A new arXiv preprint introduces a method for making web agents more token-efficient by modeling how web applications behave. The approach aims to reduce the cost of processing tokens, which the authors note is rising quickly alongside investment in agentic infrastructure. The paper is titled "Token Efficient Task Execution via Application Behavior Modeling for Web Agents."
A new arXiv paper proposes Pull, a session router that keeps conversation history in an addressable store instead of injecting all prior turns into the context window. Rather than summarizing or truncating old dialogue, the system materializes only the historical state needed at each step, aiming to avoid the quadratic token growth of full-context prompting without permanently losing information. The approach targets multi-hundred-turn, stateful LLM sessions.
A new arXiv paper examines how the way post-training updates are parameterized influences the length of text that large language models produce. The authors propose LOCUS, a task-aware low-rank adaptation method intended to curb the verbosity that standard preference alignment tends to introduce without sacrificing usefulness. The work targets serving costs, which grow with output length.
A new arXiv paper introduces VikingRAG, a retrieval-augmented generation approach aimed at cutting the number of tokens spent on structural context when working with structured documents. The authors note that current state-of-the-art RAG systems use document structure to gather better evidence but pay a high token cost for it, and their method targets that trade-off. The abstract excerpt provided does not detail the full technique or report benchmark results.
Researchers have introduced TEFM, a framework that tackles two core challenges in deploying large language models in high-stakes domains: token efficiency and faithfulness. The method is aimed at structured data and addresses both constraints together rather than treating them separately. The paper was posted on arXiv in computer science with cross-listings in machine learning.