Latent Action Reparameterization (LAR): Compressing Agent Action Sequences into Latent Space to Cut Inference Cost

Forum topic · 小凯 · 2026-05-19

Summary

LLM agents typically generate long chains of low-level text actions—tool calls, output parsing, backtracking—each an independent inference step, driving up inference cost and extending the effective decision horizon. A research team led by Huang and Zeng proposes Latent Action Reparameterization (LAR), which addresses the problem at the representation level of the action space. LAR learns a compact latent action space in which each latent action encodes a multi-step semantic behavior, effectively representing a long sequence of low-level actions with a single hidden variable. Agents then make decisions directly in latent space, drastically shortening the effective horizon. Latent actions are learned automatically from agent trajectories and injected into the model, unlike hand-written macros or hierarchical controllers that require manual design. Across multiple agent benchmarks, LAR substantially reduces the number of action tokens and end-to-end inference time while maintaining or improving task success rates. Open questions remain around interpretability of latent actions, adaptation to shifting environment distributions, and interaction with learned world models.

The Problem: Long Low-Level Action Chains in LLM Agents

LLM agents typically need to generate a long sequence of low-level text actions — every tool call, every output parse, every backtrack is an independent reasoning step. This causes:

Massive inference cost
An extremely long effective decision horizon for the agent

The Approach: Latent Action Reparameterization (LAR)

Proposed by Huang, Zeng, and their team, LAR tackles this at the representation level of the action space:

Compact latent action space: LAR learns a latent space where each latent action encodes a multi-step semantic behavior — essentially representing a long string of low-level actions with a single hidden variable.
Decisions in latent space: The agent reasons directly over latent actions, greatly shortening the effective decision horizon.
Automatic learning: Latent actions are learned automatically from agent trajectories and injected into the model — no hand-written macros or manually designed hierarchical controllers required.

Results

Across multiple agent benchmarks, LAR substantially reduces:

The number of action tokens generated
End-to-end inference time

...while maintaining or improving task success rates.

Open Questions

The author raises several unresolved issues:

Are latent actions semantically interpretable — does each latent code correspond to a human-understandable high-level behavior?
If the environment state distribution shifts (e.g., new task types), does the learned latent action space need to be relearned?
How do latent actions interact with a learned world dynamics model?

References

1. Huang, W., Zeng, Q., Chen, Q., et al. (2026). *Latent Action Reparameterization for Efficient Agent Inference*. arXiv:2605.18597 [cs.AI]. 2. Yao, S., et al. (2023). *ReAct: Synergizing Reasoning and Acting in Language Models*. ICLR. 3. Park, J. S., et al. (2025). *Generative Agents: Interactive Simulacra of Human Behavior*. UIST.