Research map / Multi-agent systems and games / LLM agents

Agent-Omit: Adaptive Context Omission for Efficient LLM Agents

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Agent-Omit, which trains LLM agents (Qwen3-4B/8B) via cold-start SFT on omission data and omit-aware agentic RL to adaptively omit redundant thoughts and tool observations, improving the effectiveness-efficiency trade-off.

Problem

LLM agents accumulate redundant thoughts and observations over multi-turn interactions, inflating token cost; existing context management methods treat all turns equally, ignoring that thought necessity and observation utility vary across turns.

Approach

Turn-level Monte Carlo analysis identifies omittable thoughts/observations; the agent is trained (SFT on single- and multi-turn omission data, then GRPO with dual full/partial trajectory sampling and a token-saving omission reward) to emit explicit omission commands adaptively.

Key results

• Agent-Omit-8B-RL reaches Pass@1 of 87.00 on TextCraft, 84.36 on BabyAI, and 23.57 on WebShop, the best among compared models, with lower average tokens than most frontier baselines.
• Against seven efficient-agent methods on Qwen3-8B, Agent-Omit-8B-RL achieves the highest Pass@1 and lowest average token cost across all five benchmarks (e.g., DeepSearch 24.56 vs 22.28 for ReSum).
• The trained agent omits on average 3-4 turns of thought/observation per trajectory, with omissions concentrated in intermediate turns (3-10).

Surprising findings

• Omitting thoughts or observations in some intermediate turns improves both accuracy and token cost, while omitting in initial or late turns hurts performance.
• Without the omission reward, the RL-trained model does not reduce tokens beyond the SFT version.

Limitations

• Additional training samples are needed during RL, limiting scalability (stated by authors).
• Evaluated only on Qwen3-4B/8B backbones; larger models not explored (future work).
• Theoretical bound relies on strong assumptions (semantic Lipschitz continuity, bounded semantic space) that are loosely justified.

Experimental setup

Datasets: DeepSearch queries (NQ, TriviaQA, PopQA, HotpotQA, 2wiki, Musique, Bamboogle) WebShop TextCraft BabyAI SciWorld

Benchmarks: DeepSearch WebShop TextCraft BabyAI SciWorld

Metrics: Pass@1 average token count

Baselines: DeepSeek-R1-0528 DeepSeek-V3.2 OpenAI o3 OpenAI o4-mini Qwen3-235B-A22B Qwen3-Next-80B-A3B Qwen3-32B Thinking-Retention DEPO Tool-Light Observation-Mask DeepMiner

Builds on: GRPO AgentGym-RL Qwen3 verl

Closest papers by meaning

• Agentic Abstention: Do Agents Know When to Stop Instead of Act? (NeurIPS 2026 · similarity 0.88)
• Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation (ICLR 2026 · similarity 0.87)
• $\texttt{Multi}^2$: Hierarchical Multi-Agent Decision-Making with LLM-Based Agents in Interactive Environments (ICML 2026 · similarity 0.87)
• Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use (ICML 2026 · similarity 0.87)
• Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)
• Pushing Forward Pareto Frontiers of Proactive Agents with Behavioral Agentic Optimization (ICML 2026 · similarity 0.87)