Today for AI
HOT RADAR
agentHEAT 0.0°

AI CURATED HOT STORY · 10/12/2026

Self-Compacting Agents: Lightweight Rubrics Enable Autonomous Context Compaction Without Fine-Tuning

1 reports archived0 independent sourcesupdated 10/12/2026, 08:00:00
Synthesis & Latest Updates

To address the loss of critical intermediate results caused by fixed-threshold compaction in long agent trajectories, this paper proposes the Self-Compacting framework. By pairing a compaction tool with a lightweight rubric, it enables models to autonomously decide when to summarize context based on sub-task resolution or trajectory convergence, rather than relying on external triggers. Experiments across six benchmarks and seven models show significant cost reduction per question, revealing a meta-cognitive gap where unprompted models cannot detect context rot, which scaffolds can bridge without training.

LATEST/To address the loss of critical intermediate results caused by fixed-threshold compaction in long agent trajectories, this paper proposes the Self-Compacting framework. By pairing a compaction tool with a lightweight rubric, it enables models to autonomously decide when to summarize context based on sub-task resolution or trajectory convergence, rather than relying on external triggers. Experiments across six benchmarks and seven models show significant cost reduction per question, revealing a meta-cognitive gap where unprompted models cannot detect context rot, which scaffolds can bridge without training.

Coverage timeline

Total 1 reports · Latest first
  1. Apple Machine Learning ResearchT1·78 pts

    Self-Compacting Agents: Lightweight Rubrics Enable Autonomous Context Compaction Without Fine-Tuning

    • Fixed-interval compaction risks discarding partial results mid-derivation; Self-Compacting allows models to decide timing based on task state.
    • The core mechanism pairs a compaction tool with a lightweight rubric that defines trigger conditions (sub-task resolved/converging) and suppression cases (mid-derivation/stuck).
    • Experiments show reduced per-question costs across multiple models/benchmarks, proving scaffolds can bridge the model's inability to sense context rot without fine-tuning.