Closing the Social-Semantic Gap: SPSD for Edge-Based Prompt Compression in Cloud LLM Inference

2026-06-10

AI Summary

This study addresses the “social-semantic gap” in cloud-based large language model (LLM) inference, where user prompts often contain substantial social redundancy—such as pleasantries and repetitive phrasing—leading to high energy consumption during the prefill stage with minimal semantic gain. To tackle this, the authors formally characterize the problem and propose Sentiment-Preserving Semantic Distillation (SPSD), a method that leverages a 4-bit quantized small model (Gemma-2-2B-Instruct) at the edge to compress prompts while preserving essential sentiment. A rule-based gating mechanism ensures safety, and the distilled input is then forwarded to the cloud LLM. Experiments demonstrate that SPSD reduces input tokens by an average of 99.9 per invocation (with positive net savings across all 146 tested cases), maintains non-inferior response quality (71% of outputs rated equal or better on a 15-point scale), and achieves net energy savings of 70–270 µWh per call.

Abstract

The prefill stage of Large Language Model (LLM) inference is a growing contributor to cloud-scale energy cost. Many consumer-support and conversational prompts contain social scaffolding: politeness markers, apologetic preamble, repetition, and rapport-building language that is important for human communication but carries low marginal information for machine reasoning. We call this discrepancy the Social-Semantic Gap. We present SPSD (Sentiment Preserving Semantic Distillation), an edge-based pipeline that compresses user prompts using a 4-bit quantised Small Language Model before transmission to a cloud-deployed LLM. Evaluation on a 248-prompt corpus using Gemma-2-2B-Instruct (Q4_K_M) as the SLM and Llama-3.1-8B-Instruct as the cloud evaluation model yields a mean input token saving of 99.9 tokens per distilled call, with all 146 distilled calls yielding positive savings. Response quality, assessed by blind LLM-as-judge scoring across 121 pairs, is non-inferior to the raw path within a pre-specified 1-point margin on a 15-point rubric; the judge awarded 43 percent ties, 28 percent distilled wins, and 29 percent raw wins. Cosine similarity is mixed: mean 0.682, median 0.712, with 54.1 percent of pairs above the 0.70 reference threshold. Safety-critical domains are conservatively routed to passthrough via rule-based gates. Per-call net energy saving is estimated at 70-270 uWh under stated assumptions. SPSD shows that on-device prompt distillation can reduce cloud LLM input-token cost while preserving response quality within a practical non-inferiority margin.

Problem

• Social-Semantic Gap
• LLM inference
• prompt compression
• energy cost
• cloud computing

Innovation

• prompt compression
• edge-based distillation
• social-semantic gap
• sentiment preservation
• energy-efficient LLM inference