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

2025/09/11 by Akshit Sinha, Sinha, Akshit, Arvindh Arun +7 · 14 voices · 13 citations
Computer Science · #Topic Modeling #Explainable Artificial Intelligence (XAI) #Natural Language Processing Techniques

paper · pdf · doi:10.48550/arxiv.2509.09677

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.

Citations

• Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
• Vending-Bench: A Benchmark for Long-Term Coherence of Autonomous Agents
• Self-Improving Transformers Overcome Easy-to-Hard and Length Generalization Challenges
• GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models
• Training Language Models to Self-Correct via Reinforcement Learning
• GAIA: a benchmark for General AI Assistants
• Toolformer: Language Models Can Teach Themselves to Use Tools
• Training Compute-Optimal Large Language Models
• Thinking Like Transformers
• Strips: A new approach to the application of theorem proving to problem solving

Cited by

• Multi²: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments
• ODYSSE: Episode-wise Policy Optimization for Personalized Agentic Reasoning
• Generalization of RLVR Using Causal Reasoning as a Testbed
• Achieving Olympia-Level Geometry Large Language Model Agent via Complexity Boosting Reinforcement Learning
• HalluHard: A Hard Multi-Turn Hallucination Benchmark
• Pessimistic Verification for Open Ended Math Questions
• Automating Structural Engineering Workflows with Large Language Model Agents
• Iterated Agent for Symbolic Regression
• h1: Bootstrapping LLMs to Reason over Longer Horizons via Reinforcement Learning
• EVALUESTEER: Measuring Reward Model Steerability Towards Values and Preferences
• Bridging Reasoning to Learning: Unmasking Illusions using Complexity Out of Distribution Generalization
• Benchmarking the Residual: What Long-Horizon Evaluations Add Beyond Matched Short-Task Performance
• Benchmarking Open-Ended Multi-Agent Coordination in Language Agents

Discussions

• Paper argues that diminishing returns to AI scale are an illusion Economic value comes from completing long projects, not single questions. And accuracy drives how long a project AI does: small gains [bsky, 62 points, 2 comments]
• Why do LLMs fail at long horizon tasks? Because errors in execution, not planning i.e. an error made early on conditions the LLM into a bad state later. Newer agentic LLMs deal with early errors bette [bsky, 53 points, 3 comments]
• Here's another paper folks are inclined to site, that helpfully includes the entire experimental setup on Github, you can reproduce it in an afternoon. "The Illusion of Diminishing Returns: Measuring [bsky, 9 points, 1 comments]
• diminishing returns to AI scale are an illusion Economic value comes from completing long projects, not single questions. And accuracy drives how long a project AI does: small gains compound exponenti [bsky, 4 points, 0 comments]
• The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs [hn, 4 points, 0 comments]
• The illusion of diminishing returns in LLM progress [hn, 3 points, 0 comments]
• The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs [hn, 3 points, 1 comments]
• The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs [hn, 2 points, 0 comments]
• The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs [hn, 1 points, 0 comments]
• For the record, I've only skimmed the Oolong paper above But I immediately started thinking of this other paper about long conversations with LLMs arxiv.org/abs/2509.09677 [bsky, 1 points, 0 comments]
• The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs [hn, 1 points, 0 comments]
• maybe a better way to think about this arxiv.org/pdf/2509.09677 [bsky, 1 points, 1 comments]
• [2509.09677] The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs arxiv.org/abs/2509.09677 - AI가 짧은 문제는 잘 해결하는데 긴 작업에서 실수하는 이유를 연구 - 간단한 덧셈 실험으로 확인한 결과, AI가 과거 실수를 보고 더 많은 실수를 [bsky, 0 points, 1 comments]
• Bueno, dices que el ejemplo es de coña pero hace poco ha salido un paper que explica que si en una conversación el LLM comete errores, afecta negativamente a su desempeño posterior. Así que la convers [bsky, 0 points, 1 comments]

Related

• I Learn Better If You Speak My Language: Understanding the Superior Performance of Fine-Tuning Large Language Models with LLM-Generated Responses
• When More is Less: Understanding Chain-of-Thought Length in LLMs
• When Fewer Layers Break More Chains: Layer Pruning Harms Test-Time Scaling in LLMs
• The Illusion of Procedural Reasoning: Measuring Long-Horizon FSM Execution in LLMs
• Line Goes Up? Inherent Limitations of Benchmarks for Evaluating Large Language Models