REA Reverse-Engineering Test: Gemini 4 Argon and GLM 5.3 Recreate the Palace Museum Famous Paintings Chronicle

Reverse engineering tools can help models identify a piece of software's structure and behavior, but which features should be kept and which details cannot be omitted still have to be decided by the model.

In early October, well-known AI tech blogger Matthew Berman left this exclamation on X, pushing an open-source reverse engineering analysis tool called REA (Reverse Engineer Anything) into the spotlight.

REA (Reverse Engineer Anything) is one of the hottest projects on GitHub in October. What makes it formidable is that it gives AI Agents a "disassembly" perspective, directly extending the goal of AI programming to off-the-shelf software. It hands the ability to analyze binaries, applications, and runtime behavior to Agents through a unified interface: native programs can connect to engines such as Ghidra, IDA, and Hopper, while web pages can be inspected for structure, scripts, network, and screenshots. Analysis results come with evidence and known limitations for the model to trace.

What makes it attractive is that it connects the study of off-the-shelf software into coding sessions, so that when developers see a feature, they can ask the Agent to investigate the components and (Leiphone) behavior, and then write the findings into the implementation spec. This means that materials that previously had to be passed back and forth manually now have the chance to continue being used within the same task.

This greatly lowers the barrier to reverse engineering. On October 16, Perplexity co-founder and CEO Aravind Srinivas announced that REA Skill had been integrated into Computer, along with trial credits, making reverse engineering analysis begin to be treated as one of the features of Agent products; soon afterward, a developer demonstrated how to use Codex and REA to rebuild a Claude Slides editor with nearly complete functionality.

02

No matter how powerful the tool is, it still comes down to the "person"

Faced with the controversy, we must recognize an engineering truth: using tools to find clues and using code to rebuild a system have never been the same thing.

REA is like a powerful "X-ray machine." It can illuminate the skeleton of existing software with complete clarity, dumping tens of thousands of DOM nodes, scattered API endpoints, and messy low-level scripts in front of you all at once. But at best, this is only an unpolished and possibly trap-filled "architectural blueprint."

Faced with such a massive amount of contextual information, can the blueprint be restored one-to-one into a precisely operating "skyscraper," or will only a drafty thatched hut with an empty shell be hastily built? All of this depends entirely on the "construction crew" doing the work, that is, the hard strength of the large model itself in long-text information filtering, code architecture understanding, and multi-step logical reasoning. No matter how powerful the tool is, in the end it still depends on which "brain" is behind the controls.

At exactly this moment, Google, which had been quiet for a while, made a heavyweight return with its new-generation flagship reasoning model Gemini 4 Argon. This model makes no secret of its ambition, aiming directly at complex software engineering, professional knowledge work, and long-horizon Agent tasks that require continuous tool calls; meanwhile, in the domestic large model camp, GLM 5.3 has also recently shown extremely strong dominance in code generation and long-context processing.

On one side is the wildly popular reverse engineering "artifact," and on the other are top Chinese and foreign large models that claim a major leap in coding intelligence.

If the same blueprint were handed to them, what would happen? With this question in mind, in OpenCode, together with REA and Reverse-skill, we had Gemini 4 Argon and GLM 5.3 separately disassemble and replicate the Palace Museum's online famous painting appreciation platform, "Palace Museum Famous Paintings."

This website not only includes more than a thousand ancient paintings, but also supports high-definition zooming, artwork filtering, and viewing detailed information. It is not just a simple image display page.

Homepage of the Palace Museum Famous Paintings website

The intersection of REA and Argon adds another application scenario to AI programming: investigating off-the-shelf software and rebuilding existing features. At the same time, this will further change the industry division of labor: the competitive pressure on software companies will further shift to data, services, and continuous maintenance; model vendors, meanwhile, must prove that their models can turn analytical clues into reliable delivery.

The target of this test is locked on Palace Museum Famous Paintings, with the core being to have the models reverse engineer and replicate the website's artwork display functionality, while not involving login, favorites submission, comments, or the complete annotation service. On the original Famous Paintings website, the artwork list has filters for dynasty, author, school, and subject; clicking an artwork enters the detail page; once inside, zooming, dragging, the navigation map, and the information panel all need to maintain a reasonable state after continuous operations.

To observe the models' capabilities more intuitively, we chose GLM 5.3 as the control model for Gemini 4 Argon, having both models use REA to observe and analyze the site's functionality, then use Reverse-skill to reverse engineer and restore the site, and finally, after the models complete the test, provide a replicated version of the Famous Paintings website.

First, let's look at how Gemini 4 Argon performs at handling REA. Judging from the homepage screenshot, it preserved the page skeleton of Famous Paintings; the parts consistent with the official site include the three-column artwork layout, four filters, clear options, artwork titles, and author and dimension information. The header texture was simplified, and the fonts and whitespace also differ, but the page's hierarchy was not rearranged.

Gemini 4 Argon's replica of the Famous Paintings artwork display homepage

Furthermore, as can be seen from the image below, comparing the original site on the left with the replicated version on the right, the top navigation, filter area, and card text positions basically correspond one-to-one. Although the images were replaced with geometric patterns, it is still evident that the model followed the original site's structure rather than building a separate generic gallery.

Comparison between the Gemini 4 Argon replica and the original site (left: original site, right: replica)

Entering the detail page, the replica also retained the top banner, centered artwork information, the view-details entry, and related recommendations. However, in terms of details, the button and text styles were not fully aligned; the replica roughly restored the original site's functional design and content, but did not achieve complete consistency in the details.

Comparison between the Gemini 4 Argon replica and the original site on the artwork detail page (left: original site, right: replica)

Overall, the Gemini 4 Argon combination did a good job replicating the artwork display functionality of the Famous Paintings site, with corresponding core functions and detailed design.

As the control group, GLM 5.3 likewise produced a browsing path of "artwork list—detail—view painting," but once the list opened, the replication direction differed slightly: the three-column cards became four columns, the light-colored page was spread with brown, and the original site's large filter area disappeared. Unexpectedly, the model clearly noted in the header that this was a minimal browsing-flow version using local test images.

GLM 5.3's replica of the Famous Paintings artwork display homepage

Further comparison with the original site reveals that the four filters, top entry points, and card favorite icons did not appear in their corresponding positions; GLM 5.3's replica clearly made enormous changes to the site's structure.

Comparison between the GLM 5.3 replica and the original site (left: original site, right: replica)

Entering the detail page, the changes are even greater: the original site's organization of a large top image, centered title, and introduction area was replaced with a white content box; view details and related recommendations are still present, but the entire page is closer to a streamlined browsing prototype.

Comparison between the GLM 5.3 replica and the original site on the artwork detail page (left: original site, right: replica)

Overall, although GLM 5.3 completed a reverse-engineered replica, there are still considerable differences from the actual original site.

In short, the two runs delivered versions that can continue to be iterated; in terms of reverse-restoration effect, Gemini 4 Argon is closer to the original site, while GLM 5.3 retained the browsing flow but omitted and rewrote many details.

On September 30, Google announced that Argon would be opened to trusted cybersecurity partners, and on October 13 it expanded to the developer API and corresponding subscription products, with key uses including complex software engineering, professional knowledge work, and multi-step Agents.

Screenshot of Google's first announcement about Argon

It can be seen that what Google is seeking is developers willing to hand long tasks to the model, and the leaderboard gives Argon a positive signal. According to the intelligence index of the Artificial Analysis (AA) leaderboard, Argon's high tier scores the same as GPT-6 Astra and Fable 5.1's max tier (both 53 points after rounding), with Opus 5.5's 58 points and Sonnet 5.5's 56 points ahead. This means Argon already has reason to enter the flagship candidates.

Argon scores 53 points on the AA leaderboard

In addition, from the model evaluation results released by Google, the AA leaderboard scores are interpreted in more detail:

1. Most striking is knowledge work, where Argon leads in all four results: AutomationBench is 51.3%, higher than Astra's 41.4% and Opus's 42.5%; the financial Agent test is 65.4%. This is consistent with Box's positive assessment and also gives Google a reason to pursue enterprises' complex workflows.

2. Long-context scenarios are another advantage; facing GraphWalks' 256,000 to 1 million Token evaluation, Argon's F1 score is 84.2%, Astra's is 71.8%, and Opus's is 66.8%. This shows that Argon excels at tracking graph-structure relationships in long contexts and can still reason after reading in large amounts of information.

3. Coding results fluctuate noticeably: Argon scores 77.9% and 91.9% on DeepSWE and Vibe Code Bench respectively, leading its opponents in the table; switching to FrontierSWE, its 55.0% is lower than Astra's 65.5%; on Terminal-Bench 4.0 it also trails Opus's 66.4% with 57.4%. These differences show that Argon's coding ability is affected by task type and tool environment.

Argon evaluation results reported by Google officially

Overall, a more suitable industry positioning for Argon is a general flagship that excels in knowledge work and long context and has top-tier coding competitiveness. Therefore, Google already has reason to return to flagship competition, and developers have one more option worth seriously trying.

05

Leaderboard wins bring trials; only delivery retains users

At present, toolchain capabilities are leveling out, while model capabilities are diverging. REA lowered the threshold for reverse analysis; for the same site, both models could disassemble and rebuild it, and the gap is concentrated at the delivery layer: Argon retained the filter area and three-column cards, while GLM 5.3 simplified the structure into its own interpretation.

Both sets of results can open and can complete the browsing path of "list, detail, view painting"; the difference lies in the details. Argon's version followed the original site's hierarchy, with the four filters, three-column layout, detail-page banner, and related recommendations corresponding one-to-one; GLM 5.3's version is closer to a streamlined prototype, with the filter area gone and the detail page replaced by a white content box. Between being runnable and being restored properly lie precisely these easily overlooked correspondences.

What REA can do is bring the investigation phase into the coding session: structure, scripts, network, and screenshots can all be traced, and analysis results come with evidence and known limitations. What it lowers is the threshold for understanding a piece of software; the clues it analyzes are still far from reliably replacing a product. DeepTech divides understanding the mechanism, rewriting functionality, and reliable replacement into three levels, while Daqi Guancha simply questions whether this is another AI hallucination; the two voices remind us of the same thing: a page that can run is not the same as software that can be maintained long-term.

General functions are easier to redo, and products that charge for features are coming under pressure. Display-type sites like Famous Paintings have already been replicated quite thoroughly, and pure page functionality is increasingly difficult to price separately. What users are willing to keep paying for is compatibility, data services, and long-term maintenance; these things that require long-term investment will increasingly become the chips that determine whether they stay or go.

For Google, 53 points is already enough to put Argon back in the flagship candidate position. The reverse replication of Famous Paintings is proof of this capability on a real site: Argon scores in benchmarks and, when facing a specific site, also delivered something close to the original site. Trials won by the leaderboard must ultimately be retained through qualified deliveries time and again; in the future, whoever can steadily deliver long tasks to the finish line will truly stay at the table.

Click the keywords to view the corresponding hands-on content

English text is being prepared.

vivago R1

Seedance 2.5 vs Kling 3.0

Minimax H3 vs Seedance 2.0 Fast