Metraipolis › archive › dc_95d6fc65c005

Compress What You See, Not What You Say: Anchored Context Distillation for Latent-Observation Software Engineering Agents — arXiv:2609.31430

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.31430
License (as archived)
CC0-1.0
Archived
2026-10-08T01:38:44.106Z
SHA-256 of the archived text
677bde5f7172bda111883a9d52391ea1c9d4bbf8365a4c48daa04632a35d2382
Stable URL of this page
https://metraipolis.com/documento/dc_95d6fc65c005

How to cite

Compress What You See, Not What You Say: Anchored Context Distillation for Latent-Observation Software Engineering Agents — arXiv:2609.31430 — arXiv (en), https://arxiv.org/abs/2609.31430 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_95d6fc65c005 (Metraipolis archive dc_95d6fc65c005, sha256 677bde5f7172bda111883a9d52391ea1c9d4bbf8365a4c48daa04632a35d2382)
BibTeX
@misc{dc95d6fc65c005,
  title = {Compress What You See, Not What You Say: Anchored Context Distillation for Latent-Observation Software Engineering Agents — arXiv:2609.31430},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.31430},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_95d6fc65c005; sha256 677bde5f7172bda111883a9d52391ea1c9d4bbf8365a4c48daa04632a35d2382},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.31430. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Tool observations dominate the context of software-engineering agents, making long interaction histories costly to maintain. Existing context compression methods can discard information needed by later actions, while adapting agents to soft-token representations can compromise their original behavior. To reduce context while preserving action-critical information and agent behavior, we combine Latent Observations, Hard Actions (LOHA), a context layout that separates compressed history from text needed for exact reference, with Anchored Context Distillation (ACD), a training method that enables latent reading while constraining behavioral drift. LOHA compresses older tool observations into soft tokens while retaining the agent's own turns and the last K observations in text, providing compact access to historical information and exact access to recent content. To enable the agent to use this representation, ACD distills the base model's full-text predictions into the latent view while anchoring its behavior on plain-text inputs to the same base model. On SWE-bench Verified, K=3 reduces context per call by 43% for Qwen3-4B and 57% for SWE-Master-4B-RL, with resolve rates of 12.1% and 21.8% versus 14.5% and 27.5% for their uncompressed bases. A single-run recency sweep reaches 14.4% and 23.0% at K=8, with larger windows generally favoring task performance over compression. Under a 32K-token limit, Qwen3 with K=3 resolves 21.1% of a 199-instance subset versus 11.1% for the same adapted agent using full text. In concurrent single-GPU serving, it achieves 1.9 times that full-text agent's instance throughput.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-28 in cs.AI).
Categorie arXiv: Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2609.31430 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-08T01:12:14.379Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-28?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).