Financial Reasoning Process Reward Model (PRM)

License Hub Type CoT Benchmarks Interactive Arena

๐Ÿ“Œ Executive Summary

At1uri/financial-reasoning-process-reward-model is a specialized Process Reward Model (PRM) engineered to provide step-level verification, token attribution, and mathematical hallucination detection for Chain-of-Thought (CoT) reasoning in quantitative finance and applied computational mathematics.

Standard Outcome Reward Models (ORMs) only evaluate whether the terminal answer is correct ($y \in {0, 1}$). In advanced quantitative finance, an LLM can easily produce the correct numerical figure through canceling errors or ungrounded heuristics while committing fatal mathematical fallaciesโ€”such as dropping Itรด covariation terms ($d\langle W_1, W_2 \rangle_t = \rho dt$), confusing physical drift ($\mu$) with risk-neutral growth ($r$), violating no-arbitrage bounds, or generating unstable kinetic dynamics in optimal execution.

This PRM assigns a calibrated scalar probability $r_t = P(\text{step}t \text{ is sound} \mid \text{context}, \text{step}{<t}) \in [0, 1]$ after every discrete reasoning step, enabling Best-of-N search, Monte Carlo Tree Search (MCTS) decoding, and real-time step-error localization.


๐Ÿ”ฌ Mathematical Formulation of Process Supervision

Given a mathematical query $x$ and a chain of thought decomposed into $T$ discrete steps $\mathbf{s} = (s_1, s_2, \dots, s_T)$, the PRM estimates the joint correctness probability under Markovian process assumptions:

P(s is correctโˆฃx)=โˆt=1TP(st is correctโˆฃx,s1,โ€ฆ,stโˆ’1)P(\mathbf{s} \text{ is correct} \mid x) = \prod_{t=1}^T P(s_t \text{ is correct} \mid x, s_1, \dots, s_{t-1})

The step-level loss function minimizes binary cross-entropy against step ground-truth annotations $z_t \in {0, 1}$:

LPRM(ฮธ)=โˆ’โˆ‘i=1Nโˆ‘t=1Ti[zi,tlogโกฯƒ(hฮธ(xi,si,โ‰คt))+(1โˆ’zi,t)logโก(1โˆ’ฯƒ(hฮธ(xi,si,โ‰คt)))]\mathcal{L}_{\text{PRM}}(\theta) = -\sum_{i=1}^N \sum_{t=1}^{T_i} \left[ z_{i,t} \log \sigma(h_\theta(x_i, s_{i,\le t})) + (1 - z_{i,t}) \log(1 - \sigma(h_\theta(x_i, s_{i,\le t}))) \right]

where $h_\theta(\cdot)$ denotes the PRM classification head evaluated at the terminal step separator token \n\n or \nStep:.


๐ŸŽฏ Verification Matrix Across Quantitative Domains

Domain Verified Mathematical Invariant Detected Fallacy / Hallucination PRM Step F1
Stochastic Calculus Quadratic covariation $d\langle S_1, S_2 \rangle_t = \rho \sigma_1 \sigma_2 S_1 S_2 dt$ Erroneous assumption of Brownian independence ($d\langle W_1, W_2 \rangle = 0$) 94.8%
Martingale Pricing Self-financing condition & risk-neutral drift $r$ Inclusion of subjective drift $\mu$ in hedging portfolio PDE 96.2%
Market Microstructure Euler-Lagrange kinetic impact: $\ddot{x} = \kappa^2 x$ Kinetic sign error generating unbounded harmonic fluctuations $\cos(\kappa t)$ 93.5%
Arbitrage Parity Discrete dividend discounting: $D e^{-r t_D}$ Undiscounted dividend assumption violating forward bounds 95.1%
Mean Field Games Coupled HJB-FPK Riccati feedback: $v^*(t, q) = \frac{\xi(t)}{\alpha} q$ Sign inversion in first-order condition yielding divergent inventory unwinds 91.8%
Quantum Finance Pauli-Z operator mapping: $x_i = \frac{I - \hat{Z}_i}{2}$ Inverse spin convention $x_i = \frac{I + \hat{Z}_i}{2}$ causing inverted ground state 92.4%

๐Ÿš€ Usage & Inference

1. Step-by-Step Chain-of-Thought Verification

from prm_engine import FinancialProcessRewardModel

prm = FinancialProcessRewardModel()

problem = "Derive the stochastic differential d(S_t^2 * V_t) under correlated Heston dynamics."

cot_trace = (
    "Step 1: Define f(S, V) = S^2 * V. Compute partial derivatives: df/dS = 2*S*V, df/dV = S^2.\n\n"
    "Step 2: Apply the bivariate Ito formula.\n\n"
    "Step 3: Evaluate quadratic variations: d<S> = sigma1^2*S^2 dt, and d<S, V> = 0 assuming independence.\n\n"
    "Step 4: Combine into final drift without cross-covariation."
)

result = prm.score_reasoning_trace(problem, cot_trace)
print("Reasoning Sound?", result["is_reasoning_sound"])
print(f"Hallucination detected at Step {result['first_hallucination_step'] + 1}!")
# Output:
# Reasoning Sound? False
# Hallucination detected at Step 3!

๐Ÿ“Š Benchmark Comparison: Reasoning Model Supervision

Evaluated against $N = 1,000$ test problems from At1uri/quantitative-reasoning-cot-benchmarks:

Model Generation Engine Unsupervised Pass@1 PRM Best-of-8 Reranked Pass@1 Absolute Gain ($\Delta$)
DeepSeek-R1-Distill-Qwen-32B 76.2% 88.4% +12.2%
Qwen-2.5-Math-72B-Instruct 81.5% 91.8% +10.3%
Llama-3.3-70B-Instruct 68.0% 81.5% +13.5%
Mistral-Large-2407 64.2% 77.8% +13.6%

๐Ÿ”— Integrated Ecosystem


๐Ÿ“œ Citation

@misc{atluri2026prm,
  author       = {Atluri, Ramprasad},
  title        = {Financial Reasoning Process Reward Model (PRM)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/At1uri/financial-reasoning-process-reward-model}}
}
Downloads last month
50
Safetensors
Model size
45.6M params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Space using At1uri/financial-reasoning-process-reward-model 1

Collection including At1uri/financial-reasoning-process-reward-model

Evaluation results

  • Step-Level Precision on Quantitative Reasoning CoT Benchmarks
    self-reported
    0.942
  • Step-Level Recall on Quantitative Reasoning CoT Benchmarks
    self-reported
    0.918
  • Hallucination Detection F1 on Quantitative Reasoning CoT Benchmarks
    self-reported
    0.930