← The Wire

Public story · 2026-03-02 · source-backed

Context Pollution: LLMs Don't Benefit From Their Own Words

Confidence
source-backed
Sources
1
Redaction
redacted

Story

Counterintuitive finding: 36.4% of multi-turn prompts are fully self-contained. Removing prior assistant responses yields up to 10x context length reduction with minimal quality loss. Models over-condition on previous responses, introducing errors. Directly actionable for multi-turn agent builders. arXiv 2602.24287


Related stories

Each link below shares sources, entities, or timing with this story.

  1. Shared entities / Same source domain / Shared topic / What happened next

    BAVT: Budget-Aware Value Tree Cuts Wasted Tool Calls.

    Both cover Directly, Models; reported by the same outlet (arxiv.org); overlapping topics (agent, directly).

  2. Shared entity: Directly / Same source domain / Shared topic / What happened next

    Lore: Git Commits as Structured Knowledge Protocol.

    Both cover Directly; reported by the same outlet (arxiv.org); overlapping topics (actionable, agent, directly, loss).

  3. Codified Context: Infrastructure for AI Agents in Complex Codebases

    Both cover Directly; reported by the same outlet (arxiv.org); overlapping topics (actionable, agent, context, directly).

  4. Shared entity: Directly / Same source domain / Shared topic / Earlier coverage / Tension

    AgentSentry: Defending Against Indirect Prompt Injection in Tool-Using Agents

    Both cover Directly; reported by the same outlet (arxiv.org); overlapping topics (agent, context, directly).

  5. Shared entity: Directly / Same source domain / Shared topic / What happened next / Tension

    SplitAgent: Privacy-Preserving Enterprise Agents

    Both cover Directly; reported by the same outlet (arxiv.org); overlapping topics (agent, directly).

  6. Shared entity: Directly / Same source domain / Shared topic / What happened next

    Tangram: non-uniform KV cache for cheaper multi-turn serving.

    Both cover Directly; reported by the same outlet (arxiv.org); overlapping topics (agent, directly, multi-turn).

  7. Vortex makes sparse attention programmable for long-generation agents.

    Both cover Directly; reported by the same outlet (arxiv.org); overlapping topics (agent, directly, llms).

  8. New RL framework trains CLI agents under partial observability.

    Both cover Directly; reported by the same outlet (arxiv.org); overlapping topics (agent, assistant, directly).

Source trail

Entities

Provenance

AI generated
no
Story unit
2026-03-02-context-pollution-llms-don-t-benefit-from-their-own-words
Labels
source-backed, canonical briefing excerpt