Typed Blocks
Conversation, tool calls, file reads, and derived state become addressable workspace blocks.
arXiv 2026
VISTA: Self-Managed Context for Long-Horizon LLM Agents
Paper: LLM Agents Are Latent Context Managers: Eliciting Self-Managed Context via State Proprioception
1The Chinese University of Hong Kong (CUHK) 2LIGHTSPEED, Tencent
context dashboard
pressure 72%
Abstract
Long-horizon tool agents are bottlenecked by how their context grows toward the limits of the context window. VISTA represents working memory as typed, addressable blocks, surfaces runtime state through a dashboard, and archives blocks as recoverable full-fidelity payloads. The same untrained interface transfers across LOCA-Bench, BrowseComp-Plus, and GAIA.
Key Idea
Give the agent visibility into its own context state.
Proprioceptive Dashboard
The model sees token usage, recency, access history, and context pressure before acting.
Recoverable Archive
Archived blocks are moved out of the active window without destroying full-fidelity evidence.
Methodology Overview
VISTA turns an opaque transcript into a managed workspace.
Case Study Trace
VISTA archives large evidence and recovers payloads when needed.
In one 128K LOCA-Bench run, VISTA keeps the live context compact relative to a no-archive counterfactual, archives large evidence, and recovers payloads when the agent needs them.
Results
Training-free gains under context pressure.
22.7 → 50.7%
Gemini-3-Flash on LOCA-Bench
3
Benchmarks: LOCA-Bench, BrowseComp-Plus, GAIA
0
Training examples or learned compression policy
Main Results
Accuracy and trajectory cost across long-horizon agent benchmarks.
| Method | State | Control | Recover | LOCA Acc. | LOCA Traj. | BrowseComp Acc. | BrowseComp Traj. | GAIA Acc. | GAIA Traj. |
|---|---|---|---|---|---|---|---|---|---|
| Fixed external policy | |||||||||
| ReAct | - | - | - | 22.7 | 3.51M | 39.3 | 163K | 61.2 | 23K |
| Tool-result Clearing | - | - | - | 26.7 | 2.60M | 42.7 | 161K | 65.5 | 24K |
| Stale-obs. Masking | - | - | - | 28.0 | 3.32M | 38.0 | 112K | 61.8 | 24K |
| Skeleton Compression | - | - | - | 33.3 | 2.84M | 40.0 | 139K | 70.3 | 28K |
| Agent-mediated / lossy | |||||||||
| SLIM (summary) | - | partial | - | 29.3 | 3.76M | 49.3 | 162K | 67.9 | 30K |
| Active Context Compression | - | yes | - | 36.0 | 3.20M | 42.7 | 162K | 71.5 | 39K |
| Context-Folding | - | yes | - | 34.7 | 3.41M | 43.3 | 166K | 64.8 | 39K |
| Lossless external store | |||||||||
| Auto-Archive + Recover | - | - | yes | 44.0 | 2.73M | 45.3 | 133K | 63.6 | 20K |
| Claude Code | - | yes | partial | 42.7 | 6.72M | 52.0 | 247K | 73.9 | 44K |
| VISTA and ablations | |||||||||
| VISTA w/o dashboard | - | yes | yes | 37.3 | 5.25M | 50.0 | 423K | 68.5 | 24K |
| VISTA w/o recovery | yes | yes | - | 45.3 | 2.99M | 43.3 | 161K | 72.1 | 28K |
| VISTA (full) | yes | yes | yes | 50.7 | 2.86M | 58.0 | 135K | 73.3 | 33K |
Run
Download data, add an API key, and run.
git clone git@github.com:binyxu/VISTA.git
cd VISTA
cp .env.example .env
# edit .env with LOCA_OPENAI_API_KEY and LOCA_OPENAI_BASE_URL
bash run_loca_self_managed.shCitation
@article{xu2026vista,
title = {LLM Agents Are Latent Context Managers: Eliciting Self-Managed Context via State Proprioception},
author = {Xu, Binyan and Li, Haitao and Zhang, Kehuan},
journal = {arXiv preprint arXiv:2606.30005},
year = {2026}
}