LLM Agents Are Latent Context Managers: Eliciting Self-Managed Context via State Proprioception
VISTA enables self-managed context in LLMs via a training-free, model-agnostic layer that visualizes and archives internal states, boosting long-horizon task performance.
Key Findings
Methodology
This paper introduces VISTA (Visible Internal State for Tool Agents), a training-free, model-agnostic layer that represents working memory as typed, addressable blocks. It surfaces a real-time dashboard displaying token usage, recency, archive status, and remaining budget. Blocks can be losslessly archived externally with stable handles and recovered exactly when needed. The system integrates these features into the model’s action space, allowing it to autonomously decide what to keep, archive, or recover based on the dashboard signals. Theoretical analysis confirms that both recoverability and informative proprioception are necessary for effective context management. Empirical tests across multiple benchmarks demonstrate significant performance gains and transferability across models and scales.
Key Results
- In LOCA-Bench, VISTA improved Gemini-3-Flash success rate from 22.7% to 50.7%, achieved 58.0% on BrowseComp-Plus, and maintained competitiveness on GAIA. Across trajectories of 1M, 100K, and 10K tokens, it consistently outperformed baselines. Ablation studies confirmed the dashboard's critical role in enabling self-management, especially under high context pressure. The approach showed strong transferability across different backbone models, demonstrating its general applicability.
- VISTA’s model-agnostic design allows seamless integration with various architectures without additional training. It effectively mitigates context overflow by enabling explicit state perception and external archiving, leading to higher accuracy and efficiency in complex tasks. The experimental results validate that explicit internal state visualization significantly enhances the model’s ability to manage long sequences, outperforming traditional compression or external memory methods.
- Ablation results highlight that the dashboard’s visibility of token costs, recency, and archive status is crucial beyond mere archiving tools. The system’s theoretical guarantees on recoverability and information trade-offs underpin its robustness. These findings suggest that equipping models with explicit proprioceptive states is a promising direction for scalable, autonomous long-horizon reasoning.
Significance
This work addresses a fundamental bottleneck in large language models: the inability to self-manage their growing context. By introducing a model-agnostic, training-free interface that visualizes and externalizes internal states, it enables models to autonomously decide what information to retain or discard. This approach not only improves task success rates but also reduces computational overhead associated with context truncation. It paves the way for more autonomous, scalable, and efficient AI systems capable of handling complex, long-term reasoning tasks, with broad implications for industrial automation, scientific research, and intelligent assistants.
Technical Contribution
The paper’s key contribution is the design of VISTA, a universal, training-free layer that couples a typed, addressable block storage system with a real-time dashboard. It formalizes the necessity of both lossless recoverability and informative proprioception for effective context management, providing rigorous theoretical bounds. The system’s model-agnostic nature allows deployment across diverse architectures without retraining, representing a significant step forward in autonomous context management. Empirically, it demonstrates consistent performance improvements and transferability, establishing a new paradigm for self-managed long-horizon reasoning.
Novelty
This is the first work to formalize and implement a model-agnostic, training-free internal state visualization and management layer that couples lossless external archiving with real-time proprioceptive signals. Unlike prior approaches relying solely on learned compression policies or external memory modules, VISTA emphasizes explicit, recoverable internal states and dynamic management via a dashboard. Its hierarchical, type-based block storage and external archiving mechanisms provide a novel solution to the long-standing challenge of context overflow in large models, enabling autonomous, self-aware management.
Limitations
- VISTA’s performance may degrade under extreme long-sequence scenarios with frequent archiving, due to storage and retrieval overheads. Its effectiveness depends on the granularity of block segmentation and archiving policies.
- Real-time dashboard accuracy can be affected by high-pressure conditions, potentially leading to misjudgments in context management decisions.
- The current framework does not incorporate reinforcement learning for adaptive management, which could further enhance efficiency and decision quality in dynamic environments.
Future Work
Future research will explore integrating reinforcement learning to optimize archiving and recovery policies dynamically. Extending VISTA to multi-modal data, such as images and audio, is also promising. Additionally, deploying the system in real-world applications like autonomous agents, scientific data analysis, and industrial automation will test its scalability and robustness. Further theoretical work on information trade-offs and resource allocation will refine the framework, aiming for fully autonomous, self-regulating AI systems.
AI Executive Summary
Long-horizon tasks in large language models (LLMs) face a critical challenge: managing an ever-expanding context that exceeds the fixed window size. Existing solutions either rely on external compression policies or layer-based context control, which often discard evidence or operate outside the model’s direct perception. This paper introduces VISTA, a novel, training-free, model-agnostic layer that fundamentally rethinks context management by making the model’s internal state visible and controllable.
VISTA represents working memory as typed, addressable blocks, each with associated metadata such as token usage, recency, archive status, and remaining budget. It surfaces this information through a real-time dashboard, providing the model with proprioceptive signals that guide keep-or-archive decisions. Blocks can be losslessly archived externally with stable handles, ensuring exact recovery when needed. This design addresses the core limitations of prior methods by coupling recoverability with informative state perception, enabling the model to autonomously manage its context.
Empirical results across benchmarks like LOCA-Bench, BrowseComp-Plus, and GAIA demonstrate VISTA’s effectiveness. It significantly boosts success rates, e.g., from 22.7% to 50.7% on Gemini-3-Flash in LOCA-Bench, and achieves 58.0% on BrowseComp-Plus, outperforming baselines. The system’s transferability across models and scales underscores its generality. Ablation studies confirm that the dashboard’s visibility of token costs, recency, and archive status is crucial for self-management, especially under high context pressure.
This work advances the field by providing a universal, training-free interface that empowers models with explicit internal state awareness. It opens new avenues for autonomous, scalable AI capable of handling complex, long-term reasoning tasks efficiently. Future directions include integrating reinforcement learning for adaptive management, extending to multi-modal data, and deploying in real-world scenarios, promising a new era of intelligent, self-regulating systems.
Deep Dive
Abstract
Long-horizon tool agents are bottlenecked by how their context grows toward the limits of the context window. Recent systems make context management agent- or system-controlled, but they either learn compression policies that discard evidence or manage context in a layer the agent never sees. We argue that both miss a more basic gap: frontier language models are proprioceptively blind to their own context. From the prompt alone they cannot reliably infer block size, recency, or the remaining budget, all of which are needed for keep-or-archive decisions. We introduce VISTA (Visible Internal State for Tool Agents), a training-free, model-agnostic layer that represents working memory as typed addressable blocks, surfaces a runtime dashboard of token usage, recency, archive status, and remaining budget, and archives blocks as recoverable full-fidelity payloads. On LOCA-Bench, BrowseComp-Plus, and GAIA, the same untrained interface transfers across 1M-, 100K-, and 10K-scale trajectories. On LOCA-Bench it lifts Gemini-3-Flash from 22.7 to 50.7%, reaches 58.0% on BrowseComp-Plus, and remains competitive on GAIA. Gains grow with context pressure and transfer across backbones, while ablations confirm that the dashboard matters beyond archive and recovery tools.