Tag
该论文提出一个可复现的离线评估框架，用于在生产环境的可变企业数据上评测对话智能体，并将其作为 GEPA 提示优化的适应度函数，在正确性约束下实现提示压缩 3.5-7.6% 而不损失任务准确率。
论文提出 CQ2-CARE 方法，在固定 token 预算下为编码智能体选择完整文件而非压缩或检索片段，从而在 SWE-rebench 上显著提升问题修复率，超过压缩器、检索器及专用训练方法。
The paper introduces a training-free, deterministic pipeline for lexical prompt compression in large language models, featuring empirical Pareto analysis across eleven task categories.
This paper presents HSDN, a hierarchical semantic distillation network for compressing chain-of-thought reasoning in financial tasks, achieving 91.0% accuracy with 68.4% compression on the AFAC2025 benchmark.
This paper audits extractive prompt compressors across ten languages, revealing that English-trained models exhibit significant performance gaps on non-English text at high compression rates and proposes a translate-then-compress pipeline as a more effective alternative.
IterCOMP is a training-free prompt compression framework that iteratively integrates essential evidence to improve efficiency and accuracy in multi-hop question answering.
This paper identifies a structural failure in hard prompt compression called 'referential dangling', where independent scoring splits dependent evidence pairs, removing the context needed to interpret retained answers. Experiments show this affects many compressors and datasets, and automatic restoration of missing references improves QA accuracy.
This paper introduces SkillSV, a structure-aware Shapley-style framework for valuing the internal units of agent skills, accounting for dependencies and hierarchy. It demonstrates improved faithfulness, actionability, and explanation quality on agentic benchmarks.
This paper proposes using linguistic rules alone as prompt compressors without LM forward passes, achieving performance similar to advanced strategies under light-to-moderate compression.
Proposes Cache-Aware Prompt Compression (CAPC), a method that combines query-agnostic compression with caching to reduce LLM API costs, demonstrating significant savings over existing approaches on Anthropic's Sonnet API and production workloads.
This paper unifies memory compaction techniques across LLMs and agents under a rate-distortion framework, proposing a taxonomy and benchmark for evaluating compression across different layers.
This paper proposes compressing instruction prompts into a single activation vector via learned weighted sums of intermediate layer activations, achieving under 2% accuracy drop and revealing insights into LLM activation space structure.
This paper proposes RAGP, a prompt compression method that models text as a multiplex graph and uses Lévy walks for redundancy-aware graph pruning, achieving superior performance on LongBench compared to existing LLM-based compression methods.
This paper systematically investigates when and which parts of the prompt KV cache become dispensable during LLM decoding, showing that redundancy primarily involves chat template scaffolding rather than task content, and replacement with neutral filler preserves accuracy.
AGORA introduces an inference-free step-level prompt compressor for LLM agents that avoids the 'action-grammar destruction' failure mode of token-level compressors. It retains ≥75% uncompressed performance in 8 of 9 environments across backbones, using a structural parser, an always-keep floor, and a learned relevance scorer.
This paper introduces MedTPE, a method for efficient, lossless prompt compression of electronic health records for large language models, significantly reducing token length and inference latency in clinical prediction tasks.