Thinking as Compression: Your Reasoning Model is Secretly a Context Compressor

Researchers introduce Thinking as Compression (TaC), a novel approach that leverages language model reasoning traces as a natural context compression mechanism without requiring dedicated compression modules. The method demonstrates significant performance gains, outperforming existing compression baselines by 17-23% across long-context QA benchmarks at high compression ratios.

This research addresses a fundamental challenge in large language model deployment: processing long contexts efficiently without sacrificing performance. Traditional context compression methods rely on specialized components trained specifically for compression tasks, creating architectural complexity and computational overhead. The TaC framework shifts this paradigm by recognizing that the thinking or reasoning process inherent to advanced LLMs already performs compression naturally—organizing task-relevant information while filtering noise.

The innovation builds on recent advances in reasoning models that generate explicit thinking traces before producing outputs. Rather than treating these traces as intermediate steps to discard, TaC repurposes them as compressed context representations. This approach requires minimal architectural changes, making it broadly applicable across different LLM implementations. The constrained variant (TaC-C) introduces reward-driven optimization to prevent common failure modes like budget overruns and shortcuts, ensuring reliable compression performance.

The empirical results across four benchmarks represent meaningful progress in a competitive space. At 4x compression, the 17.4% F1 improvement over prior methods suggests that thinking-based compression captures more task-relevant information than alternative approaches. This efficiency gain carries practical implications for inference costs, latency-sensitive applications, and deployment on resource-constrained systems.

Looking forward, the generalizability of this approach across different model sizes, training procedures, and domain-specific tasks remains an open question. Integration with existing production inference stacks and understanding the interplay between reasoning depth and compression quality will determine real-world adoption.

• →Thinking traces naturally compress context by organizing task-relevant information without dedicated compression modules
• →TaC-C outperforms competing methods by 17-23% F1 score at 4-8x compression ratios on QA benchmarks
• →The approach reduces architectural complexity by leveraging reasoning capabilities already present in modern LLMs
• →Reward-driven optimization constrains thinking output to prevent budget overruns and ensure controllable compression
• →Results suggest reasoning models compress information more effectively than task-agnostic compression techniques