AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Autoencoding-Free Context Compression for LLMs via Contextual Semantic Anchors

Xin Liu, Runsong Zhao, Pengcheng Huang, Xinyu Liu, Junyi Xiao, Chunyang Xiao, Tong Xiao, Shengxiang Gao, Zhengtao Yu, JingBo Zhu

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add bidirectional attention in the compressor encoder (ablation)SAC · MRQA out-of-domain (BioASQ, TextbookQA avg)F1: 45.11 -> 48.45Improved
add bidirectional attention in the compressor encoder (ablation)SAC · MRQA in-domain (TriviaQA, HotpotQA avg)F1: 63.62 -> 66.24Improved
add learnable anchor embedding to selected tokens (ablation)SAC · MRQA in-domain (TriviaQA, HotpotQA avg)F1: 64.58 -> 66.24Improved
add learnable anchor embedding to selected tokens (ablation)SAC · MRQA out-of-domain (BioASQ, TextbookQA avg)F1: 47.65 -> 48.45Improved
select anchors uniformly (middle of chunks) instead of randomly (ablation)SAC · MRQA in-domain average, 5xF1: 59.54 -> 63.63Improved
select anchors with LLMLingua-2 instead of uniform (ablation)SAC · MRQA in-domain average, 5xF1: 63.63 -> 63.26No clear effect
add autoencoding objective to pretraining (with LM) (ablation)SAC · MRQA in-domain average, 5xF1: 63.63 -> 62.04Hurt
pretrain SAC with AE only instead of LM only (ablation)SAC · MRQA in-domain average, 15xF1: 54.95 -> 49.93Hurt
add autoencoding objective to LM pretraining (ablation)500xCompressor · MRQA in-domain average, 5x/15x/51xF1: 53.23 -> 55.26 (5x); 49.76 -> 49.35 (15x); 44.46 -> 43.19 (51x)Mixed
replace EPL with SAC (anchors from context tokens)Llama-3.2-3B compressor · MRQA in-domain average, 15xF1: 68.60 -> 69.93Improved
replace EPL with SAC (anchors from context tokens)Llama-3.1-8B compressor · MRQA out-of-domain average, 15xF1: 59.88 -> 65.23Improved
replace EPL with SACLlama-3.2-1B compressor · MRQA in-domain average, 51xF1: 43.22 -> 46.37Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy