Quantifying the Stability of Multi-Step Reasoning via Error Amplification
Abstract
We consider the stability of multi-step reasoning processes, which have extensive applications in language models, including chain-of-thought and algorithmic reasoning. While longer sequences of reasoning can improve a model's generation capability at test time, the errors due to intermediate reasoning steps can accumulate in autoregressive generation, and thus grow substantially at the end. In this paper, we ask: What are the key factors determining the stability of multi-step reasoning? First, we show an inference error bound governed by the product of spectral norms of the Jacobians taken through the input space across generation steps. This product can be viewed as an error amplification factor, which could scale exponentially with the number of reasoning steps, serving as a quantitative measure of reasoning stability. Second, we analyze this measure in transformer models trained to predict simple tasks like linear and quadratic functions. We theoretically prove that the transformer model converges to a solution where the stability measure decays, thus yielding nearly zero inference loss over (arbitrarily) long steps. Finally, the stability analysis leads to several algorithmic implications for controlling the stability, through (i) chain-of-thought length compression that reduces the sensitivity of each step, and (ii) quantization-aware training that regularizes the input Jacobian norms. We validate the proposed algorithms by fine-tuning language models on graph-algorithmic reasoning tasks and symbolic state-tracking tasks. Across seven evaluations, our algorithms improve over baseline comparisons by 3.5% on average, and by 8.2% for longer-length inputs. Ablation analysis validates that the stability measure is drastically reduced by 3-8×, confirming the regularization effect on the spectral norms of the (input space) Jacobians.
Related
- Universal Local Error and Realized Amplification for the First-Order EDM Predictor
- Error Amplification in Koopman Predictive Control: Computable Bounds, a Conditional Lyapunov Certificate, and Horizon-Consistent Estimation
- Stability and Error Estimates with Optimal ε-Dependence for Strang Splitting of the Allen--Cahn Equation
- Energy Stability and Error Estimates for a Globally Divergence-Free DG Method of Arbitrary Order
- Optimal stability estimates and preasymptotic error analysis of the CIP-FEM for the Helmholtz equation with uniaxial perfectly matched layers
BibTeX
@misc{li2026quantifying,
title = {{Quantifying the Stability of Multi-Step Reasoning via Error Amplification}},
author = {Dongyue Li and Ziniu Zhang and Minxuan Duan and Hongyang R. Zhang},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.06404},
eprint = {2610.06404},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.06404},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Quantifying the Stability of Multi-Step Reasoning via Error Amplification
AU - Dongyue Li
AU - Ziniu Zhang
AU - Minxuan Duan
AU - Hongyang R. Zhang
PY - 2026
DA - 2026/10/05
PB - arXiv
DO - 10.48550/arxiv.2610.06404
UR - https://arxiv.org/abs/2610.06404
AB - We consider the stability of multi-step reasoning processes, which have extensive applications in language models, including chain-of-thought and algorithmic reasoning. While longer sequences of reasoning can improve a model's generation capability at test time, the errors due to intermediate reasoning steps can accumulate in autoregressive generation, and thus grow substantially at the end. In this paper, we ask: What are the key factors determining the stability of multi-step reasoning? First, we show an inference error bound governed by the product of spectral norms of the Jacobians taken through the input space across generation steps. This product can be viewed as an error amplification factor, which could scale exponentially with the number of reasoning steps, serving as a quantitative measure of reasoning stability. Second, we analyze this measure in transformer models trained to predict simple tasks like linear and quadratic functions. We theoretically prove that the transformer model converges to a solution where the stability measure decays, thus yielding nearly zero inference loss over (arbitrarily) long steps. Finally, the stability analysis leads to several algorithmic implications for controlling the stability, through (i) chain-of-thought length compression that reduces the sensitivity of each step, and (ii) quantization-aware training that regularizes the input Jacobian norms. We validate the proposed algorithms by fine-tuning language models on graph-algorithmic reasoning tasks and symbolic state-tracking tasks. Across seven evaluations, our algorithms improve over baseline comparisons by 3.5% on average, and by 8.2% for longer-length inputs. Ablation analysis validates that the stability measure is drastically reduced by 3-8×, confirming the regularization effect on the spectral norms of the (input space) Jacobians.
ER -