ResearchResearch paperLarge Language Models · Reinforcement Learning1 source · Oct 6, 2026

TAP: Efficient Long-Horizon Agent Pruning via Trajectory-Anchored Recovery

In this work, we propose Trajectory-Anchored Pruning (TAP), the first structural pruning framework for reinforcement learning (RL)-trained agents.

Emerging long-horizon agentic tasks require repeated model calls, worsening the inference cost of already-costly language models.
While narrow agentic tasks suggest potential for aggressive model pruning without performance drop, empirical results show existing methods proposed for question answering tasks severely degrade task performance when applied to agentic models.
TAP couples structural pruning with efficient on-policy recovery, anchoring interactions to teacher trajectories while allowing the student to generate each reasoning-action response.
These results demonstrate effective structural compression of long-horizon agents under a limited recovery budget.

[1]TAP: Efficient Long-Horizon Agent Pruning via Trajectory-Anchored Recovery
arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 6, 08:18 PM
In this work, we propose Trajectory-Anchored Pruning (TAP), the first structural pruning framework for reinforcement learning (RL)-trained agents.
Emerging long-horizon agentic tasks require repeated model calls, worsening the inference cost of already-costly language models.

Extractive summary: sentences quoted from the sources.

Before this