Huawei’s Xiaoyi scored 28.6 points on the IPhO 2026 theoretical exam, surpassing the theoretical gold-medal cutoff. The result is impressive, but its true significance still depends on the system configuration, tool usage, and evaluation process.

Xiaoyi Surpasses the Physics Olympiad Gold-Medal Threshold: How Meaningful Is the Result?

Huawei announced today (July 30) that Xiaoyi had completed the theoretical exam for the 56th International Physics Olympiad (IPhO 2026), scoring 28.6 out of 30 points and exceeding this year’s theoretical gold-medal threshold.

That is unquestionably a very high score. IPhO theory problems are not ordinary exam questions where a few formulas are embedded in some context and answers can be produced through pattern matching. They often span multiple branches of physics and involve complex diagrams, long chains of derivation, and intensive symbolic computation. A single incorrect sign or boundary condition early on can invalidate several pages of subsequent calculations.

However, the result should not be interpreted simply as “Xiaoyi won an IPhO gold medal.” A more accurate statement is that Xiaoyi exceeded the corresponding gold-medal benchmark in the assessment of this year’s theoretical exam. The official IPhO competition also includes an experimental exam, and medals are generally awarded based on combined theory and experiment scores. Nor did the AI enter the competition as an officially registered student contestant. This distinction does not diminish the technical significance of the 28.6-point score, but it does affect how the result should be understood.

What Makes This Exam So Difficult?

This year’s theoretical exam consisted of three major problems spanning an unusually eclectic range of topics:

Torque equilibrium in hydrostatics, as well as relativistic scattering;
Multiple reflections in optics, vignetting, and derivation of a family’s envelope, including Taylor expansions;
Magnetic refrigeration cycles, proofs involving Maxwell relations, and integral calculations of cooling time.

Viewed individually, all of these topics can be found in university physics courses. The real difficulty is that the problems do not tell solvers which formulas to use, nor do they usually present the given information in the structured format of textbook examples. Solvers must first reconstruct the physical model from natural-language descriptions and diagrams, then determine the appropriate approximations, choose a coordinate system, formulate conservation relations, and finally complete the symbolic derivations and numerical calculations.

This is more like taking over a complex engineering project that comes with only a requirements document and an architectural sketch, but no test cases, than solving an algorithm problem with a clearly defined interface.

Consider the multiple-reflection optics problem. The model must not only “see” the diagram, but also distinguish how many reflections each ray undergoes, which paths are constrained by the aperture, and how the vignetting boundary changes with the parameters. If the visual analysis stage misidentifies a single intersection or angular relationship, even a perfectly executed Taylor expansion afterward will still produce a meaningless final answer.

The magnetic refrigeration problem tests a different set of abilities. The model must connect thermodynamic relations, cyclic processes, and temporal evolution into a single causal chain. This involves both conceptual reasoning and operations with partial derivatives and integrals. A common problem for large language models is not that they “do not know the Maxwell relations,” but that they quietly swap the variables being held constant during a long derivation, drop a minus sign, or become inconsistent about the integration bounds.

The 28.6-point score therefore indicates, at minimum, that the Xiaoyi system used in the assessment went well beyond simple knowledge retrieval. It had to maintain a high degree of consistency across diagram understanding, model construction, long-horizon reasoning, and precise computation.

The System, Not Just the Model, Is What Really Matters

Huawei attributes this performance to Xiaoyi’s agentic system. Based on the information disclosed so far, its capabilities can broadly be divided into three stages: multimodal visual perception reads diagrams and spatial constraints; the reasoning module constructs the solution step by step and backtracks when it detects contradictions; and symbolic computation handles exact algebraic transformations and formula derivations.

There is nothing mysterious about this approach, but it is highly practical.

A large language model generating a response in a single pass is like a student with broad knowledge who writes quickly but rarely checks their scratch work. An agentic system adds roles for problem interpretation, verification, and review: one module ensures that the diagram is read correctly, another produces an initial solution, and a third checks dimensional consistency, boundary conditions, and logical closure. If the solution fails these checks, the errors are fed back to the solving module for another attempt.

Over the past year, the focus of competition in complex scientific reasoning has shifted from “whose model can guess correctly on the first try” to “who can build a stable generate–verify–correct loop.” SU-01, previously unveiled by the Shanghai AI Laboratory, as well as multi-agent research aimed at physics competitions, has emphasized a similar path: specialized training teaches models the structure of proofs, while verifiers and iterative correction improve reliability during inference.

This means Xiaoyi’s result may not be attributable to a foundation model suddenly acquiring near-perfect physical intuition. A more likely explanation is that the foundation model, vision module, symbolic tools, search strategy, and verification workflow collectively form a problem-solving system designed for scientific tasks.

For developers, this is actually more instructive than another increase in parameter count. Complex production tasks should never place all their reliability on a single model sample. Coding agents need compilers and tests, data agents need validation against SQL execution results, and scientific reasoning agents likewise need dimensional checks, symbolic verification, and independent review.

28.6 Is an Impressive Score, but a Complete Evaluation Card Is Still Missing

It is important to emphasize that the information currently available comes mainly from Huawei’s official disclosures and media reports. The result deserves attention, but developers and researchers need more evaluation details to reproduce it accurately and compare it with other systems.

At least the following questions remain unanswered:

What was the input format? Did the system receive the original PDF, scanned images, or manually prepared text and formulas? Were the diagrams preserved in their original form?
Which external tools were used? Did it call a computer algebra system, a code execution environment, search tools, or a specialized physics solver? Using tools is not inherently “cheating,” but they must be documented in the evaluation methodology.
How many sampling and correction attempts were allowed? Was 28.6 the result of a single end-to-end run, or was the best answer selected from multiple generations? The two scenarios imply entirely different levels of system reliability.
Who performed the grading? Was each subproblem graded according to the official rubric? Were the graders independent, and how were disputed steps reviewed?
What were the cost and latency? The number of inference tokens consumed, the total runtime, and the number of parallel agents used for a single exam all affect whether the technology can be deployed in real products.
Was there any data contamination? If the problems or solutions had entered the training data, retrieval pipeline, or cache before the evaluation, the score would not be equivalent to generalization on entirely novel problems.

Huawei has stated that Xiaoyi used the same exam as the contestants and was graded according to the official evaluation criteria. That is an important foundation. However, “the same exam” does not mean “identical testing conditions.” Human contestants face limits on time, tools, and the on-site environment, while an AI system may have access to longer reasoning time, parallel sampling, and symbolic tools. The quality of their solutions can be compared, but their overall competitive capabilities should not be equated directly.

Our assessment is therefore: 28.6 points represents a substantial demonstration of system capability, but it is not yet a sufficiently transparent and directly reproducible scientific benchmark report. If Huawei subsequently publishes the answers to each problem, the inference configuration, the list of tools, and the grading records, the result will carry considerably more industry value than a one-off corporate announcement.

From the IMO to the IPhO, AI Testing Is Entering the Systems Engineering Stage

Academic olympiads are becoming a favored arena for large-model vendors to demonstrate their capabilities. On the one hand, the problems are highly difficult and the grading standards are relatively clear, preventing the rapid saturation seen in ordinary question-answering benchmarks. On the other hand, mathematical and physical reasoning transfer strongly to coding, scientific research, and engineering design, making them more representative of high-value tasks than measures of “how much encyclopedic knowledge a model has memorized.”

However, there are still clear differences between mathematics and physics olympiads. Mathematics problems are largely self-contained within abstract symbolic systems, whereas physics problems first require real-world descriptions to be translated into models. Diagrams, units, approximations, boundary conditions, and experimental intuition all become part of the reasoning chain. Even a model with strong algebraic capabilities may produce an answer that is elegant in form but physically absurd because of a modeling error.

The significance of Xiaoyi’s result lies in crossing this barrier: it needed not only to calculate, but first to determine “what to calculate.”

Nevertheless, competition problems are ultimately closed problems that have been carefully designed and generally provide sufficiently complete conditions. Real scientific research often has no unique answer; data may be noisy, assumptions may fail, and the problem definition itself may need repeated revision. The ability to solve difficult olympiad problems does not mean that AI can autonomously discover new laws of physics, much less replace scientists.

It is like a coding model passing a highly difficult programming competition: that demonstrates powerful local reasoning and implementation abilities, but does not prove that it can independently maintain a large production system for ten years.

What Does This Mean for Ordinary Users?

Huawei says it will gradually integrate the core technical architecture demonstrated here into Xiaoyi’s broader capabilities, providing features such as formula derivation, theoretical verification, chart analysis, and personalized guidance in educational scenarios.

This is a sensible deployment path. Rather than simply having AI provide answers, a more valuable product would put verification capabilities in users’ hands:

Recognize formulas in handwriting or images and identify missing variable definitions and conditions;
Check each step of a derivation rather than merely judging the final numerical result;
Cross-check conclusions through dimensional analysis, limiting cases, and numerical simulations;
Generate hints based on where a student made a mistake instead of immediately displaying the complete answer;
In engineering scenarios, convert natural-language requirements into computable models while preserving a record of assumptions.

In June this year, Xiaoyi Claw integrated the open-source Pangu model openPangu 2.0 Pro, with a focus on improving system-level task execution in HarmonyOS. Huawei says the model has been optimized for Ascend computing hardware and can achieve single-card throughput up to twice that of some mainstream open-source models in the industry. Combining this adaptation to local computing hardware with agent execution and scientific reasoning chains may have greater commercial value than merely setting a new exam benchmark. Xiaoyi already spans devices including smartphones, tablets, PCs, vehicles, and smart screens, allowing it to turn reasoning results into system actions.

Of course, a near-perfect olympiad solution is still far removed from reliably serving hundreds of millions of device users. Hallucination rates, latency, cost, privacy, and fault-tolerance mechanisms all remain challenges. Competition results demonstrate the upper bound of capability; product experience tests the lower bound. The latter is often more difficult.

Conclusion

Xiaoyi’s score of 28.6 on the IPhO 2026 theoretical exam is a substantial demonstration of Chinese AI capabilities in complex scientific reasoning. It shows that a combination of multimodal understanding, symbolic computation, and iterative verification can already handle long-chain physics problems once thought solvable only by elite students.

However, whether this result becomes an industry benchmark will be determined not by the slogan “surpassing the gold-medal threshold,” but by whether sufficiently complete evaluation conditions are disclosed and whether the relevant capabilities can be transformed from a costly one-off demonstration into derivation, verification, and engineering tools that users can rely on continuously.

Olympiad results show that AI is becoming increasingly capable of solving difficult problems. The next stage is to prove whether it can reliably recognize when it has solved them incorrectly.