A paper by Abhay Chaturvedi, Shreya Bhattacharya, Rashmika Gopalkrishnan, and Peter van der Putten introduces Dynamic Tool Output Compression (DTOC), letting agents hide or unhide tool outputs each turn to manage context. Experiments on DeepSWE tasks and internal ablations show token, step, and cost savings while improving solve rate, though results vary by model and task.
A new arXiv paper introduces FOCUS, a training-free context compression framework for LLM agents that treats compression as causal decision preservation over discrete interaction units. It runs entirely at test time with no offline data or fine-tuning and attaches to any closed-API frontier model. Evaluations across API/tool-calling, QA, web, and dialogue benchmarks report new state-of-the-art results. Practitioners get a modular, architecture-agnostic way to cut agent context costs without retraining.
A new arXiv paper introduces FOCUS, a training-free context compression framework for LLM agents that treats compression as causal decision preservation over discrete interaction units. It runs entirely at test time with no offline data or fine-tuning and attaches to any closed-API frontier model. Evaluations across API/tool-calling, QA, web, and dialogue benchmarks report new state-of-the-art results. Practitioners get a modular, architecture-agnostic way to cut agent context costs without retraining.
A new arXiv paper argues that task-completion metrics miss the interaction costs of context compression in agents. The authors introduce a controlled runtime protocol measuring reacquisition cost in a bounded-horizon tool-using agent, showing compression can force state reacquisition without changing completion stats.
A new arXiv paper argues that task-completion metrics miss the interaction costs of context compression in agents. The authors introduce a controlled runtime protocol measuring reacquisition cost in a bounded-horizon tool-using agent, showing compression can force state reacquisition without changing completion stats.
Two arXiv papers address when long-horizon agents can safely forget past reasoning. ICLR (Interaction Aware Compression) is a training-free online method using frozen proxy entropy, improving reward and cutting tokens on WorkBuddyBench; StateComp trains a router to decide when interactions are safe to compress, halving total tokens while preserving performance.
Two arXiv papers address when long-horizon agents can safely forget past reasoning. ICLR (Interaction Aware Compression) is a training-free online method using frozen proxy entropy, improving reward and cutting tokens on WorkBuddyBench; StateComp trains a router to decide when interactions are safe to compress, halving total tokens while preserving performance.
A new arxiv paper, AttnCompress, introduces dynamic attention-guided compression of interaction trajectories for Autonomous Software Engineering (ASE) agents. It targets context-window and cost bottlenecks caused by agents' lengthy trial-and-error trajectories, which prior static pruning methods handle poorly.
A new arxiv paper, AttnCompress, introduces dynamic attention-guided compression of interaction trajectories for Autonomous Software Engineering (ASE) agents. It targets context-window and cost bottlenecks caused by agents' lengthy trial-and-error trajectories, which prior static pruning methods handle poorly.
A new arXiv paper introduces ContextRender, a context management method for LLM agents that tracks how later steps reuse earlier tool results via a persistent execution-dependency graph. On AppWorld and 8-objective QA, it matches or beats passing full history while cutting mean inference cost by 10.2%-32.2% under a 6K history budget.
A new arXiv paper introduces ContextRender, a context management method for LLM agents that tracks how later steps reuse earlier tool results via a persistent execution-dependency graph. On AppWorld and 8-objective QA, it matches or beats passing full history while cutting mean inference cost by 10.2%-32.2% under a 6K history budget.