Research map / Reinforcement learning / Visual and representation RL

LAFP: Preserving Latent Action Structure in Latent Policy Learning via Flow Matching

NeurIPS 2026 · Reinforcement learning › Visual and representation RL · approach: Skill learning and diffusion policies · 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.

Replaces behavior cloning in LAOM's latent-policy distillation with conditional flow matching and adds an inference-time interpolation scheme for training the action decoder, evaluated on 16 Procgen games.

Problem

Distilling a pretrained latent action space into a latent policy via behavior cloning collapses multimodal action distributions, and naively applying flow matching makes latent-to-physical action decoding misaligned due to sampling stochasticity.

Approach

Use conditional flow matching as the latent policy so the pretrained latent action structure is preserved, freeze it during post-training, and train the action decoder on flow-generated latents obtained by interpolating from Gaussian noise toward the target latent action distribution to reduce stochasticity-induced misalignment.

Key results

• LAFP averages 62.6% success across 16 Procgen environments versus 54.4% for LAOM (reported as +8.1 points in the text).
• On Miner, LAFP reaches 87.0% versus 36.4% for LAOM in the appendix table.
• A 3-step LAFP inference budget averages 62.1% success, compared with 53.5% for LAOM behavior cloning in the step ablation.

Surprising findings

• Fine-tuning the flow-matching latent policy during post-training degrades performance (LAFP Fine-tuned averages 50.6% vs 62.6% frozen), while fine-tuning helps behavior cloning.
• Reusing the IDM-pretrained action decoder without retraining causes almost no loss for LAFP but substantially hurts LAOM (44.6% vs 53.9%).

Limitations

• Iterative flow sampling adds inference latency relative to one-step policies.
• Evaluated only in offline imitation learning on discrete-action Procgen environments.
• Fine-tuning the flow policy in post-training degrades performance, so the method relies on freezing the latent policy.

Experimental setup

Datasets: Procgen

Benchmarks: Procgen

Metrics: success rate

Baselines: LAOM LAOM (Frozen) LAOM behavior cloning Direct-Sample AD

Builds on: LAOM LAPO flow matching VQ-VAE

Closest papers by meaning

• Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations (NeurIPS 2025 · similarity 0.91)
• Act to See, See to Act: Diffusion-Driven Perception-Action Interplay for Adaptive Policies (NeurIPS 2025 · similarity 0.89)
• Latent Action Learning Requires Supervision in the Presence of Distractors (ICML 2025 · similarity 0.88)
• Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.88)
• Latent Diffusion Planning for Imitation Learning (ICML 2025 · similarity 0.87)
• Latent Action Reparameterization for Efficient Agent Inference (NeurIPS 2026 · similarity 0.87)