Fast-ThinkAct: Efficient Vision-Language-Action Reasoning via Verbalizable Latent Planning

AI summary

Overview

Research area: Vision-Language-Action (VLA) models for embodied AI, specifically efficient reasoning for robotic manipulation and embodied question answering.

Technical level: Advanced. The paper assumes familiarity with chain-of-thought reasoning, reinforcement learning fine-tuning (GRPO), knowledge distillation, diffusion policies, and latent reasoning.

Scope: The paper introduces Fast-ThinkAct, a framework that compresses lengthy textual chain-of-thought reasoning into compact "verbalizable" continuous latent tokens for VLA tasks, achieving up to 89.3% inference latency reduction over state-of-the-art reasoning VLAs while improving task success rates.

What This Paper Is About

Reasoning VLAs produce explicit chain-of-thought traces that improve generalization on robotic tasks, but these traces are long (roughly 250 tokens) and take several seconds per decision, conflicting with the 1-15 Hz control frequency that embodied applications require. Fast-ThinkAct's goal is to preserve the planning benefits of reasoning while making it compact enough for real-time control, by distilling explicit textual and visual planning from a teacher model into a small set of continuous latent tokens that a student model can generate quickly.

Key Contributions

1. Fast-ThinkAct framework: An efficient embodied reasoning framework that compresses reasoning into verbalizable latent thoughts while maintaining expressive planning ability.
2. Preference-guided distillation with manipulation trajectory alignment: A method that compresses both linguistic and visual planning into compact continuous latents, using reward preferences derived from the teacher's GRPO training to favor high-quality reasoning patterns over low-quality ones.
3. Reasoning-enhanced policy learning: A mechanism that bridges high-level visual planning to low-level action execution by feeding latent planning representations from the student VLM into an action model.
4. Efficiency and performance results: Up to 89.3% inference latency reduction over state-of-the-art reasoning VLAs while maintaining strong performance across embodied manipulation and reasoning benchmarks.

Main Findings

• Latency reduction: Fast-ThinkAct achieves 89.3% and 88.0% latency reduction compared to ThinkAct-7B and MolmoAct-7B respectively, and 7× faster inference than ThinkAct-3B. Figure 1 reports 9.3× faster inference than ThinkAct-7B while delivering improved performance on SimplerEnv-Google.
• Robot manipulation success: Fast-ThinkAct consistently outperforms all baselines on LIBERO subtasks (Spatial, Object, Goal, Long) and SimplerEnv-Google, beating foundation VLAs such as OpenVLA and reasoning VLAs including CoT-VLA, ThinkAct, and MolmoAct.
• Bimanual manipulation (RoboTwin2.0): Fast-ThinkAct reaches 65.7 average success on easy and 26.4 on hard settings, versus RDT (56.4 / 22.8), ThinkAct (62.4 / 24.7), π0 (52.9 / 16.3), ACT (45.5 / 3.5), and DP (43.1 / 0.6). Compared to RDT this is 9.3% and 3.6% higher on easy and hard; against ThinkAct it is 3.3% and 1.7% higher.
• Embodied reasoning: Fast-ThinkAct-3B scores 46.4 on EgoPlan-Bench2, 60.8 on RoboVQA, and 51.2 on OpenEQA (52.8 average), surpassing ThinkAct-3B (44.0 / 55.3 / 48.9; 49.4 average), RoboBrain2.0-3B (41.8 / 46.5 / 50.1; 46.1), and proprietary models GPT-4V and Gemini-2.5-Flash. It exceeds the runner-up by 2.4% on EgoPlan-Bench2, 5.5 BLEU on RoboVQA, and 1.1 points on OpenEQA.
• Long-horizon planning: On RoboTwin2.0 long-horizon tasks (average length exceeding 270 steps), Fast-ThinkAct scores 48.8 and 16.8 on easy and hard settings, surpassing RDT (35.0 / 12.3) and ThinkAct (42.8 / 15.3).
• Failure recovery: On RoboFAC, Fast-ThinkAct outperforms the second-best baseline RoboFAC-3B by 10.9 points on the simulation split and 16.4 points on the real-world split.
• Few-shot adaptation: Fine-tuning with only 10 demonstrations per task on RoboTwin2.0 significantly enhances the RDT action model and outperforms π0 and ThinkAct on medium and long-horizon tasks, while using lower reasoning latency than ThinkAct.
• Ablation results: Removing the verbalizer loss ℒ_verb drops the three-benchmark average from 52.8 to 48.5; further removing ℒ_distill drops it to 47.7. The textual teacher scores 49.8, SFT + CoT-SFT scores 45.0, and SFT only scores 46.5. CoT-SFT underperforms SFT on EgoPlan-Bench2 and RoboVQA but improves on OpenEQA.
• Verbalization quality: Comparing the teacher's textual reasoning to the student's verbalized latent reasoning on RoboVQA, the student produces more concise and focused responses, filtering out less relevant content while still capturing task-relevant information.

Methodology in Plain English

The approach uses a teacher-student setup built on the Qwen2.5-VL 3B backbone.

Teacher: A textual teacher VLM is initialized from the same SFT and CoT-SFT checkpoint as the student and is trained with GRPO using action-aligned visual rewards and QA-style rewards. Because GRPO produces multiple rollouts per input, the highest-advantage trace (τ⁺) and lowest-advantage trace (τ⁻) in each group form preference pairs that serve as a quality signal.

Student: Instead of generating text tokens, the student VLM autoregressively generates a small number of continuous latent vectors (M latents of hidden size d, described in Figure 1 as compact continuous tokens, e.g., 6) and additionally uses K learnable spatial tokens for parallel waypoint prediction. A verbalizer LLM (Qwen3-0.6B with cross-attention layers inserted at each layer) decodes the latents into natural language. A DPO-inspired objective (ℒ_verb, with β = 0.1) trains the verbalizer to assign higher likelihood to the high-quality trace than the low-quality trace, which in turn pushes the student's latents toward high-quality reasoning. A second loss (ℒ_distill) minimizes the L2 distance between the teacher's and student's hidden states at the <answer> token, transferring spatial planning. A third loss (ℒ_ans) supervises the K spatial tokens to predict ground-truth waypoints via an MLP, allowing parallel trajectory prediction rather than autoregressive waypoint text (the teacher tokenizes waypoints into 60-70 tokens when K = 5).

Policy learning: The visual latent plan c_t is extracted from the KV cache of the spatial tokens in earlier VLM layers and concatenated with KV pairs from the action model's state encoder, so cross-attention attends to both planning context and state observations. The student VLM and state encoder are frozen while only the action model π_φ is updated with an imitation learning objective. π_φ is initialized from DiT-Policy (pre-trained on OXE) for SimplerEnv and from RDT for LIBERO and RoboTwin2.0; a linear projection adapts the VLM KV cache to the action model dimension (1,024 for DiT-Policy, 2,048 for RDT).

Training details: SFT runs 1 epoch with batch size 64 and learning rate 1e-5, followed by CoT-SFT for 15K iterations. Teacher-student training runs 4,500 iterations with batch size 128 and learning rate 1e-6; the verbalizer is first warmed up for 3,000 iterations with standard language modeling loss, then frozen for the remaining 1,500 iterations using ℒ_verb. Policy training runs 20K iterations with batch size 256 and learning rate 1e-4. All experiments use 16 NVIDIA A100 GPUs with 80 GB memory.

Inference: Only the student VLM and the action model are required; the verbalizer is used only during training and optionally for interpretability.

Why This Matters

Impact on research: The paper shows that explicit, token-heavy chain-of-thought is not a requirement for reasoning-based VLA performance. By demonstrating that compact continuous latents can carry both linguistic and visual planning signal and even outperform longer textual reasoning, it challenges the assumption that reasoning length and reasoning quality must scale together. It also offers a concrete instantiation of latent reasoning adapted to embodied control, where prior LLM latent-reasoning techniques did not directly transfer.

Real-world applications:

• Real-time robotic manipulation, where control loops operate at 1-15 Hz and multi-second reasoning delays are impractical.
• Bimanual and long-horizon assembly or household tasks, such as the RoboTwin2.0 handover and stacking tasks requiring coordinated two-arm control.
• Failure detection and recovery in manipulation, where a robot must notice a dropped object mid-execution and generate corrective steps.
• Fast adaptation to new environments or tasks with limited demonstrations (10 demonstrations per task in the reported few-shot experiments).

Industry relevance: The framework is developed by NVIDIA researchers and targets embodied AI deployment, where inference cost directly determines whether a model can run at control frequency. The reported latency reductions (up to 89.3% versus ThinkAct-7B and 88.0% versus MolmoAct-7B) speak to serving-cost and throughput concerns, and the use of A100-class hardware plus existing action model backbones (DiT-Policy, RDT) makes the approach compatible with common robotics stacks.

Future Directions

• Reducing hallucination in verbalization: The verbalizer is built on a pre-trained LLM and inherits language model limitations including hallucination, occasionally producing plausible but inaccurate descriptions. The authors suggest incorporating grounding-aware objectives or hallucination suppression techniques. They note this does not affect action execution at inference, since the verbalizer serves only for interpretability.
• Improving verbalized reasoning faithfulness: Beyond suppressing hallucination, making the decoded text a more faithful readout of the latent plan remains open.
• Extending latent reasoning to settings not covered: The paper evaluates only the benchmarks listed (LIBERO, SimplerEnv, RoboTwin2.0, EgoPlan-Bench2, RoboVQA, OpenEQA, RoboFAC); additional real-world deployment validation is not reported.
• Additional ablations: The paper states that further ablation studies are provided in the supplementary material, indicating the current analysis is partial.

Target Audience

Researchers and engineers working on embodied AI, robotic manipulation, and vision-language-action models who need reasoning capabilities under real-time constraints. It is also relevant to those studying efficient reasoning, chain-of-thought distillation, and latent reasoning in multimodal models, as well as practitioners deploying robot policies that must run at control frequencies where multi-second inference is unacceptable.

Authors’ abstract

Vision-Language-Action (VLA) tasks require reasoning over complex visual scenes and executing adaptive actions in dynamic environments. While recent studies on reasoning VLAs show that explicit chain-of-thought (CoT) can improve generalization, they suffer from high inference latency due to lengthy reasoning traces. We propose Fast-ThinkAct, an efficient reasoning framework that achieves compact yet performant planning through verbalizable latent reasoning. Fast-ThinkAct learns to reason efficiently with latent CoTs by distilling from a teacher, driven by a preference-guided objective to align manipulation trajectories that transfers both linguistic and visual planning capabilities for embodied control. This enables reasoning-enhanced policy learning that effectively connects compact reasoning to action execution. Extensive experiments across diverse embodied manipulation and reasoning benchmarks demonstrate that Fast-ThinkAct achieves strong performance with up to 89.3% reduced inference latency over state-of-the-art reasoning VLAs, while maintaining effective long-horizon planning, few-shot adaptation, and failure recovery.