Beyond Textual Repository Exploration: Dual-Modal Structural Reasoning for Agentic Issue Resolution
This paper introduces DUALVIEW, a dual-modal framework that enhances agentic issue resolution in large-scale repositories by replacing fragmented text-based navigation with persistent visual representations of code dependencies across four complementary graph views, thereby improving long-horizon exploration and structural reasoning.
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
Imagine you are a detective trying to solve a mystery in a massive, sprawling city (the software repository). Your job is to find a specific broken streetlight (the bug) and fix it.
The Old Way: The "Text-Only" Detective
Currently, most AI detectives work like this: They are given a list of text clues. They have to read a file, then read another file, then search for a keyword, then read a third file. They are trying to build a mental map of the city just by reading street signs one by one.
The problem? The city is huge. The connections between buildings (code dependencies) are complex. By the time the detective reads enough text to understand how two distant buildings are connected, they get lost, confused, or "drift" off track. They are trying to reconstruct a 3D map from a 2D list of words. It's slow, and they often miss the big picture.
The New Way: DUALVIEW
The paper introduces a new tool called DUALVIEW. Instead of just giving the detective a list of text clues, DUALVIEW gives them two things at once:
- A Visual Map: A clear, colorful diagram showing how the city's neighborhoods, streets, and buildings connect.
- A Textual Legend: A detailed written guide that explains exactly what each building is and where it is located.
The detective can look at the map to instantly see the "shape" of the problem (e.g., "Oh, this broken light is connected to three other buildings over here!") and then use the text to find the exact address to fix it.
The Four "Maps" DUALVIEW Uses
To make this work, DUALVIEW doesn't just show one map; it shows four different types of maps, depending on what the detective needs to see:
The Neighborhood Map (Module Coupling Graph):
- Analogy: A map showing which city districts are connected to which.
- Use: Helps the detective figure out which general area of the city to look in first, rather than searching the whole city blindly.
The Phone Directory (Function Call Graph):
- Analogy: A chart showing who calls whom on the phone.
- Use: If a specific function (a worker) is acting up, this map shows who called them and who they called next. It traces the flow of actions.
The Family Tree (Class Hierarchy Graph):
- Analogy: A family tree showing parents, children, and cousins.
- Use: In programming, some code is a "parent" and other code is a "child" that inherits traits. If the parent has a rule, the child follows it. This map helps the detective find the real source of the rule, even if the error report only mentions the child.
The Blueprint (Program Dependence Graph):
- Analogy: A detailed blueprint of a single room showing how the pipes and wires connect inside the walls.
- Use: Once the detective is inside the specific room (function), this map shows exactly how the data flows from one step to the next to find the exact broken wire.
How It Works in Practice
The paper tested this new method on real-world software repair tasks (using benchmarks called SWE-bench). Here is what they found:
- Faster Fixes: The AI agents using DUALVIEW solved more problems than those using the old text-only method.
- Less Confusion: The agents made fewer mistakes and didn't get "lost" in the code as often.
- Cheaper to Run: Surprisingly, even though the AI had to look at images (which usually costs more computer power), it actually saved money. Because the visual maps helped the AI find the answer faster, it didn't have to ask as many questions or read as many files overall.
- The Best Combo: The AI performed best when it had both the map and the text. The map helped it see the big picture quickly, while the text gave it the precise details needed to make the actual repair.
The Bottom Line
The paper argues that software repositories are naturally structured like complex networks (graphs), but we've been forcing AI to read them as simple lists of text. By adding a visual layer that lets the AI "see" the structure of the code, we can help it solve bugs much faster and more accurately. It's the difference between trying to navigate a city by reading a phone book versus looking at a GPS map.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.