Show HN: Reducing LLM Context Complexity from O(N^2) to O(k)

Samama Tariq published a whitepaper on Zenodo (DOI 10.5281/zenodo.23839670) describing a Deferred Cognitive Architecture (DCA) that aims to cut LLM context cost from O(N^2) to O(k) by invoking a high-parameter model only after a Tier 1 routing engine and Tier 2 asynchronous producer-consumer relays filter the payload down to k relevant tokens, where k is much smaller than N. The architecture's Tier 1 uses deterministic point-to-point routing with O(1) lookup, and Tier 2 decouples retrieval latency from inference through non-blocking I/O queues and out-of-band indexing that updates vector stores without re-indexing context windows. The proposal targets context degradation and compute bottlenecks in large language models, though the source reports no benchmarks or measured results.

Author: Samama Tariq Whitepaper PDF: Download Full Whitepaper https://github.com/samamatariq3527/Deferred-Cognitive-Architecture/blob/main/dca whitepaper.pdf Official Citation DOI: https://doi.org/10.5281/zenodo.23839670 https://doi.org/10.5281/zenodo.23839670 Modern Large Language Models LLMs incur quadratic computational complexity 1. Tier 1: Mechanical I/O & Dynamic Routing Engine - Deterministic point-to-point routing using ultra-lightweight models or exact key-matching. - Low compute overhead $O 1 $ lookup time . 2. Tier 2: Asynchronous Producer-Consumer Relays - Non-blocking I/O queues decoupling retrieval latency from inference engines. - Decoupled out-of-band indexing allowing live, dynamic vector store updates without re-indexing context windows. 3. Tier 3: Deferred Reasoning Engine System-2 - High-parameter LLM invocation triggered only after context payload is filtered and structured. - Reduces context overhead from $N$ total tokens to$k$ relevant tokens $k \ll N$ . 4. High-parameter LLM invocation triggered If you use or reference DCA in your research or system design, please cite: @publication{tariq samama 2026 dca, author = {Samama Tariq}, title = {Deferred Cognitive Architecture DCA : Resolving Context Degradation \& Compute Bottlenecks in Large Language Models}, year = 2026, publisher = {Zenodo}, doi = {10.5281/zenodo.23839670}, url = { https://doi.org/10.5281/zenodo.23839670 https://doi.org/10.5281/zenodo.23839670 } }