based on the input image and question.
  2. Verify: The understanding model performs a binary check to determine if the edit contains the necessary visual information to answer the question.
  3. Reason: If verification passes, the model reasons using the edited image; otherwise, it falls back to the original image to prevent noisy edits from contaminating the final answer.

2.2. Training Strategy

ETCHR is trained in two stages to address the identified gaps:

Stage I: Reasoning Imitation (Supervised Fine-Tuning)

• Goal: Convert the editor from a passive instruction-follower into a question-conditioned system.
• Data: A large-scale corpus of edit trajectories across five task families: fine-grained perception, chart understanding, logic reasoning, jigsaw restoration, and 3D understanding.
• Technique: The editor is fine-tuned using Low-Rank Adaptation (LoRA) on pairs of (Input Image, Question, Ground-Truth Edited Image). A task-level meta-prompt is prepended to the question to act as a soft router, partitioning the latent space and reducing cross-task interference.

Stage II: Reasoning Enhancement (Reinforcement Learning)

• Goal: Align the editor with downstream reasoning utility and improve edit correctness.
• Reward Design: Two complementary rewards are derived from a VLM-as-Judge:
  1. Editing Guidance Reward ($r_{guide}$): Measures if the downstream MLLM answers correctly using the edited image.
  2. Editing Correctness Reward ($r_{correct}$): Measures if the edit itself contains the necessary visual evidence (e.g., correct bounding box) regardless of the final answer.
• Optimization: The authors use Pref-GRPO (a pairwise-preference extension of GRPO) to optimize the editor. The combined reward is a convex sum of the two signals ($\alpha = \beta = 0.5$), balancing the "ceiling-bound" nature of guidance rewards with the "ceiling-lifting" but noisy nature of correctness rewards.

3. Key Contributions

1. Mechanism Proposal: A "think with images" mechanism utilizing a dedicated image-to-image editor as the source of intermediate visual evidence, coupled with an inference-time verification step to reject unreliable edits.
2. ETCHR Implementation: A reasoning-aware editor built on FLUX.2-klein-base-9B, trained via a two-stage recipe (Reasoning Imitation SFT + Reasoning Enhancement RL) using VLM-derived rewards.
3. Training-Free Integration: Due to the decoupled design, ETCHR can be paired with diverse open- and closed-source MLLMs (e.g., Qwen3-VL, Gemini, Kimi K2.5) without fine-tuning the MLLM, demonstrating broad applicability across fine-grained perception, spatial reasoning, puzzle restoration, and 3D understanding.

4. Experimental Results

The authors evaluated ETCHR across five task families and nine benchmarks, comparing it against tool-based baselines (DeepEyes, Thyme) and unified models (Bagel-Zebra-CoT, ThinkMorph).

• Performance Gains: ETCHR consistently improved Pass@1 scores across all tested MLLMs:
  • Qwen3-VL-8B: Increased from 55.95 to 60.77 (+4.82).
  • Gemini-3.1-Flash-Lite: Increased from 65.08 to 70.55 (+5.47).
  • Kimi K2.5 (1T MoE): Increased from 76.55 to 81.16 (+4.61).
• Task-Specific Improvements: Significant gains were observed in logic reasoning (e.g., Maze solving) and 3D understanding, areas where tool-based methods failed due to limited action spaces and unified models struggled with generation fidelity.
• Comparison with Closed-Source Editors: When compared against a closed-source editor (Nano Banana 2), ETCHR showed comparable performance on perception tasks but superior results on structural tasks (Logic, Jigsaw, 3D), suggesting that reasoning-aware training is more critical than model scale for these tasks.
• Ablation Studies:
  • Both SFT and RL stages contributed to performance, with SFT providing the primary boost in reasoning capability.
  • Combining both reward signals (Correctness and Guidance) outperformed using either alone.
  • The "Verify" step in the inference pipeline provided consistent gains on tasks with high baseline accuracy (Perception, Chart), while its utility varied on tasks with low baseline accuracy (Logic, Jigsaw).

5. Significance and Limitations

Significance:
The paper argues that the "think with images" paradigm requires a shift from fixed toolkits or unified backbones to a decoupled, specialist editor capable of inferring visual transformations from abstract questions. By addressing the language-side and generation-side gaps through a two-stage training recipe and a verification mechanism, ETCHR provides a robust, plug-and-play solution that enhances visual reasoning without compromising the general capabilities of the underlying MLLM.

Limitations:

• Sampling Granularity: The GRPO-based RL training exhibits limited semantic diversity for complex structural edits (e.g., jigsaw restoration), constraining policy exploration in those specific domains.
• Downstream Ceiling: The end-to-end performance remains bounded by the reasoning capabilities of the downstream MLLM; even optimal edits cannot overcome the inherent reasoning limits of the base model.
• Temporal Cost: Image editing introduces higher time overhead compared to text-based reasoning, which may be inefficient for samples requiring only short Chain-of-Thought.