A model of errors in transformers

Abstract: We study the error rate of LLMs on tasks like arithmetic that require a deterministic output, and repetitive processing of tokens drawn from a small set of alternatives. We argue that incorrect predictions arise when small errors in the attention mechanism accumulate to cross a threshold, and use this insight to derive a quantitative two-parameter relationship between the accuracy and the complexity of the task. The two parameters vary with the prompt and the model; they can be interpreted in terms of an elementary noise rate, and the number of plausible erroneous tokens that can be predicted. Our analysis is inspired by an effective field theory'' perspective: the LLM's many raw parameters can be reorganized into just two parameters that govern the error rate. We perform extensive empirical tests, using Gemini 2.5 Flash, Gemini 2.5 Pro and DeepSeek R1, and find excellent agreement between the predicted and observed accuracy for a variety of tasks, although we also identify deviations in some cases. Our model provides an alternative to suggestions that errors made by LLMs on long repetitive tasks indicate thecollapse of reasoning'', or an inability to express ``compositional'' functions. Finally, we show how to construct prompts to reduce the error rate.

• The paper introduces a two-parameter model that quantifies error accumulation in transformers via small deviations in self-attention.
• It validates the model across multiple tasks and transformer models, showing close empirical fits even at higher task complexities.
• Findings indicate that targeted prompt engineering and design innovations can mitigate error rates in deterministic and repetitive tasks.

Modeling Error Accumulation in Transformer-Based LLMs

Overview and Motivation

The paper "A model of errors in transformers" (2601.14175) presents a systematic framework for understanding and quantitatively modeling the error rates observed in transformer-based LLMs on deterministic, repetitive tasks. Tasks such as arithmetic, dynamic programming, list reversal, and nested application of linear transformations are typically solved via repetitive local operations on sequences of tokens from small finite sets. Despite their apparent simplicity, state-of-the-art LLMs display degradation in accuracy on such tasks as the complexity (task length) increases.

The core hypothesis advanced is that errors primarily arise from small, systematic deviations in the self-attention mechanism; these inaccuracies accumulate as context length increases, ultimately pushing the system past a threshold where incorrect token predictions become likely. The authors construct a simple "effective field theory"-inspired model, reducing the LLM’s error characteristics on these tasks to just two interpretable parameters. This model is then validated across several LLMs and tasks at significant empirical scale.

Theoretical Model: Effective Noise and Error Directions

The framework starts by positing an effective model that mirrors the architectural structure of the idealized transformer capable of perfect performance on the chosen task, but incorporates parameter deviations reflecting learned imperfections. The input–output transformation can thus be decomposed into an idealized component and an error term.

Critical assumptions in this model include:

• Error Accumulation: Deviations in the self-attention parameters induce vector errors that accumulate across sequence length. Correlations between errors (due to context and token repetition) result in variance scaling quadratically with sequence length.
• Thresholding: Incorrect tokens are emitted when the accumulated error surpasses a geometric threshold in embedding space.
• Parameterization: The observed accuracy versus task complexity is ultimately governed by just two parameters:
  • $r$: an effective per-token "noise rate"
  • $q$: the number of plausible "error directions" (i.e., competing tokens)
• Accuracy Function: Given the above, the probability of perfect task execution as a function of complexity $c$ is

$$a = \frac{1}{\Gamma(q/2)} \gamma\left(\frac{q}{2}, \frac{q}{2 r c^{2}}\right)$$

where $\gamma$ denotes the lower incomplete gamma function, and $\Gamma$ is the standard gamma function.

This reduced model is reminiscent of effective field theory in the physical sciences, where high-dimensional microscopic dynamics are effectively encapsulated by a few macroscopic parameters governing observable phenomena.

Empirical Validation

The authors conduct a comprehensive empirical study, spanning 8 task types, 3 modern LLMs (Gemini 2.5 Flash, Gemini 2.5 Pro, DeepSeek R1), and approximately 0.2 million unique prompts. Tasks include:

• List reversal
• Nested linear transformations
• Dynamic programming (non-adjacent subsequence sum)
• Tower of Hanoi
• Decimal, binary addition, and multiplication
• Algorithmic addition (forcing a fixed decomposition algorithm)

Core Findings

• Excellent Fit: For most models and tasks, the two-parameter model tracks empirical accuracy curves closely, across large complexity ranges.
• Parameter Interpretability: Best-fit $q$ values are $\mathcal{O}(1)$ and can be aligned with the number of plausible erroneous tokens at each generation step, supporting the semantic validity of the model.
• Deviations and Insights: Systematic failures or deviations (e.g., Gemini Pro on vanilla addition) provide evidence where the core assumptions break down—such as when the model switches between inconsistent algorithms as complexity varies—demonstrating the diagnostic utility of the approach.
• Prompt Engineering: Accuracy can be improved by altering prompts to enforce focus or explicit token tagging, reducing accumulation of irrelevant contextual error.

Figure 1: Accuracy degradation of Gemini 2.5 Flash on list reversal as sequence length increases, with model fit superimposed.

Figure 2: Accuracy for nested linear transformations; quadratic scaling of variance leads to rapid decay at higher complexity for all models tested.

Figure 3: LLM accuracy on dynamic programming; modeled and observed accuracies are tightly matched, discounting reasoning collapse as primary error source.

Figure 4: Tower of Hanoi sequence prediction demonstrates accuracy plateau at low complexity and collapse at higher $c$, consistent with accumulation-of-noise model.

Figure 6: Comparison of multiplication accuracy using baseline and polynomial-tagged prompts, illustrating the impact of focused attention mappings in reducing error rate.

Implications for Transformer Analysis and AI Development

Theoretical Implications

• Error as Emergent Phenomena: Despite the vast parameter space of large transformers, error characteristics on repetitive deterministic tasks with low-entropy token spaces are governed by a fundamentally low-dimensional and interpretable parameterization—analogous to macroscopic physical laws emerging from complex microscopic systems.
• Attention Noise as Dominant Factor: The accumulation of noise in the attention mechanism, rather than fundamental architectural expressivity limitations or "collapse of reasoning," drives loss in accuracy for the relevant class of tasks.
• Effective Modeling Tools: Physics-motivated, effective-theory approximations can provide practical tools to understand and predict neural model failures, bypassing intractable microscopic analysis.

Practical Implications and Prospective Directions

• Prompt Engineering: Incorporating explicit token tagging or reformulated problem statements (such as intermediate representations) can substantially mitigate error accumulation in LLMs, a tractable approach until architectural solutions emerge.
• Architectural Innovation: Insights suggest value in augmenting attention mechanisms with inductive biases or constraints that reduce context dilution, such as explicit hierarchical attention, learnable context windowing, or value-key filtering modules.
• Evaluation and Diagnostics: The two-parameter characterization offers a principled way to benchmark and compare models for systematic error rates and to identify cases where qualitative regime change (e.g., inconsistent behavior) or model-specific peculiarities arise.
• Extension Scope: While the analysis is robust for deterministic, repetitive, small-token-space tasks, extending this modeling framework to general natural language understanding, higher-entropy tasks, or those requiring long-range semantic integration remains an open frontier.

Future Perspectives

A promising avenue is the construction of a broader mathematical framework generalizing the "effective theory" paradigm for neural architectures. Such efforts would bridge empirical machine learning practice with rigorous statistical or physical modeling, offering both explanatory and predictive power. Additionally, architectural advances inspired by these findings may improve the reliability of LLMs on algorithmic reasoning and other complex structured tasks.

Conclusion

This work provides an incisive, physics-inspired error model for transformer-based LLMs on deterministic, repetitive tasks, empirically demonstrating that their error rates can be effectively described by just two interpretable parameters. The findings challenge prior conjectures attributing such errors to reasoning collapse or expressivity failure, instead highlighting the central role of noise accumulation in the attention mechanism. This paradigm both explains present limitations and points the way to targeted improvements in both evaluation and model design.