AI

Curating Always-Loaded Context for LLM Agents: A Capacitated Assortment Model with Censored Feedback

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/10/2026, 04:03:25 UTC.

Briefings

This article proposes a capacitated assortment model for LLM agent context curation, where instructions are selected from a finite pool to maximize net value while respecting token attention limits. The authors prove an upper bound on optimal file size and demonstrate that blindly appending all positive-value instructions can yield worse results than selecting an optimal subset. They also analyze how censored feedback (where missing instructions only cause harm) affects learning, showing that deleting ignored instructions inevitably removes helpful ones. Empirical results indicate that irrelevant rules from real context files significantly reduce language-model compliance.

Key points

• The authors formulate context curation as a capacitated assortment problem where instructions consume tokens under a finite attention capacity, and retained instructions incur a per-session setup cost.
• The paper proves an upper bound on the optimal file size regardless of available candidate instructions and demonstrates that appending every positive-value instruction can be arbitrarily worse than selecting an optimal subset.

Context and limitations

Feedback is inherently censored, meaning benefits and harms of loaded instructions are observable while missing instructions only cause harm when their absence occurs. The study also bounds regret when human reviewers can inspect only a limited number of edits per period.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 09/10/2026, 04:32:42 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeidlzkzuav5iyybuokc6svcnig4xrdbeyruzn2aslnjadllwmkt3bi