Reasoning Efficiently Through Adaptive Chain-of-Thought Compression: A Self-Optimizing Framework

2025-09-17

AI Summary

Chain-of-Thought (CoT) reasoning improves large language models’ accuracy and robustness but incurs substantial computational overhead—increased latency, KV cache bloat, and heightened risk of output truncation—particularly detrimental in deterministic domains like software engineering. This work empirically demonstrates, for the first time, that excessively long reasoning chains degrade performance: failed instances exhibit significantly longer chains (up to 5× higher latency than successful ones). To address this, we propose a self-optimizing CoT framework that dynamically adjusts compression thresholds via pre-inference prediction, integrating Best-of-N sampling with task-aware adaptive filtering to jointly optimize reasoning length and accuracy. Experiments show our method compresses reasoning chains by 42.1% on average, markedly reducing truncation rates and infinite-loop occurrences while improving accuracy and decreasing KV cache memory footprint—achieving both efficiency and robustness under resource constraints.

Technology Category

• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Natural Language Processing: Safety and Robustness
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Chain-of-Thought (CoT) reasoning enhances Large Language Models (LLMs) by prompting intermediate steps, improving accuracy and robustness in arithmetic, logic, and commonsense tasks. However, this benefit comes with high computational costs: longer outputs increase latency, memory usage, and KV-cache demands. These issues are especially critical in software engineering tasks where concise and deterministic outputs are required. To investigate these trade-offs, we conduct an empirical study based on code generation benchmarks. The results reveal that longer CoT does not always help. Excessive reasoning often causes truncation, accuracy drops, and latency up to five times higher, with failed outputs consistently longer than successful ones. These findings challenge the assumption that longer reasoning is inherently better and highlight the need for adaptive CoT control. Motivated by this, we propose SEER (Self-Enhancing Efficient Reasoning), an adaptive framework that compresses CoT while preserving accuracy. SEER combines Best-of-N sampling with task-aware adaptive filtering, dynamically adjusting thresholds based on pre-inference outputs to reduce verbosity and computational overhead. We then evaluate SEER on three software engineering tasks and one math task. On average, SEER shortens CoT by 42.1%, improves accuracy by reducing truncation, and eliminates most infinite loops. These results demonstrate SEER as a practical method to make CoT-enhanced LLMs more efficient and robust, even under resource constraints.

Problem

• Reducing computational costs of Chain-of-Thought reasoning
• Addressing accuracy drops from excessive reasoning steps
• Optimizing CoT length while maintaining performance accuracy

Innovation

• Adaptive CoT compression framework
• Best-of-N sampling with filtering
• Dynamic threshold adjustment pre-inference