Full Marks in Every Subject, Zero When Combined: XDomainBench Reveals AI's 'Cross-Domain Reasoning Collapse'

If a student scores 100 in physics, 100 in chemistry, and 100 in biology, you'd assume they're an all-rounder, right?

But drop them into a "new drug development" project that requires physics, chemistry, and biology simultaneously, and they can't even take the first step—fumbling formulas and using a microscope as a hammer. That would be bizarre.

Today's top large language models (LLMs) are exactly this kind of "high-scoring, low-capability" fake polymath.

In May 2026, a paper titled 《XDomainBench: Diagnosing Reasoning Collapse in High-Dimensional Scientific Knowledge Composition》 caused a stir on arXiv. The researchers built a "magic mirror" called XDomainBench specifically to test AI's cross-disciplinary ability.

The result was an almost comical finding: as the number of involved disciplines increases from 1 to 4, the AI's brain undergoes an irreversible "Reasoning Collapse."

Why Does AI Have "Subject Imbalance Syndrome"?

Feynman once quipped that the greatest failure of education is producing nerds who can only solve specific equations without knowing the world is interconnected.

Previously, we tested AI one subject at a time (e.g., with MMLU): this question is physics, that one is history. AI performs brilliantly in such solo battles. But real science (AI for Science) is never isolated.

To test AI's true level, XDomainBench introduces a metric called "composition dimension (k)":

• When k=1 (pure physics question), the AI chats with ease.
• When k=3 (combining thermodynamics, cell biology, and statistics), the AI instantly becomes a fool.

The Three Culprits Behind the Collapse

The researchers dissected the collapse like forensic examiners and identified three fatal flaws:

1. Domain Confusion

Every discipline has its own "rules of the trade." In physics, "mass" is absolute; but in some social sciences, "mass" might be a subjective evaluation. When AI is asked to think across both fields simultaneously, it's like someone playing chess and poker at the same table—it quickly plays the rook as a joker.

2. Composition Overhead

Even the strongest AI has limited attention within its working memory (context window). When it must forcibly combine knowledge from different dimensions (say, mathematical formulas and biological phenomena), its brain burns enormous compute searching for connection points. This "mental overdraft" makes it botch even simple arithmetic.

3. Snowballing Error Accumulation

In multi-turn reasoning, if a tiny flaw appears in the first round when combining chemistry and physics (say, a wrong unit), by the second round when biology enters, that error gets amplified 100x, and the entire reasoning chain snaps in an instant.

Why Does This Matter?

Feynman spent his life breaking down disciplinary walls—he could explain quantum mechanics using the sound of a tapping water glass. Because the universe doesn't come pre-divided into subjects; classification is just a compromise of human cognitive limits.

This paper sounds an alarm: We can no longer measure AI's intelligence by exam scores. True intelligence isn't about how many books you stuff into a database, but whether you can build a bridge called "insight" between seemingly unrelated fields. If the bridge is broken, knowledge is just a pile of dead matter.

Summary

Assembling is not integration, and erudition is not wisdom.

XDomainBench reveals one of the hardest chasms on the road to AGI (artificial general intelligence): the organic fusion of high-dimensional knowledge.

Next time an AI claims to "master all disciplines," give it a chained problem requiring physics, history, and economics at once. You'll find that, facing this complex real universe, it's nothing more than a cyber-nerd clutching three different textbooks, at a loss.

Breaking down the walls between disciplines is where true intelligence begins. This is 2026 cognitive science's deepest diagnosis of "real versus fake polymaths."

---

*Translation note: The original title of the paper contains the Chinese publication framing; the English paper title is XDomainBench: Diagnosing Reasoning Collapse in High-Dimensional Scientific Knowledge Composition (arXiv, May 2026, as cited in the source post).*