Compressing old tool outputs cuts coding-agent context with modest accuracy loss

A hybrid layout preserves recent observations as exact text while encoding older tool outputs as compact learned representations.

Top University
Zhensheng Zou (Peking University) · Guoqing Wang (Peking University) · Dan Hao (Peking University)
Research Digest··2 min read
Zou, Wang and Hao introduce a context scheme for software-engineering agents that compresses older tool observations into soft tokens while retaining the agent’s actions and recent observations as text.

The authors combine Latent Observations, Hard Actions (LOHA) with Anchored Context Distillation (ACD).

Why this paper

From Peking University · Part of Agent Harness Optimization, now 67 papers

In one line

Older tool observations can be compressed into soft tokens without losing action-critical detail, as long as recent observations and the agent's own turns stay in text.

What we could check

  • ·No code link found
  • ·No weights link found
  • ·No dataset link found
  • ·No compute details found
  • ·No stated limitations found
  • ✓Reports numbers on named benchmarks (2 benchmarks)

Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.

§
newspaper

Research Digest

Articles published under the Zotpaper byline are synthesized from multiple source publications by our AI editor and reviewed by our editorial process. Each story combines reporting from credible outlets to give readers a balanced, comprehensive view.