Qwen2.5-Math: CPMöbius RL, Geometry & Libraries

• CPMöbius is a multifaceted framework uniting data-free reinforcement learning for LLM math reasoning, Möbius-invariant differential geometry, and computational libraries.
• In reinforcement learning, it employs a cooperative Coach–Player loop with adaptive task generation, yielding up to +4.9% accuracy gains on models like Qwen2.5-Math-7B-Instruct.
• Its geometric applications range from solving the conformal-to-Einstein equation in 2D to enabling efficient symbolic-numeric operations in non-Euclidean spaces.

CPMöbius (often rendered as CPMobius) denotes several advanced concepts and frameworks spanning machine learning, differential geometry, and computational geometry. These include (1) a data-free reinforcement learning paradigm for LLMs, (2) the conformal-to-Einstein operator on Möbius surfaces in two-dimensional conformal geometry, and (3) a C++ (plus Python/GUI) library suite for symbolic and numerical computations in Möbius-invariant (non-Euclidean) geometry. Each instance is described with reference to its primary formulation, technical details, and domain-specific significance.

1. Data-Free Reinforcement Learning via the CPMöbius Paradigm

The CPMöbius method (Li et al., 3 Feb 2026) introduces a collaborative, fully data-free reinforcement learning framework designed for improving the mathematical reasoning ability of LLMs. Drawing inspiration from real-world coach–player dynamics, CPMöbius replaces adversarial self-play with a two-agent cooperative optimization loop, enabling scalable curriculum learning without reliance on external datasets.

Architecture and Algorithmic Workflow

CPMöbius orchestrates an iterative four-stage loop between a Coach and a Player:

1. Coach Task Generation:
   • The Coach policy $\pi_C$ samples $m$ candidate task instructions $x_i$.
   • Each $x_i$ is filtered by Player rollout accuracy: accepted only if accuracy lies in $[0.2,0.8]$ (ensures neither trivial nor impossible tasks).
2. Player Execution:
   • For each instruction, the Player policy $\pi_P$ generates $n$ responses $y_{i,j}$.
   • Pseudo-label $y_i^*$ is the majority of the $y_{i,j}$.
   • Binary rewards assigned: $r_{i,j}=1[y_{i,j}=y_i^*]$.
   • $\pi_P$ is updated via group relative policy optimization (GRPO).
3. Player Evaluation:
   • Post-update, $\pi_P$’s accuracy on a held-out validation set is measured.
   • Performance gain $$A_t = \mathrm{Acc}(\pi_P^{(t+1)}, D_{\mathrm{val}}) - \mathrm{Acc}(\pi_P^{(t)}, D_{\mathrm{val}})$$.
4. Coach Update:
   • Each $x_i$ receives reward $R_C(x_i) = R_P(x_i) \cdot A_t$, where $R_P(x_i) = \frac{1}{n}\sum_j r_{i,j}$.
   • Coach is optimized via REINFORCE on instruction-level rewards.

The full pseudocode is detailed in (Li et al., 3 Feb 2026), Appendix A.1. This design ensures that the curriculum continually targets the Player’s evolving “zone of proximal development,” while Coach learning is directly incentivized by Player improvement, not mere task difficulty.

Reinforcement Learning Formulation

• Player:

Employs GRPO, where roll-out rewards $r_{i,j}$ are normalized for each instruction, and the loss includes variance clipping and KL regularization:

$$\mathcal{L}_{\mathrm{GRPO}}(\pi_P) = \mathbb{E} \left[ \frac{1}{n} \sum_{j=1}^{n} \min(r_{i,j}A_{i,j}, \mathrm{clip}(r_{i,j}, 1-\epsilon, 1+\epsilon)A_{i,j}) \right] - \beta D_{\mathrm{KL}}(\pi_P\|\pi_\mathrm{ref})$$

• Coach:

Standard REINFORCE updates are performed, where the policy gradient is:

$$\nabla_{\theta_C} J(\pi_C) = \frac{1}{m} \sum_{i=1}^m R_C(x_i) \nabla_{\theta_C} \log \pi_C(x_i)$$

Empirical Results

On the Qwen2.5-Math-7B-Instruct model, CPMöbius achieves:

Model/Approach	Overall Avg	OOD Avg
Base	35.8%	37.4%
R-Zero (adversarial)	36.9%	38.1%
RENT (entropy min RL)	39.2%	38.8%
CPMöbius	40.7%	38.3%

Compared to established baselines, CPMöbius yields a +4.9% absolute accuracy gain overall and +5.4% on out-of-distribution benchmarks (Li et al., 3 Feb 2026). Similar robust improvements were observed across multiple Player architectures.

Ablation Analyses and Limitations

• With a frozen Coach, overall accuracy drops 3.5 points and OOD by 3.7.
• Without Coach warm-up, performance degrades to near base-level.
• Omitting difficulty filtering also yields substantial performance drops.

The efficacy of CPMöbius is thus tightly coupled to adaptive Coach training and zone-specific curriculum. Limitations include current restriction to mathematical reasoning and dependency on a well-initialized Coach. Prospective work addresses Coach auto-initialization and generalization to multi-modal or open-ended tasks.

2. The Conformal-to-Einstein Operator on Möbius Surfaces

CPMöbius also refers to the conformal-to-Einstein operator studied in two-dimensional Möbius geometry (Randall, 2013). This operator connects conformal geometry, overdetermined PDE systems, and tractor bundle theory.

Möbius Structures and Tractor Bundles

• A Möbius surface $(M^2, [g], P)$ is a two-manifold equipped with a conformal structure $[g]$ and a symmetric tensor $P_{ab}$ ("Rho tensor") with specific transformation laws under conformal rescalings.
• The standard tractor bundle $$\mathcal{E}^A \cong E[1] \oplus E_a[1] \oplus E[-1]$$ admits a canonical metric and connection, providing the correct setting for Möbius-invariant differential geometry.

Conformal-to-Einstein Equation and Prolongation

The operator $D_{ab}: E[1] \to E_{(ab)_0}[1]$ acts as:

$$(D_{ab} - \nabla_{(a}\nabla_{b)}) \sigma = P_{ab} \sigma$$

The conformal-to-Einstein equation is:

$$(\nabla_{(a}\nabla_{b)}\sigma + P_{ab}\,\sigma)_0 = 0.$$

Its prolonged system for a scale $\sigma$ and derived quantities $H_a = \nabla_a\sigma$, $A$ (trace component) is:

$$\begin{aligned} \nabla_a \sigma &= H_a, \ \nabla_a H_b &= -P_{ab}\sigma - g_{ab}A, \ \nabla_a A &= -Y_a \sigma + P_a{}^b H_b, \end{aligned}$$

where $Y_{abc}$ is the Cotton–York tensor. Critically, in $n=2$, the prolongation includes $-Y_a \sigma$, a term absent in higher dimensions, and requires modifying the tractor connection to the prolongation connection.

Parallel Tractors and Solution Space

There is a one-to-one correspondence between (i) nowhere-zero solutions $\sigma$ of the conformal-to-Einstein equation above, and (ii) parallel sections of the prolonged tractor connection with nonvanishing scale component.

Local Obstructions and Classification

Through differentiation, algebraic constraints (invariants in curvature, $P_{ab}$, and Cotton–York tensor) determine existence and dimension of the solution space:

• For flat Möbius structures ($Y_a\equiv0$): $\dim \ker D = 4$.
• For generic non-flat structures: $\dim \ker D = 1$.
• For special non-generic ODE-reducible cases: $\dim \ker D = 2$. No intermediate kernel dimensions occur.

3. CPMöbius/“MoebInv”: Computational Geometry Libraries

MoebInv (Kisil, 2019), referenced here as CPMöbius in a computational context, comprises “cycle” and “figure” C++ libraries (with Python and GUI frontends) for symbolic and numerical analysis in Möbius-invariant, non-Euclidean geometries.

Library Architecture and API

• cycle: Manages single cycles (quadratic forms in $\mathbb{R}^{p,q}$), Clifford-algebra embeddings, transformations, and invariants (orthogonality, tangency, normalization).
• figure: Handles ensembles of cycles linked by Möbius-invariant relations, automated solving (including Apollonius-type problems), and generation tracking. Supports multi-parameter symbolic and numeric computations.

Key class/methods include:

• Cycle(k, ℓ, m), CycleSpace(p, q)
• addCycle, getCycle, addRelation, checkRelation, with support for relation types such as Orthogonal, Tangent, Incidence.

Mathematical Foundations

• A cycle in signature $(p, q)$ is defined by:

$$k\,\langle x,x \rangle_{p,q} + 2\,\langle \ell, x \rangle_{p,q} + m = 0$$

• Fillmore–Springer–Cnops construction (FSCc) encodes cycles as $2 \times 2$ Clifford matrices; Möbius transformations act by conjugation.
• Invariants and relations are determined by a single Clifford inner product:
  • Orthogonality: $\langle C_1, C_2 \rangle = 0$
  • Tangency: $$|\langle C_1, C_2 \rangle|^2 = \langle C_1, C_1 \rangle \langle C_2, C_2 \rangle$$
  • Incidence (points, lines, circles): via self-orthogonality.

GUI and Jupyter Integration

• Provides a Qt5-based standalone GUI, as well as Jupyter/Python bindings.
• Interactive manipulation: add cycles by relations, inspect equations graphically, real-time Asymptote/SVG rendering.
• Automated solution workflow for complex hierarchical geometric configurations.

Performance and Use Cases

• Numeric cycle operations execute in $\leq100\,\mu$s, symbolic solves for a quadratic plus several linears in 20–200 ms.
• Best suited for 2D/3D, moderate parameter symbolic problems.
• Typical application: automated generation and analysis of Apollonian circle or sphere packings.

4. Technical Interrelations, Distinctions, and Naming

The name CPMöbius appears in disparate advanced research contexts connected only through notational similarity and Möbius invariance:

• RL/LLM context: CPMöbius is the name of an iterative cooperative curriculum approach for LLMs (Li et al., 3 Feb 2026).
• Differential geometry: CPMöbius designates the conformal-to-Einstein operator on Möbius surfaces (Randall, 2013).
• Computational geometry: CPMöbius (via MoebInv) encodes a software suite for Möbius-invariant constructions (Kisil, 2019).

This suggests the name is chosen for its association with symmetry, invariance, or recursive/cooperative structures inspired by the Möbius strip or group actions. There is no substantive methodological relation among the three domains.

5. Impact, Limitations, and Future Directions

In Machine Learning:

CPMöbius represents the first fully data-free, cooperative RL framework for LLM reasoning, outperforming prior methods in accuracy and curriculum stability. Its dependence on Coach warm-up and focus on mathematical reasoning are identified limitations. Extension to unsupervised Coach initialization, broader domains, and deeper investigation into two-agent co-evolution are cited as open problems (Li et al., 3 Feb 2026).

In Differential Geometry:

The CPMöbius operator provides a complete Möbius-invariant framework for understanding the conformal-to-Einstein equation in dimension two, yielding new classification theorems for kernel dimensions and explicit local curvature obstructions (Randall, 2013).

In Computational Geometry:

The MoebInv/CPMöbius libraries enable rigorous, scalable symbolic-numeric exploration of Möbius-invariant relations with applications in geometric modeling, pedagogy, and research. Limiting factors include symbolic solve complexity for multi-parameter families and computational challenges in higher dimensions (Kisil, 2019).

6. Summary Table: CPMöbius Across Contexts

Domain	Primary Contribution	Representative Reference
RL/LLM curriculum	Data-free Coach-Player iterative RL	(Li et al., 3 Feb 2026)
Möbius surface geometry	Conformal-to-Einstein operator, obstructions	(Randall, 2013)
Computational geometry	Symbolic+numeric Möbius-invariant libraries	(Kisil, 2019)

Each interpretation of CPMöbius advances its respective field by leveraging Möbius-invariant principles in the design of algorithms, geometric structures, or computational frameworks.