Abstract: The present disclosure relates to a method and system for enhancing inference in large language models (LLMs) over long input sequences. A segmented inference strategy may be employed, wherein the long context can be divided and sequentially processed through a key-value (KV) cache of the LLM. At each step, the model may generate auxiliary outputs (or margins), which may include extractive summaries or intermediate signals based on the segment's relevance to an instruction. These margins may then be classified and selectively retained to guide final inference on the instruction. The retained margins may be prepended to the instruction to facilitate improved generation without modifying the model's internal weights. The disclosed approach provides efficient localization of relevant content, improves comprehension of extended contexts, and reduces computational overhead.

Abstract: The present disclosure relates to a method and system for enhancing inference in large language models (LLMs) over long input sequences. A segmented inference strategy may be employed, wherein the long context can be divided and sequentially processed through a key-value (KV) cache of the LLM. At each step, the model may generate auxiliary outputs (or margins), which may include extractive summaries or intermediate signals based on the segment's relevance to an instruction. These margins may then be classified and selectively retained to guide final inference on the instruction. The retained margins may be prepended to the instruction to facilitate improved generation without modifying the model's internal weights. The disclosed approach provides efficient localization of relevant content, improves comprehension of extended contexts, and reduces computational overhead.

Abstract: The present disclosure relates to techniques for improving inference efficiency and memory utilization in transformer-based neural networks. A modified architectural design is introduced that decouples key and value matrix generation from inter-layer dependencies, enabling statically computed or parallelizable projections across layers. The disclosed approach may eliminate the need for layer-wise prefilling, support linear-time inference, and substantially reduce the memory footprint associated with key-value (KV) caching. The disclosed architecture may use shared or layer-specific projections, with a single KV-cache serving all or subsets of layers. In some embodiments, a non-linear transformation (e.g., implemented via a feed-forward network), may preprocess input embeddings prior to query, key, and value generation. A lookup table of transformed embeddings may be precomputed to further accelerate inference.