TLDR: This research paper, “The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs,” argues that marginal gains in single-step accuracy for large language models (LLMs) can lead to exponential improvements in the length of tasks they can complete. It posits that LLM failures in long tasks are often due to execution errors, not reasoning. The study introduces a ‘self-conditioning effect’ where models become more prone to errors after observing their own past mistakes, a problem not solved by just scaling model size. However, ‘thinking models’ (using Chain-of-Thought reasoning) effectively mitigate this issue, enabling them to execute significantly longer tasks. The paper concludes that continued investment in LLM scaling is justified by the economic value of long-horizon task completion.
In the rapidly evolving world of artificial intelligence, a central question often arises: are we reaching a point of diminishing returns with the continuous scaling of large language models (LLMs)? A new research paper, “The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs”, by Akshit Sinha, Arvindh Arun, Shashwat Goel, Steffen Staab, and Jonas Geiping, offers a compelling counter-narrative, suggesting that even small improvements in single-step accuracy can lead to massive, exponential gains in the length of tasks an LLM can successfully complete.
The authors argue that when LLMs falter on seemingly simple but extended tasks, the issue often lies not in their ability to reason, but in their capacity for sustained execution. They propose a novel approach to isolate and measure this ‘execution capability’ by explicitly providing LLMs with all the necessary knowledge and a clear plan to solve a long-horizon task. This setup allows them to observe how well models can carry out instructions over many steps, free from the complexities of planning or knowledge retrieval.
The Challenge of Long-Horizon Execution
The research reveals that even when small models achieve perfect accuracy on a single step, their performance rapidly declines as the task length increases. This highlights that long-horizon execution is a significant hurdle for LLMs. However, the study also brings good news: scaling up model size leads to substantial improvements in the number of turns a model can correctly execute. This benefit is non-diminishing, meaning larger models are consistently more reliable for longer tasks, even when the task doesn’t require more complex reasoning or knowledge.
The Self-Conditioning Effect: A Unique LLM Flaw
One of the paper’s most intriguing findings is the ‘self-conditioning effect’. Unlike humans who tend to improve with practice, LLMs’ per-step accuracy degrades as a task progresses. The researchers found that models become more likely to make mistakes when their past errors are present in the conversation history. Essentially, an LLM can ‘condition’ itself on its own faulty outputs, leading to a cascade of further errors. Crucially, this self-conditioning effect is distinct from limitations imposed by long context windows and, surprisingly, is not mitigated by simply scaling up model size.
Thinking Models: A Solution to Self-Conditioning
The paper identifies a promising solution in ‘thinking models’ – LLMs that employ sequential test-time computation, often referred to as Chain-of-Thought (CoT) reasoning. These models, especially those trained with reinforcement learning, demonstrate a remarkable ability to overcome the self-conditioning effect. Their accuracy remains stable regardless of past errors in the context, suggesting they process each turn more independently or are better at focusing on the task’s success rather than merely continuing the context. The study benchmarks frontier thinking models, with GPT-5 (codename “Horizon”) leading the pack, capable of executing over 1000 steps in a single turn, far surpassing its competitors.
Also Read:
- The Length of AI’s Reasoning: Not Always a Sign of Deeper Thought
- Navigating Complexity: How New AI Framework Guides LLMs to Smarter Reasoning
Implications for the Future of LLMs
This research offers a crucial perspective on the ongoing debate about LLM capabilities and investments. While traditional metrics might show diminishing returns on single-step performance, the ability to reliably complete long, multi-step tasks is where the true economic value of LLMs lies. The paper suggests that continued investment in scaling compute for LLMs is economically justified, as it directly translates to an exponential increase in the length of tasks they can handle. By decoupling planning and execution, the authors provide a clearer understanding of why LLMs can excel at complex reasoning yet stumble on extended simple tasks, paving the way for more robust and reliable AI agents in the future.
