Research map / Large language models / Agents, code and math
ACON: Optimizing Context Compression for Long-horizon LLM Agents
Minki Kang, Wei-Ning Chen, Dongge Han, Huseyin A Inan, Lukas Wutschitz, Yanzhi Chen, Robert Sim, Saravan Rajmohan
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add optimized compression guideline (UTCO) for history compression | gpt-4.1 agent with gpt-4.1 compressor · AppWorld test-normal | accuracy: 43.5 -> 56.5 | Improved |
| add compression maximization step (CO) after utility maximization (UT) (ablation) | ACON history compression · AppWorld test-normal | accuracy: 51.2 -> 56.5 | Improved |
| add compression maximization step (CO) after UT (ablation) | ACON history compression · OfficeBench | accuracy: 74.74 -> 72.63 | Hurt |
| add compression maximization step (CO) after UT (ablation) | ACON history compression · 8-objective QA | EM: 0.373 -> 0.335 | Hurt |
| use ACON history compression instead of no compression | gpt-4.1 agent · 8-objective QA | peak tokens: 10.35 -> 4.71 (10^3) | Improved |
| remove task contrastive feedback from guideline optimization (ablation) | ACON UT, o3 optimizer · AppWorld | average accuracy: 51.2 -> 50.6 | Hurt |
| use gpt-4.1 as optimizer instead of o3 (ablation) | ACON UT · AppWorld | average accuracy: 51.2 -> 47.6 | Hurt |
| use gpt-5 as optimizer instead of o3 (ablation) | ACON UT · AppWorld | average accuracy: 51.2 -> 50.6 | Hurt |
| lower the compression length threshold (ablation) | ACON history and observation compression · AppWorld, gpt-4.1 | accuracy and peak tokens | Mixed |
| distill compressor into Qwen3-14B, Qwen3-8B, Phi-4 students | gpt-4.1 teacher compressor · AppWorld, OfficeBench, 8-objective QA | accuracy: retains over 95% of teacher performance | No clear effect |
| use gpt-4.1-mini as compressor without distillation (ablation) | ACON history compression · OfficeBench and 8-objective QA | accuracy | No clear effect |
| add ACON compression for small distilled agent | Qwen3-14B agent · AppWorld | accuracy (%): 25.6 -> 33.9 | Improved |
Closest papers by meaning
- Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression (ICML 2025 · similarity 0.93)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.91)
- SimpleMem: Efficient Lifelong Memory for LLM Agents (ICML 2026 · similarity 0.90)
- EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.90)
- LOCA-bench: Benchmarking Language Agents Under Controllable and Extreme Context Growth (ICML 2026 · similarity 0.90)
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.89)
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.
