Tag
VisionWeave introduces elastic visual representation weaving as a native capability of MLLMs, combining a gated spatial pooler and a granularity router to adaptively allocate visual tokens. Built on Qwen3.8-27B, it saves 43.0% of tokens while retaining 98.9% of performance and delivers 2.3x throughput gains on SGLang serving.
The paper proposes HasMem, an adaptive memory compression method for long-term LLM agents that adjusts memory widths while maintaining frozen LLM parameters, demonstrating improved performance on reconstruction and QA benchmarks.
This paper proposes telemetry-informed adaptive compression for edge-based RAG systems, showing experimental evidence that intermediate compression can reduce GPU energy by up to 53.2% with negligible quality loss.
This paper introduces OPTD, an on-policy transition distillation method with consistency-guided adaptive compression for few-step diffusion language models, improving quality-efficiency trade-offs across four reasoning and code-generation benchmarks.
This paper introduces an adaptive video tokenisation method that exploits temporal redundancy in latent space to allocate tokens dynamically, achieving efficient compression without auxiliary networks. The proposed Latent Inpainting Transformer reconstructs dropped positions, delivering 31x speedup over ElasticTok-CV and 2x over InfoTok.