Chinese Physics Olympiad (CPhO)

Updated 20 November 2025

CPhO is a pre-university physics competition defined by complex multi-stage problems and rigorous quantitative modeling.
The contest integrates advanced concepts like regime analysis, diagram interpretation, and precise symbolic reasoning for national and international benchmarks.
It serves as a critical testbed for AI systems, enabling structured solution decomposition and sequential error detection to evaluate scientific reasoning.

The Chinese Physics Olympiad (CPhO) is a premier pre-university physics competition characterized by exceptionally challenging problems that demand not only mastery of advanced physics concepts but also multi-stage symbolic reasoning, interpretation of technical diagrams, and precise quantitative modeling. The CPhO serves as both a national selection mechanism for the International Physics Olympiad and an independent testbed for the development and benchmarking of advanced scientific reasoning methodologies, including AI systems. As of 2025, the CPhO has emerged as the canonical evaluation environment for AI frameworks targeting expert-level performance in multi-step physics problem solving (Li et al., 11 Nov 2025, Jian et al., 13 Nov 2025).

1. Structural Features and Problem Typology

The CPhO theory exam consists of a set of seven multi-part problems, each with a high point value (individual problems weighted up to 50 points), for a total of 320 points. The problem set covers the full breadth of high school and introductory undergraduate physics, including, but not limited to, electrostatics, oscillations and waves, classical mechanics, thermodynamics, electromagnetism, modern physics, and applied topics such as plasma physics and astrophysics. Standard components include:

Complex multi-stage modeling: Problems commonly require decomposing the system into regimes, formulating governing equations, and synthesizing stepwise derivations.
Integration of diagrams and data: Many problems demand the extraction of quantitative or topological information from provided figures or schematics for symbolic or numerical work.
Regime analysis and case branching: Tasks frequently call for the detection and distinct treatment of different physical regimes (e.g., underdamped/overdamped, high/low field, relativistic/non-relativistic limits).
High-precision computation and symbolic algebra: Both intermediate symbolic manipulations and numeric verifications are necessary for full credit.

2. Recent AI Benchmarks on the CPhO

The CPhO has become the principal benchmark for AI systems seeking generalistic scientific reasoning in physics. Two major frameworks have established state-of-the-art results:

System	Total Score (out of 320)	Error Rate	Notable Features
LOCA-R	313	2.2%	Atomic stepwise review, structured decomposition, dedicated problem interpretation (Jian et al., 13 Nov 2025)
SciAgent	264	17.5%	Hierarchical multi-agent system, adaptive sub-agent orchestration, reviewer verification (Li et al., 11 Nov 2025)
Human gold-medalist	199	37.8%	Official highest human score in 2025 exam (Jian et al., 13 Nov 2025)

LOCA-R achieves near-perfect performance, surpassing both baseline LLM prompting (error rate 12–8.8%) and the top human gold-medalist (204 points; 36% error) by substantial margins (Jian et al., 13 Nov 2025).

3. AI Methodological Advances Driven by the CPhO

Recent approaches targeting the CPhO emphasize rigorous decomposition and verification tailored to the exam's features:

LOCA-R Framework: Implements a pipeline with (1) problem interpretation to extract structure and symbols, (2) logical chain augmentation to perform stepwise decomposition ensuring atomic reasoning acts, and (3) atomic & sequential review that evaluates each inductive move with targeted feedback. By representing solutions as structured tuples $(P_j, D_j)$ , separating principles from their explicit derivations, LOCA-R eliminates latent misapplications of concepts and numerical errors. Notably, the review step is fully sequential, and feedback loops continue until no errors persist. This atomic validation is essential for detecting sign errors, boundary condition misapplications, and unit inconsistencies that previously prevented AI or human solvers from reaching near-perfect scores (Jian et al., 13 Nov 2025).
SciAgent Multi-Agent System: Organizes reasoning via a three-tiered architecture: the Coordinator Agent (parses domain and complexity), Worker Systems (specialized for the Olympiad setting, employing the ReAct framework), and Sub-agents for generation, image analysis, review, and summarization. The Reviewer Agent in particular enforces consistency by checking dimension analysis, sign conventions, and correcting logic errors, as illustrated in cases where naive LLMs (e.g., Gemini Pro 2.5) produced incorrect linewidth estimates in quantum physics problems (Li et al., 11 Nov 2025).

4. Illustrative Problem Decomposition

The following stylized exemplars mirror modern CPhO questions and highlight how recent AI systems parse and solve them:

Electrostatics in Layered Media: Given a point charge in a dielectric sphere (permittivity $\varepsilon_1$ ) surrounded by another dielectric ( $\varepsilon_2$ ), the pipeline identifies regions, applies boundary conditions for potential and electric displacement continuity at interfaces, proposes ansatz solutions, solves for coefficients, and computes surface charge densities. Symbolic manipulations are verified at each algebraic step to avoid sign or factor-2 discrepancies (Li et al., 11 Nov 2025).
RLC Oscillations with Damping: For an RLC circuit, required steps include constructing a symbolic circuit diagram, formulating and classifying the differential equation by discriminant $\Delta=R^2-4L/C$ into underdamped, overdamped, or critically damped regimes, finding closed-form solutions for transient currents $i(t)$ , and expressing the quality factor as $Q = (1/R)\sqrt{L/C}$ , with dimensions and limits checked by a reviewer agent (Li et al., 11 Nov 2025).
Astrophysical and Plasma Physics: Advanced problems such as the Yarkovsky effect (thermal recoil on asteroids) and Debye screening in plasmas are solved via sequential decomposition into principle-derivation tuples, explicit symbolic manipulation, and atomic review that validates every assumption and intermediate result (Jian et al., 13 Nov 2025).

5. Formal Mathematical Pipeline and Error Metrics

Quantitative assessment and error tracing are central to recent research on CPhO-aligned AI:

Atomic Chain Decomposition: Solutions constructed as $S_{\mathrm{aug}}=((P_1, D_1),\ldots,(P_m, D_m))$ enforce single-reasoning-act steps, isolating potential sources of error (Jian et al., 13 Nov 2025).
Sequential Error Detection: For each step $j$ , review yields $(v_j, f_j)=\mathcal{R}(s'_j|C_{j-1})$ where $v_j \in \{\mathrm{Correct}, \mathrm{Wrong}\}$ with $f_j$ as feedback. Global solution validity $V$ and cumulative feedback $F$ are compiled for targeted revision.
Scoring: The CPhO utilizes a fine-grained rubric that assigns partial credit to atomic sub-steps, not merely final answers. The error rate is defined as $\frac{320-\mathrm{Score}}{320}\times 100\%$ (Jian et al., 13 Nov 2025).

6. Impact on AI and Scientific Reasoning Research

The CPhO’s complexity and scoring rigor have catalyzed the development of domain-specialized, interpretable reasoning algorithms. The structured problem decomposition and multi-level review inherent in LOCA-R and SciAgent have enabled detection and correction of subtle conceptual flaws and arithmetic errors, revealing the limitations of direct prompting or generic chain-of-thought strategies in complex physics. A plausible implication is that frameworks devised for the CPhO can generalize to other STEM domains where layered reasoning and error localization are critical (Li et al., 11 Nov 2025, Jian et al., 13 Nov 2025).

7. Limitations, Open Problems, and Future Directions

While current systems such as LOCA-R nearly close the gap to perfect scores, the challenge of persistent errors in heavy arithmetic or when synthesizing multi-part constraints remains. Expansion to non-physics STEM domains will require embedding domain-tailored principles. Computational cost, especially with multiple review rounds and complex problem parsing, poses additional constraints. There are prospective directions in symbolic and numeric tool integration (e.g., computer algebra systems), adaptive review budgeting for high-risk subproblems, and cross-problem consistency mechanisms to eliminate global sign or unit mismatches. The CPhO thereby continues as a driving benchmark for progress in generalistic scientific intelligence (Jian et al., 13 Nov 2025).