The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs

Abstract

Does continued scaling of large language models (LLMs) yield diminishing returns? In this work, we show that short-task benchmarks may give an illusion of slowing progress, as even marginal gains in single-step accuracy can compound into exponential improvements in the length of tasks a model can successfully complete. Then, we argue that failures of LLMs when simple tasks are made longer arise from mistakes in execution, rather than an inability to reason. So, we propose isolating execution capability, by explicitly providing the knowledge and plan needed to solve a long-horizon task. First, we find that larger models can correctly execute significantly more turns even when small models have near-perfect single-turn accuracy. We then observe that the per-step accuracy of models degrades as the number of steps increases. This is not just due to long-context limitations -- curiously, we observe a self-conditioning effect -- models become more likely to make mistakes when the context contains their errors from prior turns. Self-conditioning does not reduce by just scaling the model size. But, we find that thinking mitigates self-conditioning, and also enables execution of much longer tasks in a single turn. We conclude by benchmarking frontier thinking models on the length of tasks they can execute in a single turn. Overall, by focusing on the ability to execute, we hope to reconcile debates on how LLMs can solve complex reasoning problems yet fail at simple tasks when made longer, and highlight the massive benefits of scaling model size and sequential test-time compute for long-horizon tasks.

Figures (17)

• Figure 1: A summary of our contributions. Our work measures long-horizon execution, finding large benefits from scaling model size and sequential test-time compute. We identify a failure mode where models self-condition on their own errors, degrading future performance.
• Figure 2: Growth of Horizon Length. The length of task a model can perform grows hyperbolically in the high accuracy regime.
• Figure 3: Overview of our framework. (Left) Our framework models long-horizon tasks as a sequence of retrieve-then-compose steps. (Right) We design a simple task that decouples planning from execution: in each turn, we provide the model the plan as key(s), asking it to retrieve their value(s), and compose them to maintain a running sum. We control the number of turns and turn complexity (keys per query).
• Figure 4: Scaling model size has non-diminishing improvements in the number of turns it can execute. The first-step accuracy for our task is near-perfect for all except the smallest models (a). Yet, as the model size is scaled, the horizon length increases significantly (b). We also see the effect of scaling in widening the gap between small and large models in task accuracy (c) and turn accuracy (d) as the number of turns increases. The shaded region is the mean $\pm$ one standard deviation over 100 samples; the solid line is the moving average over 5 turns; the dotted line is a hypothetical baseline model with constant step-accuracy of 0.99.
• Figure 5: Models self-condition on their previous mistakes, leading to more mistakes in subsequent turns. By manipulating the chat history, we counterfactually vary the fraction of errors in previous turns. We find this increases the likelihood of errors in future turns (left). This shows a source of degradation in turn-wise model accuracy beyond long-context, as in the turn 100 slice (right) model accuracies are much higher when we provide a fully correct history. Scaling model size increases self-conditioning, even for frontier non-thinking models.

Theorems & Definitions (3)

• Proposition 1
• Proposition 1
• proof