The paper "CompressionBench: How can your coding agents compress LLMs?" investigates whether autonomous coding agents can compress large language models (LLMs) while staying within a specified storage budget. The study frames compression as a terminal task where agents receive a source checkpoint, storage limit, inference environment, and time budget, then generate compressed weights and executable inference code, guided by perplexity feedback and verified by independent checks. Across six agents and six model-size settings, 29 of 36 submissions passed verification, achieving an average 3.4% increase in perplexity over BF16 and 77.33% downstream accuracy, with detailed experiments revealing strengths and failure points in agent-driven compression.
By Xinrui Zhong, Zhifeng Li, Zhiming Wang, Yongqi Liu, Shanglin Yang, Yang Zhou, Zikun Li, Zhuoming Chen, Beidi Chen
Long horizon language model agents continually accumulate reasoning history, increasing context length and inference cost even after earlier decisions have been executed and observed. Unlike static Chain of Thought compression, removing historical reasoning can change future actions and the resulting interaction trajectory.
The paper introduces Coca-Cola, a continuous latent context compression model that transforms raw tool outputs into latent tokens for LLM agents, bypassing the need for explicit text representation. By embedding these latent tokens into the agent’s input space and employing Action‑Preserving Training, the method ensures that decision‑critical information is retained while reducing token consumption by 70.01%–89.24%. Experiments on web‑search agents across multiple QA benchmarks demonstrate significant reductions in runtime (24.45%–55.97%) and consistent improvements in task performance.
By Yuetong Liu, Zhehao Zhou, Yitong Zhang, Haorui Chen, Jia Li
The paper introduces Dynamic Tool Output Compression (DTOC), a framework that manages context in large‑language‑model agents by storing full tool outputs in external memory and inserting compact placeholders into the active context. DTOC treats context updates as explicit, reversible operations within the agent’s reasoning loop, allowing selective reconstruction of compressed outputs when needed. Experiments on the DeepSWE benchmark show that for responsive models such as Sonnet 4.6 and GPT‑5.4, DTOC reduces input tokens and agent steps while significantly improving solve rates and lowering cost per solved task, with ablation studies confirming the importance of reversibility for maintaining performance.
By Abhay Chaturvedi, Shreya Bhattacharya, Rashmika Gopalkrishnan, Peter van der Putten
The paper introduces Agentic Soft-Token Compression (ASTC), a training‑free method that uses soft tokens within an LLM agent to compress long‑horizon context. ASTC leverages the model’s recurrent dynamics and Gated Delta Network attention to identify absorbed information, forwarding it as soft tokens while keeping hard‑token values intact. Experiments show ASTC matches or surpasses uncompressed agents, reducing attention pairs by up to 50.68% and outperforming trained compressors by 15.4 to 103.7 times in attention usage.
By Xuning He, Kai Fan, Runyu Zhang, Binghuan Wang, Yuxiang Zhang, Xue Jiang, Yihong Dong
Long‑horizon language model agents accumulate reasoning history, which inflates context length and inference cost. The paper introduces Interaction Aware Compression for Long Horizon Reasoning (ICLR), a training‑free online method that ranks and removes reasoning blocks based on frozen proxy entropy while preserving actions, tool calls, and observations. On 260 WorkBuddyBench tasks, ICLR raises average reward from 0.699 to 0.718 and cuts input, output, and cache read tokens by 25.5%, 14.4%, and 33.3% respectively, while analyses show that historical reasoning becomes replaceable once task‑relevant state is externalized.
By Mingxuan Wang, Fei Luo, Bo Wang, Guorun Yao, Yinglong Guo, Chao Ning, Hongyue Chen, Yanbiao Ma, Jungong Han
arXiv:2606. 28434v1 Announce Type: cross Abstract: Long-horizon software engineering agents often need to manage lengthy and noisy interaction histories under limited context budgets.
By Shuzheng Gao, Wenhao Zeng, Zhaojian Yu, Jianqiao Wangni, Chaozheng Wang, Kai Cai, Shilin He, Michael R. Lyu
SWE-VERIFIER is a lightweight generative model that selectively triggers candidate sampling and verification for coding agents, scoring proposed actions to decide when additional sampling is needed and ranking the resulting actions before execution. Trained on Qwen3.5-4B with preference data, it improves resolution on MiniMax-M2.7 from 75.47% to 77.20% at 1.09× cost and on DeepSeek-V4-Flash from 72.60% to 74.27% at 1.13× cost, while verifying fewer than 4% of steps. The approach demonstrates that selective sampling and lightweight verification can enhance coding‑agent performance and inference efficiency with minimal overhead.
By Kaijian Zou, Vaibhav Chennareddy, Yunxiang Zhang, Naihao Deng, Longju Bai, Jiaxin Pei, Rada Mihalcea, Lu Wang
ActKV is a new KV cache compression framework designed for agentic large language model (LLM) inference. It prioritizes cache entries that contribute to action generation, using action-oriented eviction, confidence-driven budget allocation, and page-aware compression to reduce memory usage while preserving accuracy. In long-trace tasks, ActKV retains 98.53% of FullKV’s accuracy using only 25.98% of its peak memory and boosts token and task throughput by 3.97× and 3.58×, respectively.
By Zihan Wang, Cheng Tang, Lei Gong, Chao Wang, Wenqi Lou, Teng Wang, Xuehai Zhou
arXiv:2608.29363v1 Announce Type: new
Abstract: Enterprise data agents answer business queries by chaining many tool calls over multiple reasoning steps, routinely accumulating hundreds of thousands...
By Ziqi Lin, Ye Wu, Mengying Yang, Xu Liu, Yizhou Liu, Qiang Ke, Qin Guo
arXiv:2510. 00615v3 Announce Type: replace Abstract: Large language models (LLMs) are increasingly deployed as agents in dynamic real-world environments, where success depends on maintaining precise records of actions and observations.
By Minki Kang, Wei-Ning Chen, Dongge Han, Huseyin A. Inan, Lukas Wutschitz, Yanzhi Chen, Robert Sim, Saravan Rajmohan
arXiv:2610.15093v1 Announce Type: cross
Abstract: AI coding agents operate in iterative loops: they modify code, run tools, inspect the results, and decide what to do next. We call the design of this...
By Pramod Kumar, Raj Kumar Pandey