Research map / Multi-agent systems and games / Human-AI and goal-conditioned

Latent Action Reparameterization for Efficient Agent Inference

NeurIPS 2026 · Multi-agent systems and games › Human-AI and goal-conditioned · approach: Language-grounded reinforcement agents · 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 Latent Action Reparameterization (LAR), which learns compact latent action tokens for recurring low-entropy multi-step agent action segments and distills them into a LoRA-adapted LLM agent to shorten the effective decision horizon while keeping task success.

Problem

LLM agents emit long sequences of fine-grained textual action tokens, making decision horizons long and inference costly; existing efficiency methods leave the action representation unchanged.

Approach

Learn a latent action vocabulary from agent trajectories by selecting frequent, low-next-token-entropy n-gram segments, add them as new tokens, and train a LoRA student on reparameterized trajectories with KL distillation from the original-model teacher over content positions shared by both sequences.

Key results

• On Qwen3-8B, LAR reaches 80.09 on TriviaQA (vs. 67.40 Vanilla) with 27.1% fewer action tokens.
• On Llama-3.1-8B-Instruct Mind2Web, LAR reaches 28.30 vs. 24.40 Vanilla with 20.8% fewer action tokens.
• On OpenClaw TriviaQA, the Short setting raises EM from 0.4218 to 0.5358 (27.0% relative) while compressing 6.7% of the static prompt.

Surprising findings

• Abstracting low-entropy action segments improves task performance rather than degrading it, and a padded control (LAR-PT) still beats ReAct.
• Abrupt performance collapse when high-entropy parameter-binding segments such as search queries are abstracted.

Limitations

• Gains depend on the presence of compressible structural scaffolding; reasoning-heavy tasks admit smaller compressible subsets.
• Abstracting high-entropy parameter-binding content causes abrupt collapse, so thresholds must be set conservatively.
• Latent vocabularies are built per model and per benchmark, limiting cross-model transfer.

Experimental setup

Datasets: TriviaQA KodCode Mind2Web Musique MBPP HumanEval

Benchmarks: TriviaQA KodCode Mind2Web Musique MBPP HumanEval

Metrics: accuracy success rate action token reduction token throughput peak GPU memory exact match

Baselines: Vanilla CoT ReAct TokenSkip ACON ConciseHint

Builds on: ReAct Search-R1 VeRL GRPO

Closest papers by meaning

• Controlling Large Language Model with Latent Action (ICML 2025 · similarity 0.89)
• Reinforcing LLM Agents via Policy Optimization with Action Decomposition (NeurIPS 2024 · similarity 0.89)
• Enhancing Decision-Making for LLM Agents via Step-Level Q-Value Models (AAAI 2025 · similarity 0.88)
• Learning Agentic Policy from Action Guidance (NeurIPS 2026 · similarity 0.88)
• AgentRefine: Enhancing Agent Generalization through Refinement Tuning (ICLR 2025 · similarity 0.88)
• AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.88)