PaperlayerView exact revision on arXiv

Source-linked AI summary

Agent-Omit: Adaptive Context Omission for Efficient LLM Agents

Yansong Ning, Jun Fang, Naiqiang Tan, Hao Liu

arXiv:2602.04284v2cs.AIcs.LG

TL;DR

LLM agents face redundant thoughts and growing observation context, while existing methods overlook turn-level variation in their utility. The paper introduces Agent-Omit, which combines omission-data-driven fine-tuning with omit-aware agentic reinforcement learning. Across five benchmarks, Agent-Omit-8B achieves a better effectiveness–efficiency trade-off than seven efficient-agent methods, with a stated scalability limitation from additional RL training samples.

  • Problem

    Existing agent-context methods treat interaction trajectories uniformly, overlooking that thought necessity and observation utility vary across turns.

  • Method

    Agent-Omit combines cold-start omission-data synthesis with omit-aware agentic reinforcement learning for adaptive thought and observation omission.

  • Results

    Across five diverse benchmarks, Agent-Omit achieves a better effectiveness–efficiency trade-off than seven efficient-agent methods and performs strongly against frontier agents.

  • Takeaways & Limitations

    The results support turn-level omission as a training paradigm for efficient LLM agents while retaining task effectiveness within the reported evaluations.

  • Takeaways & Limitations

    The reinforcement-learning stage requires additional training samples, limiting the scalability of the approach.

Abstract

from arXiv · show

Managing agent context (e.g., thought and observation) during multi-turn agent-environment interactions is an emerging strategy to improve agent efficiency. However, existing studies treat the entire interaction trajectories equally, overlooking the thought necessity and observation utility varies across turns. To this end, we first conduct quantitative investigations into how thought and observation affect agent effectiveness and efficiency. Based on our findings, we propose Agent-Omit, a unified training framework that empowers LLM agents to adaptively omit redundant thoughts and observations. Specifically, we first synthesize a small amount of cold-start data, including both single-turn and multi-turn omission scenarios, to fine-tune the agent for omission behaviors. Furthermore, we introduce an omit-aware agentic reinforcement learning approach, incorporating a dual sampling mechanism and a tailored omission reward to incentivize the agent's adaptive omission capability. Theoretically, we prove that the deviation of our omission policy is upper-bounded by KL-divergence. Experimental results on five agent benchmarks show that our constructed Agent-Omit-8B could obtain performance comparable to seven frontier LLM agent, and achieve the best effectiveness-efficiency trade-off than seven efficient LLM agents methods. Our code and data are available at https://github.com/usail-hkust/Agent-Omit.

1. Introduction

LLM agents often incur redundant thought and observation context, while existing compression methods treat interaction trajectories uniformly. Agent-Omit instead learns selective omission and achieves strong effectiveness–efficiency results across agent benchmarks.

  • LLM agents generate redundant thought and accumulate excessive observations, limiting efficiency and practical applicability.
  • Existing thought, observation, and joint-management methods compress or modify interaction trajectories broadly, overlooking turn-varying influence.
  • Agent-Omit combines omission-behavior synthesis with omit-aware agentic reinforcement learning to adaptively omit redundant thoughts and observations.Cold-start data covers single-turn and multi-turn omission scenarios; reinforcement learning uses dual sampling and a tailored omission reward.
  • Agent-Omit-8B achieves accuracy comparable to seven frontier LLMs while substantially reducing token cost across five agent benchmarks.The evaluated benchmarks are DeepSearch, WebShop, TextCraft, BabyAI, and SciWorld.
  • The framework integrates thought-and-observation analysis, adaptive context management, extensive experiments, and theoretical analysis as its main contributions.

2. Preliminary

The paper models multi-turn agent–environment interaction through thoughts, actions, and observations. Thoughts support reasoning and planning, actions execute operations, and observations provide environmental feedback for subsequent turns.

  • At each interaction turn, the agent–environment state includes thought, action, and observation components.
  • A thought is chain-of-thought reasoning used to analyze context, plan actions, or reflect on historical feedback.
  • An action is an operation selected from a predefined action space, including tool invocation or final-response generation.
  • An observation is environmental feedback returned after executing an action and provides grounding for the next interaction turn.

3. Analysis of Thought and Observation on Agent Effectiveness and Efficiency

Thought necessity and observation utility vary by interaction turn, creating opportunities for selective omission. Experiments show that intermediate omissions can reduce token costs without sacrificing accuracy, whereas early and late context is often more important.

  • Turn-Level Variation: Thought necessity and observation utility are turn-dependent rather than equal across the interaction trajectory.Complex steps may require detailed reasoning, while some intermediate steps can execute directly; older observations may become redundant.
  • Quantitative Analysis: Qwen3-8B WebShop analysis measures turn-wise token costs and task-accuracy contributions for thoughts and observations.
  • Quantitative Analysis: Thought and observation tokens dominate agent cost at 45.1% and 52.2%, respectively, while actions account for 2.7%.
  • Quantitative Analysis: Thoughts consume more tokens in early planning turns, whereas stacked observations create increasing context burden in later turns.
  • Quantitative Analysis: Early thoughts and observations are critical for high Pass@8 accuracy, but their accuracy contribution diminishes rapidly in later turns.Later-turn contributions often fall below the Pass@1 baseline.
  • Motivation for Adaptive Omission: Because optimal omission points are dynamic and task-dependent, selective adaptive omission is needed instead of a static heuristic.
  • Thought Omission: Omitting intermediate thoughts can improve accuracy and reduce token usage, but omitting initial or final thoughts is detrimental.Initial reasoning supports high-level planning, while final reasoning supports contextual continuity.
  • Observation Omission: Observations are most omissible in intermediate turns such as Turns 2, 4, and 6, while later observations become indispensable.Later omission can reduce accuracy and trigger extra reasoning tokens to bridge missing information.

4. Efficient LLM Agents with Agent-Omit

Agent-Omit balances task success with minimal context overhead by adaptively omitting redundant thoughts and historical observations. It combines cold-start omission training with omit-aware agentic reinforcement learning using dual sampling and omission-aware rewards.

  • Agent-Omit formulates context management as selectively reducing redundant thoughts and omitting irrelevant historical observations during interaction.The policy may reduce a thought to an empty string and outputs an observation omission set for irrelevant historical turns.
  • The framework uses two-stage optimization: synthetic cold-start data establishes omission behavior, followed by omit-aware agentic RL for adaptive omission.The cold-start data covers omission execution and reasoning under omitted historical context.
  • Cold-start data is built through omission-turn identification, hierarchical behavior synthesis, and training on single-turn and multi-turn omission scenarios.Rollouts evaluate whether omitting a thought or observation reduces token use without reducing task performance; synthesized datasets are then fine-tuned with language-modeling loss.
  • Dual sampling separates full trajectories, which evaluate overall efficiency and success, from partial trajectories, which expose omission decisions to training.Each full trajectory yields a variable number p(y) of partial trajectories corresponding to omission-triggered turns.
  • The omit-aware reward combines task correctness with token savings and sets omission reward to 0 when task reward is 0, preventing reward hacking.The omission reward is based on the ratio of saved tokens, while multi-objective policy learning optimizes rewards from full and partial trajectories.

5. Theoretical Analysis

The theoretical analysis models omission quality through semantic continuity assumptions and policy distributional shift. It shows that deviations in effectiveness and efficiency are bounded by the KL-divergence between the learned and optimal omission policies.

  • The analysis asks how much effectiveness and efficiency deviate from an optimal omission strategy and measures policy distance with KL-divergence.The bound provides a guarantee relating learned-policy approximation to omission quality.
  • The theory assumes task accuracy and token cost are Lipschitz continuous with respect to semantic distance between trajectory embeddings.Under this assumption, semantically similar trajectories have similar reward and cost outcomes.
  • The bounded omission error theorem upper-bounds expected deviations in effectiveness and efficiency between the learned policy and the optimal omission policy.The total omission error is modeled as distributional shift during reinforcement learning.
  • As KL-divergence is minimized during reinforcement learning, accuracy and token cost converge toward the optimal omission frontier, subject to approximation errors.The convergence claim follows from the stated theorem and its reinforcement-learning objective.

6. Experiments

Experiments across five agent benchmarks show that Agent-Omit combines strong task accuracy with reduced token use, outperforming efficient-agent baselines and adaptively omitting context during inference.

  • Experimental Setup: The evaluation covers DeepSearch, WebShop, TextCraft, BabyAI, and SciWorld, spanning information search, web navigation, digital games, and embodied control.The framework is evaluated against frontier LLM agents and efficient agent-construction methods using Pass@1 accuracy and token efficiency.
  • Main Results: Agent-Omit-8B-RL achieves the highest Pass@1 score on four benchmarks and remains marginally below the state of the art on DeepSearch.It also requires substantially fewer tokens than several reasoning frontier models, while slightly exceeding some non-reasoning models’ token costs for higher accuracy.
  • Main Results: On Qwen3-8B, Agent-Omit achieves the highest Pass@1 accuracy and lowest average token cost across all five benchmarks.Heuristic methods reduce tokens but degrade accuracy, while summarization methods provide a more limited trade-off.
  • Ablation Study: Both SFT and RL contribute to effectiveness and efficiency gains, with single-turn omission data and partial trajectory sampling being especially important.The omission reward is the primary efficiency factor; changing its reweighting factor from 0.2 worsens the trade-off.
  • In-Depth Analysis: The trained agent omits an average of 3 to 4 thought or observation turns per trajectory, concentrated mainly in turns 3–10.This nonuniform pattern indicates that omissions predominantly occur during intermediate interaction turns.

7. Related Work

Related work reduces agent context through thought management, observation management, or joint summarization, but generally applies compression uniformly across interaction trajectories.

  • Existing Efficiency Methods: Thought Management methods compress generated reasoning, while Observation Management methods prune or mask historical observations.Examples include fine-tuning-based thought compression and heuristic observation omission strategies.
  • Existing Efficiency Methods: Thought-and-Observation Management methods jointly summarize reasoning and observations into a shorter context.These approaches use LLM-based summarization tools to compress both context components.
  • Research Gap: Existing context-management studies generally modify entire interaction trajectories equally, overlooking turn-level variation in thought influence and observation utility.This limitation motivates adaptive omission based on the differing roles of context across turns.
  • Agentic Reinforcement Learning: Agentic reinforcement learning trains agents through environment interaction and reward feedback, supporting end-to-end optimization in constructed sandboxes.Recent frameworks apply this paradigm to improve agent capabilities across tool-use tasks.

8. Conclusion, Limitation and Future Work

The paper presents a turn-level, omission-data-driven training framework and reports strong results across five benchmarks, while identifying additional RL samples as a scalability limitation.

  • Conclusion: Agent-Omit combines turn-level analysis, omission-data-driven cold-start fine-tuning, and omit-aware agentic reinforcement learning.The framework is experimentally and theoretically grounded for efficient LLM agent training.
  • Conclusion: Experiments on five diverse benchmarks show that small-sized Agent-Omit models outperform seven frontier LLM agents and seven efficient-agent methods on effectiveness–efficiency trade-offs.This conclusion is stated at the paper level across the evaluated tasks and methods.
  • Limitation and Future Work: Additional training samples required during the RL stage limit the approach’s scalability.Future work targets scaling omission-data synthesis to large-scale pre-training and extending the paradigm to larger LLMs.

Impact Statement

The paper aims to reduce redundant token use and computational cost in multi-turn LLM agents through adaptive omission, with potential energy-footprint benefits and no identified specific additional societal harms.

  • Impact: Adaptive omission is intended to reduce token consumption and computational cost during multi-turn agent interactions.The stated motivation is improving efficiency by removing unnecessary thoughts and observations.
  • Impact: Lower deployment cost may contribute to reducing the energy footprint of large language model agents.The paper presents this as a possible contribution rather than a measured outcome.
  • Impact: The paper reports no specific negative societal consequences beyond those generally associated with large language models.This is the authors’ stated impact assessment.
  • Theoretical Assumptions: The theoretical analysis assumes that reward and token-cost functions are Lipschitz continuous with respect to semantic trajectory distance.The proof framework maps interaction trajectories into token embeddings and treats effectiveness and efficiency as functions over trajectories.

A.3. Theorem 1: Bounded Omission Error

Theorem 1 states that adaptive omission bounds the expected effectiveness and efficiency deviation between the learned policy πθ and expected policy π∗ by their KL-divergence. The proof derives this result through Wasserstein distance, a transport inequality, and Lipschitz assumptions.

  • The expected deviation in effectiveness and efficiency between πθ and π∗ is upper-bounded by KL(π∗, πθ).
  • The reward-function proof bounds expectation differences using Kantorovich-Rubinstein duality and the Wasserstein-1 distance between policies.
  • A transport inequality upper-bounds the Wasserstein distance by KL-divergence when the semantic space is bounded or satisfies measure concentration.
  • For any ω > 0, the inequality √x ≤ 1/(2ω) + (ω/2)x converts the square-root bound into a linear upper bound.
  • The reward and efficiency bounds use Lipschitz constants Kr and Kc, with δr = Krε and a corresponding efficiency derivation.

B.2. Agent Training Configuration

Agent-Omit training uses environment-specific SFT and RL configurations, with adjustments for task complexity and dataset size. The reported practical guidance emphasizes limited SFT and RL training to preserve later policy stability.

  • SFT and RL hyperparameters are adjusted across five environments for task complexity, including learning rates, interaction turns, and rollout samples.
  • SFT converges in approximately 1 hour across agents, while DeepSearch and TextCraft RL require approximately 32 hours.
  • WebShop RL training extends to 36 hours because of its interactive environment and state complexity.
  • Excessive SFT can lead to collapse during RL, while extending RL to a second epoch degrades performance on complex tasks.
  • Figure 8 visualizes Qwen3-8B Agent-Omit RL training on WebShop.

C.1. SFT Training Visualization

SFT training adapts agents to omission behavior, producing an initial gradient spike followed by decreasing gradients and convergence near loss 0.6 across five tasks. The RL visualization further reports shorter trajectories and fewer task rounds as omission behavior develops.

  • The gradient norm spikes initially and then rapidly decreases as the agent adapts from its pretrained policy to omission behavior.
  • Loss converges to approximately 0.6 across five tasks, indicating that the omission policy is learnable.
  • RL training on WebShop is visualized for Qwen3-8B to analyze efficiency and behavioral evolution.
  • Trajectory length consistently decreases during training as the agent learns to omit redundant tokens and reduce token consumption.
  • The DeepSearch case study shows omitted tool responses, empty-thought turns, repeated searches, and a final answer after multiple search attempts.
  • The case study’s question asks how many episodes are in season 4 of The Affair, with the displayed final answer being 10.
Loading 2602.04284v2…