An error amplification factor governs reasoning stability
The paper asks what determines whether a long chain of reasoning stays stable. Its answer is an inference error bound governed by the product of the spectral norms of the Jacobians taken through the input space across generation steps. That product is the error amplification factor, and it serves as a quantitative measure of reasoning stability.
The distinction that matters is between per-step error and accumulated error. A model that makes small mistakes at each step can still fail badly at the end if each step multiplies the previous deviation. The amplification factor captures that multiplier, which is why it can scale exponentially with the number of reasoning steps rather than staying flat.
The bound tracks measured inference loss across steps
The paper reports that its bound closely tracks the inference loss on the Bellman-Ford algorithmic reasoning task, with both quantities able to scale exponentially with steps. In Figure 5, the loss gap and the error amplification factor are measured for T from 1 to 5 on Bellman-Ford and Prim's algorithm using Gemma-2B and Llama-1B; the inference loss increases significantly with the number of steps, and the amplification factor closely tracks that growth.
The bound is a worst-case upper bound, not an exact prediction, and the paper reports the two quantities on separate axes because they differ in absolute scale. Table 2 decomposes the bound for Qwen-1.5B, Llama-1B, and Gemma-2B on Bellman-Ford. The bound-to-gap ratio stays between 6 and 9 across all T on Qwen-1.5B, but reaches 2.3×10^6 on Llama-1B and 1.5×10^5 on Gemma-2B at T=5. The paper attributes that spread to the constants: C1 stays between 0.01 and 0.07 across the three models, while C2 differs by roughly three orders of magnitude, from around 5 on Qwen-1.5B to above 3×10^3 on Llama-1B.
The decomposition also shows a change in which term dominates. The second-order term grows faster in T than the first-order term; for Llama-1B and Gemma-2B it dominates the bound from T=3 onward, while on Qwen-1.5B the first-order term remains larger. At T=1 there is no accumulated error, so both the loss gap and the bound are zero.
Theory shows transformers can converge to decaying amplification
The paper analyzes the stability measure in transformer models trained to predict simple tasks such as linear and quadratic functions. It proves theoretically that the transformer model converges to a solution where the stability measure decays, yielding nearly zero inference loss over arbitrarily long steps.
A numerical check supports the same picture. When the norm of the Jacobian matrix product is evaluated as a function of reasoning length T, it converges to a value as T increases, consistent with the analysis of the gradient descent update rule for linear and quadratic functions. This is a result about the specific analyzed setting, not a guarantee that any trained language model will land in such a solution.
Length compression trades fewer steps against higher per-step error
The stability analysis yields two algorithmic levers. The first is chain-of-thought length compression: for a training example with T steps and a sampling factor λ in (0,1], the method uniformly samples T′=⌊λ·T⌋ intermediate steps and minimizes the training loss over that compressed sequence.
The reported effect is large. Subsampling steps reduces the error amplification factor by 92× and 181× compared with training on all steps, on the Dijkstra and Prim tasks respectively. But the paper is explicit that the trade-off cuts both ways: subsampling reduces the amplification factor while increasing the per-step errors ε_t, so the best λ is the one that balances the two effects. Intermediate numbers of steps achieve the lowest inference loss on Dijkstra and Prim.
Quantization-aware training regularizes input Jacobian norms
The second lever is quantization-aware training, which applies a uniform quantization function to the full-precision weights during the forward pass and uses a straight-through estimator for the gradient. The paper hypothesizes that this acts as a regularizer on the error amplification factor: under the pseudo-quantization noise model, the quantized model loss can be viewed as the expected loss under mean-zero randomized-rounding noise, and minimizing it penalizes the trace of the loss Hessian, which decomposes into the input-space Jacobians.
The reported reductions are 3.6× and 5.8× for the amplification factor when combining 1-bit quantization, compared with full precision, on Dijkstra and Prim. Table 3 shows the same opposing forces at T=5 for Qwen-1.5B: lowering the bit-width monotonically reduces the amplification factor and increases per-step prediction errors, so the bound and test accuracy depend on the balance. On Dijkstra and Prim, prediction errors increase by only 1.6× and 1.3×, the bound decreases by 6.8× and 2.5× at 1 bit, and test accuracy improves by 3.1% and 4.0% over full precision. On Bellman-Ford the reduction in the amplification factor is small while prediction errors double from 4 bits to 1 bit, so both the bound and test accuracy are best at 4 bits.
Gains hold across seven evaluations and longer inputs
The algorithms are validated by fine-tuning language models on graph-algorithmic reasoning tasks and symbolic state-tracking tasks. Across seven evaluations, the 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.
The comparison against named baselines is reported for Llama-1B and Gemma-2B on Bellman-Ford at T=5. The algorithm attains the lowest loss gap and the lowest bound on both models. Compared with continuous chain-of-thought (Coconut), it reduces the loss gap by 58% on average and the bound by 78% on average; compared with standard chain-of-thought fine-tuning (SFT-CoT), it lowers the bound by 330× and 65×. Within each model, the ranking of methods by the bound matches their ranking by the loss gap, even though the absolute scale of the bound differs by orders of magnitude between the two models.
Where the analysis stops and what it does not cover
The formulation targets supervised fine-tuning on ground-truth steps. The paper notes that reinforcement learning methods such as Group Relative Policy Optimization or Proximal Policy Optimization optimize over the model's own rollouts; because test-time generation is independent of the training objective, the stability analysis applies equally to models trained with reinforcement learning, but the formal extension of the bounds to that setting is left for future work.
The method regularizes Jacobian norms during training, and the paper points to test-time regularization that adapts to each input example as future work. For black-box models where gradients and Jacobians are not directly accessible, proxy stability measures based on noise perturbations may provide a practical alternative. Extending the Jacobian-based error accumulation analysis to broader language-model tasks, including tool use, generation, and question answering, also remains open.
Length generalization is a further boundary. In out-of-distribution settings the test error bound can increase in two ways: per-step prediction errors grow due to covariate shift, and the error amplification factor accumulates over the additional steps. If the step-to-step Jacobian norm exceeds one, the extra steps multiply the error geometrically, which is exactly why the regularization approach matters there.
Bit width and step count are the levers to tune
The practical reading of this paper is that stability is not a fixed property of a model but a quantity that responds to two training choices: how many intermediate steps are kept, and how aggressively weights are quantized. Both act on the same amplification factor, and both carry a countervailing cost that the paper measures rather than assumes.
The clearest decision threshold in the evidence is task-dependent. On Dijkstra and Prim, 1-bit quantization gives the best test accuracy in Table 3, while on Bellman-Ford 4 bits is best because the amplification-factor reduction is small and per-step errors double from 4 bits to 1 bit. The paper states that the bit-width at which the growth of the error levels begins to dominate depends on the task, and that it selects this bit-width on the validation set. That is the paper-specific priority: tune the compression factor and the bit width per task against the amplification factor, not against a single global setting.
Sources used in this answer
Quantifying the Stability of Multi-Step Reasoning via Error Amplification
The paper shows that multi-step reasoning stability is governed by an error amplification factor, the product of input-space Jacobian spectral norms across generation steps, which can scale exponentially with reasoning steps; it proves transformers converge to a solution where this measure decays on simple tasks, and reduces the factor in practice through chain-of-thought length compression and quantization-aware training, improving over baselines by 3.5% on average and 8.2% for longer-length inputs across seven evaluations.