CondenseFlow: Scalable Latent Space Collaboration via Semantic Compression for Multi-Agent Systems (2026.findings-acl)
Copied to clipboard
| Challenge: | Full-state latent communication in LLMs suffers from memory overhead scaling linearly with collaboration rounds. |
| Approach: | They propose a lightweight module that uses learnable semantic probes to compress KV caches into fixed-size representations. |
| Outcome: | The proposed module reduces KV cache memory by over 99% and inference latency by approximately 20% on seven benchmarks spanning six models . it outperforms text-based methods by 1.7 percentage points on average across all configurations while outperforming existing methods by 1.7%. |
Similar Papers
Latent-Condensed Transformer for Efficient Long Context Modeling (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches address these bottlenecks separately: Multi-head Latent Attention (MLA) reduces the KV cache by projecting tokens into a low-dimensional latent space, while sparse attention reduces computation. |
| Approach: | They propose a Latent-Condensed Attention mechanism that performs structured context condensation directly within MLA's latent space. |
| Outcome: | The proposed approach reduces KV cache size and attention cost without adding parameters. |
Enabling Agents to Communicate Entirely in Latent Space (2026.acl-long)
Copied to clipboard
Zhuoyun Du, Runze Wang, Huiyu Bai, Zouying Cao, Xiaoyong Zhu, Yu Cheng, Bo Zheng, Wei Chen, Haochao Ying
| Challenge: | Natural language is the de facto communication medium for LLM-based agents, but it presents a fundamental constraint . natural language downsampling limits the depth and nuance of information that can be transmitted . et al.: inter-agent latent space communication is a promising paradigm for solving complex tasks . |
| Approach: | They propose a paradigm that leverages the last hidden states of an LLM as a representation of its thought for direct communication. |
| Outcome: | The proposed paradigm outperforms fine-tuned chain-of-thought prompting and single-agent baselines even across heterogeneous models. |
Scaling External Knowledge Input Beyond Context Windows of LLMs via Multi-Agent Collaboration (2026.acl-long)
Copied to clipboard
| Challenge: | Existing context window extension methods obstruct scaling external knowledge input. |
| Approach: | They develop a multi-agent framework to overcome two core bottlenecks in existing agent orchestration designs. |
| Outcome: | The proposed framework overcomes two core bottlenecks and improves inference-time knowledge integration without longer-context training. |
Context Compression for Auto-regressive Transformers with Sentinel Tokens (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing Transformer-based LLMs have limited performance due to complexity of attention module . key-value cache is the major memory footprint and inference latency problem . |
| Approach: | They propose a plug-and-play approach that incrementally compresses token activation into compact ones . they also profile the benefit of context compression on improving the system throughout . |
| Outcome: | The proposed approach reduces memory footprint and inference latency by compressing tokens into compact ones. |
Cognitive Scaffold: From Fluid Context to Crystallized Memory for Long-Horizon DeepResearch Agents (2026.acl-long)
Copied to clipboard
| Challenge: | Scaling LLM-based agents to long-horizon deep research is constrained by context-noise trade-off . solving a single query may require hundreds of interactions with noisy environments . |
| Approach: | They propose a factorized memory architecture that decouples the cognitive state into a Fluid Working Context for immediate reasoning and a persistent Knowledge Graph for long-term retention. |
| Outcome: | The Cognitive Scaffold outperforms baselines on Xbench-DeepSearch, BrowseComp-ZH, and GAIA . it achieves 74.7% Avg@3 and 87.0% Pass@3 on xbench, browseComp, and 88.3% Pass@3. |
KV Cache Compression, But What Must We Give in Return? A Comprehensive Benchmark of Long Context Capable Approaches (2024.findings-emnlp)
Copied to clipboard
Jiayi Yuan, Hongyi Liu, Shaochen Zhong, Yu-Neng Chuang, Songchen Li, Guanchu Wang, Duy Le, Hongye Jin, Vipin Chaudhary, Zhaozhuo Xu, Zirui Liu, Xia Hu
| Challenge: | Long context capability is a crucial competency for large language models as it mitigates the human struggle to digest long-form texts. |
| Approach: | They propose to evaluate 10+ state-of-the-art approaches for long context-capable LLMs. |
| Outcome: | The proposed methods are compared against 10+ state-of-the-art approaches across seven categories of long context tasks. |
HeteroCache: A Dynamic Retrieval Approach to Heterogeneous KV Cache Compression for Long-Context LLM Inference (2026.acl-long)
Copied to clipboard
| Challenge: | Existing static compression methods suffer from coarse-grained caching and high I/O overhead. |
| Approach: | They propose a training-free dynamic compression framework that uses a sparse attention mechanism to categorize attention heads based on stability and similarity. |
| Outcome: | The proposed framework achieves state-of-the-art performance on long-context benchmarks and accelerates decoding by up to 3 compared to the original model with a 224K context. |
Adapting LLM Agents with Universal Communication Feedback (2025.findings-naacl)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have demonstrated potential for LLM agents. |
| Approach: | They propose a universal buffer and iterative pipeline to store feedback and itersative pipelines to enable LLM agents to explore and update their policy in an environment. |
| Outcome: | The proposed approach outperforms supervised instruction fine-tuning baselines on four datasets. |
AgentSlimming: Towards Efficient and Cost-Aware Multi-Agent Systems (2026.acl-long)
Copied to clipboard
| Challenge: | Automated expansion methods often result in bloated structures with redundant agents, leading to excessive token consumption. |
| Approach: | They propose a plug-and-play compression framework for graph-structured multi-agent workflows . they estimate the importance score of each agent and remove redundant agents . |
| Outcome: | Experiments show that AgentSlimming reduces average token cost by 78.9% with negligible performance degradation. |
AMA: Adaptive Memory via Multi-Agent Collaboration (2026.findings-acl)
Copied to clipboard
Weiquan Huang, Zixuan Wang, Hehai Lin, Sudong Wang, Bo Xu, Qian Li, Beier Zhu, Linyi Yang, Chengwei Qin
| Challenge: | Existing approaches to longterm memory rely on rigid retrieval granularity, accumulation-heavy maintenance strategies, and coarse-grained update mechanisms. |
| Approach: | They propose a framework that leverages coordinated agents to manage memory across multiple granularities. |
| Outcome: | The proposed framework outperforms state-of-the-art benchmarks while reducing token consumption by approximately 80%. |