VAR-MATH: Mathematical Reasoning & Risk Metrics

• VAR-MATH is a dual-purpose framework that benchmarks large language models’ mathematical reasoning using symbolic multi-instance evaluations and computes risk quantiles efficiently.
• Empirical studies reveal substantial accuracy drops when models face contamination-resistant, parameterized benchmarks, highlighting limitations of superficial pattern-based learning.
• The framework employs FFT-based transform techniques to compute Value-at-Risk and Conditional Value-at-Risk with high accuracy and remarkable computational speed.

VAR-MATH refers to a symbolic multi-instance evaluation methodology designed to rigorously probe mathematical reasoning capabilities in LLMs, as well as to a class of numerical and optimization techniques for computing Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR) using characteristic function inversion. This article synthesizes the evaluation paradigm introduced in "VAR-MATH: Probing True Mathematical Reasoning in LLMs via Symbolic Multi-Instance Benchmarks" (Yao et al., 17 Jul 2025) and the computational framework presented in "On a Transform Method for the Efficient Computation of Conditional VaR (and VaR) with Application to Loss Models with Jumps and Stochastic Volatility" (Ramponi, 2014), detailing its theoretical foundation, algorithmic implementation, and implications for both mathematical AI benchmarking and risk quantile computation.

1. Symbolic Multi-Instance Benchmarking for Mathematical Reasoning

VAR-MATH provides a systematic approach for constructing contamination-resistant and robust benchmarks in mathematical problem solving. Traditional evaluation, such as on AMC/AIME competition problems, suffers from benchmark contamination—publicly posted test instances leak into LLM pre-training data—and evaluation fragility—single-instance scoring is highly sensitive to stochastic output variance and superficial pattern matching. VAR-MATH addresses these via:

• Symbolic template abstraction: For each fixed numerical problem, the key constants are lifted to symbolic variables (e.g., turning $3x + 5 = 11$ into $a x + b = c$ with $a, b, c$ as sampled variables).
• Feasible domain instantiation: Domains $\mathcal{D}_{v_i}$ for each variable are specified, ensuring logical structure is preserved.
• Multi-instance protocol: Each symbolic template generates up to $N=5$ random instantiations over the cross-product of variable domains, requiring models to solve all instances correctly for success.
• Automated answer checking: The ground-truth is computed for each instantiation by a closed-form parametric solution.

This reframing enforces true reasoning over parameterized problem families, sharply reducing the possibility of memorization-driven “shortcut” success and variance-induced evaluation error.

2. Empirical Evaluation and Quantitative Impact

Application of VAR-MATH to AMC23 and AIME24 yielded symbolic benchmarks VAR-AMC23 and VAR-AIME24. Experimental campaigns spanning multiple RL-augmented open-source and frontier models demonstrated:

Model Group	AMC23 Accuracy	VAR-AMC23 Accuracy	AIME24 Accuracy	VAR-AIME24 Accuracy
7B RL avg	59.1%	30.7%	27.8%	11.6%

Performance drops average 48%–58%, with similar degradation observed in 32B models and nonzero decline even in frontier models. This strongly indicates that prior RL fine-tuning often induces overfitting to public test set surface statistics, not genuine mathematical generalization.

3. Principles of Efficient VaR and CVaR Computation

In risk quantification, VAR-MATH also refers to a set of fast transform-based algorithms for computing quantiles and conditional quantiles from models admitting explicit characteristic functions, such as Lévy and stochastic volatility models (Ramponi, 2014). The central objects are:

• VaR at level $\alpha$: $$\operatorname{VaR}_\alpha(L) = \inf\{x \mid \mathbb{P}(L \le x) \ge \alpha\}$$
• CVaR (Average-VaR) at level $\alpha$: $$\operatorname{CVaR}_\alpha(L) = \inf_x \left\{ x + \frac{1}{1-\alpha} \mathbb{E}[(L-x)^+] \right\}$$

For $L$ admitting characteristic function $\phi_L(u)$, the stop-loss expectation admits a Fourier integral representation: $$\mathbb{E}[(L-x)^+] = -\frac{1}{2\pi} \int_{-\infty}^{\infty} \frac{e^{-iux} \phi_L(u)}{u^2} du,$$ which can be efficiently evaluated (after damping and quadrature) using Fast Fourier Transform (FFT) or Fractional FFT (FRFT).

4. Optimization Structure and Algorithmic Details

The VaR and CVaR calculation reduces to a one-dimensional convex minimization:

• Evaluate the convex function $$G_\alpha(x)=x+\frac{1}{1-\alpha}\mathbb{E}[(L-x)^+]$$ over an equispaced grid via FFT, leveraging the Nyquist relation between frequency and space for vectorized computation.
• Minimize $G_\alpha(x)$ to extract CVaR; VaR is obtained as the minimal $x$ at which $G_\alpha(x)$ achieves its infimum.

Complexity is $\mathcal{O}(N \log N)$ per quantile, $N$ being the grid size. Error sources include truncation in the frequency integral, discretization, and periodic wrap-around in FFT.

5. Practical Applications: Jump-Diffusion Illustration

For Merton jump-diffusion loss models, the characteristic function is

$$\phi_L(u) = \exp \left( i(\mu-\frac{1}{2}\sigma^2)T u - \frac{1}{2}\sigma^2 T u^2 + \lambda T \left(e^{i a u-\frac{1}{2}b^2u^2}-1\right) \right),$$

enabling direct transform-based VaR/CVaR computation. Numerical experiments exhibit sub-millisecond run times with accuracy better than $0.001$ on high-confidence quantiles, outperforming brute-force Monte Carlo by orders of magnitude.

6. VAR-MATH in the Context of Tail Risk and Robust Estimation

VAR-MATH as a framework is neutral with respect to the risk factor dynamics; it encompasses Lévy, stochastic volatility, regime-switching, and any loss model admitting explicit or numerically tractable characteristic functions. Its relevance is further amplified in tail-risk contexts, where model misspecification, moment constraints, and importance sampling techniques converge. Key properties include:

• Rigorous convex/programmatic minimization for quantile estimation.
• Extension to discrete moment matching for robust quantile bracketing under model uncertainty.
• Applicability to loss distributions and risk pricing under heavy-tailed regimes.

7. Conclusions and Future Implications

VAR-MATH in the context of symbolic multi-instance reasoning fundamentally reshapes how AI mathematical competence is measured: shifting emphasis from isolated benchmark accuracy toward robust, contamination-resistant generalization across parameterized problem families. In risk quantile computation, transform-based VAR-MATH methods enable rapid, high-accuracy assessment for complex loss models irrespective of underlying factor dynamics, including stochastic volatility and jumps. These methodologies collectively mark a significant advancement in both mathematical AI benchmarking and risk management computation, suggesting new standards for both theoretical rigor and computational efficiency (Yao et al., 17 Jul 2025, Ramponi, 2014).