arxiv:2604.17091

GenericAgent: A Token-Efficient Self-Evolving LLM Agent via Contextual Information Density Maximization (V1.0)

Published on Apr 18

· Submitted by

liangjiaqing on Apr 21

Fudan University

Upvote

Authors:

Abstract

GenericAgent is a self-evolving large language model agent system that maximizes context information density through hierarchical memory, reusable SOPs, and efficient compression to overcome long-horizon limitations.

AI-generated summary

Long-horizon large language model (LLM) agents are fundamentally limited by context. As interactions become longer, tool descriptions, retrieved memories, and raw environmental feedback accumulate and push out the information needed for decision-making. At the same time, useful experience gained from tasks is often lost across episodes. We argue that long-horizon performance is determined not by context length, but by how much decision-relevant information is maintained within a finite context budget. We present GenericAgent (GA), a general-purpose, self-evolving LLM agent system built around a single principle: context information density maximization. GA implements this through four closely connected components: a minimal atomic tool set that keeps the interface simple, a hierarchical on-demand memory that only shows a small high-level view by default, a self-evolution mechanism that turns verified past trajectories into reusable SOPs and executable code, and a context truncation and compression layer that maintains information density during long executions. Across task completion, tool use efficiency, memory effectiveness, self-evolution, and web browsing, GA consistently outperforms leading agent systems while using significantly fewer tokens and interactions, and it continues to evolve over time. Project: https://github.com/lsdefine/GenericAgent

View arXiv page View PDF GitHub 5.51k Add to collection

Community

liangjiaqing

Paper submitter about 20 hours ago

Context information density is all a self-evolving LLM agent needs.

Current agent frameworks pursue ever-longer context windows, yet agent performance is bounded by context quality, not length. We identify three compounding failure modes: positional bias buries mid-context evidence, irrelevant content actively degrades reasoning, and the effective (hallucination-free) context length is roughly an order of magnitude below the nominal window. Together, they create a regime where more context can hurt rather than help.

We frame this as a structural trilemma between Completeness (all decision-critical information must be present) and Conciseness (everything else must be excluded), with Naturalness as a secondary representational constraint. The tension between the first two is not merely a budget problem — it persists even with unbounded context, because including more to be complete inherently dilutes conciseness, and compressing to be concise risks losing completeness.

GenericAgent (GA) resolves this by maximizing context information density through four coupled mechanisms. A hierarchical on-demand memory organizes all accumulated knowledge into layered tiers, but exposes only a minimal high-level index in the active context — not the knowledge itself, but a compact representation of its existence. This is sufficient: once the agent is aware that certain information exists, it can leverage its own agentic capabilities to retrieve it through tool calls, no matter how deeply it is stored. Completeness is thus preserved without loading anything upfront, and conciseness is maintained because the index itself is negligibly small. A self-evolution mechanism further raises the density ceiling over time: verified execution paths are consolidated into reusable operational knowledge, giving future tasks a stronger starting point without replaying lengthy exploration. Combined with tool minimality to reduce scaffolding overhead and active context truncation/compression, GA maintains a working budget of ~30K tokens — roughly 6× smaller than comparable frameworks — while consistently outperforming them.