Paper Detail

The Optimization Landscape of Learning Compacted Context Models

Thomas Villeneuve, Alex Sandomirsky, Charles O'Neill, Max Kirkby, Michael Psenka

arxiv Score 11.5

Published 2026-10-05 · First seen 2026-10-06

Research Track A · General AI

Abstract

Many works approach continual learning through the lens of infinite context windows. As an agent puts more observation into context (concretely the KV cache), compacting said context is akin to direct memory manipulation, without affecting the base model's weights. Many works pose KV compaction as an optimization problem: learn a smaller set of KV vectors that matches the behavior of the full KV cache. While this preserves base model behavior, optimizing through a frozen base model results in a highly nontrivial optimization problem with a brittle and flat loss landscape. In this paper, we characterize what makes these optimization problems difficult and demonstrate that a heavily simplified Perceiver-based architecture not only matches performance of a full Perceiver transformer in continuous context compaction, but outperforms baselines on compaction utility. Results are presented on MCQ tasks across Finance, Legal, Gutenberg, and Code.

Workflow Status

Review status
pending
Role
unreviewed
Read priority
now
Vote
Not set.
Saved
no
Collections
Not filed yet.
Next action
Not filled yet.

Reading Brief

No structured notes yet. Add `summary_sections`, `why_relevant`, `claim_impact`, or `next_action` in `papers.jsonl` to enrich this view.

Why It Surfaced

No ranking explanation is available yet.

Tags

No tags.

BibTeX

@article{villeneuve2026optimization,
  title = {The Optimization Landscape of Learning Compacted Context Models},
  author = {Thomas Villeneuve and Alex Sandomirsky and Charles O'Neill and Max Kirkby and Michael Psenka},
  year = {2026},
  abstract = {Many works approach continual learning through the lens of infinite context windows. As an agent puts more observation into context (concretely the KV cache), compacting said context is akin to direct memory manipulation, without affecting the base model's weights. Many works pose KV compaction as an optimization problem: learn a smaller set of KV vectors that matches the behavior of the full KV cache. While this preserves base model behavior, optimizing through a frozen base model results in a },
  url = {https://arxiv.org/abs/2610.05885},
  keywords = {cs.LG},
  eprint = {2610.05885},
  archiveprefix = {arXiv},
}

Metadata

{}