GenieReasoner: Unified Vision-Language Action Model

• GenieReasoner is a unified architecture that integrates vision, language, and action generation via an autoregressive Transformer for embodied robotic manipulation.
• It employs a flow-matching action tokenizer (FACT) to discretize continuous motor trajectories, ensuring sub-millimeter control fidelity.
• The architecture’s efficacy is validated through the ERIQ benchmark, demonstrating significant gains in reasoning accuracy and real-world task performance.

GenieReasoner denotes a unified architecture and conceptual paradigm for integrating high-level reasoning and precise action generation in embodied multimodal agents, particularly within open-world robotic manipulation contexts. The defining contribution of GenieReasoner is its simultaneous optimization of vision-language reasoning and motor control within a single autoregressive sequence-to-sequence Transformer backbone, leveraging a flow-matching-based action tokenizer (FACT) to bridge continuous motor trajectories with discrete token spaces. The architecture’s validity and value for benchmarking embodied reasoning capabilities are evidenced through the introduction of the Embodied Reasoning Intelligence Quotient (ERIQ), a large-scale diagnostic that isolates and quantifies the reasoning-precision trade-off, and by empirical gains over prior continuous and discrete action frameworks (Liu et al., 30 Dec 2025).

1. Architectural Integration of Vision, Language, and Action

GenieReasoner employs a single Transformer backbone receiving visual input (short video clips as sequences of frames), natural language instructions, and (during training) future action sequences corresponding to robotic end-effector states. Visual frames are embedded via a patch-based vision encoder; linguistic input is tokenized using the Transformer's word-piece scheme; and continuous action sequences are discretized into a compact code via the flow-matching action tokenizer (FACT). At each time step, the three modalities are concatenated into a single sequence of tokens and autoregressively modeled. During inference, the model receives only vision and language tokens and emits a sequence of action tokens, which the FACT decoder transforms back into continuous control signals for physical execution (Liu et al., 30 Dec 2025).

This integrated design enables tightly coupled learning of both high-level semantic reasoning (via vision-language alignment) and fine-grained motor precision, unified under a single predictive objective and token space.

2. Flow-Matching Action Tokenizer (FACT) and Action Reconstruction

The FACT tokenizer is critical for resolving the bottleneck between reasoning-rich, discrete representations and continuous, high-precision motor commands. The tokenizer comprises two modules:

• A VQ-style encoder $\mathcal{E}_\theta$ mapping action trajectories $a_{0:H}$ into a compact discrete code $c \in \{-1, +1\}^{L\times D}$ using a sign quantization function. The encoder operates on $L$ queries of dimension $D$ to produce a fixed-size code embeddable in the VLM token vocabulary.
• A flow-matching decoder $\mathcal{D}_\theta$ reconstructs trajectories from the discrete token $c$ plus time-indexed Gaussian noise, by solving an ODE $$\partial_t \hat{a}^{(t)} = \mathcal{D}_\theta(\hat{a}^{(t)}, c, t)$$ from $t=0$ (Gaussian noise) to $t=1$ (final trajectory).

The flow-matching loss

$$L_{\mathrm{flow}} = \mathbb{E}_{a, z \sim \mathcal{N}, \, t \sim U[0,1]} \left[ \| (a-z) - \mathcal{D}_\theta(a^{(t)},c,t) \|_2^2 \right]$$

ensures the decoder smoothly interpolates between noise and data. Auxiliary losses encourage codebook utilization and quantization stability. This mechanism provides sub-millimeter-level control fidelity with as few as $L\approx 20$ tokens, maintaining a compact token vocabulary (e.g., $2^{12}$ entries) (Liu et al., 30 Dec 2025).

3. Pre-Training Regime and Optimization Strategy

GenieReasoner adopts a phased training protocol:

1. FACT Pre-training: The FACT tokenizer is trained on action trajectories alone, optimizing the flow-matching, entropy, and commitment losses to ensure fidelity and code diversity.
2. Joint Pre-training: The Transformer backbone is trained to predict the next token on a mixture of general vision-language QA data, embodied VQA (drawing from the ERIQ dataset), and tokenized action sequences. All input modalities are fused as a single sequence; the model jointly learns semantic reasoning and action generation under a unified objective.
3. Post-training (Fine-Tuning): The model is further optimized on embodied VQA and action data alone to solidify embodied reasoning representations and action code alignment.

All components (vision, language, actions) are co-trained, enabling gradient flow through all branches except for the separately pre-trained FACT decoder (Liu et al., 30 Dec 2025).

4. Embodied Reasoning Intelligence Quotient (ERIQ) Benchmark

The ERIQ benchmark is developed to quantitatively decouple reasoning from execution, offering 6,052 multiple-choice and yes/no questions that interrogate four primary reasoning capabilities:

1. Spatial Perception & Grounding (scene understanding, relative position, etc.)
2. Planning & Monitoring (action understanding, progress tracking, trajectory analysis)
3. Error Detection & Recovery (recognizing and classifying mistakes, planning recovery)
4. Human Intent Understanding (inferring intention and human-robot interaction)

Data span five domains (household, restaurant, supermarket, industrial, office) and three visual modalities (single images, sequential images, interleaved image+text). ERIQ accuracy has predictive value for end-to-end generalization and task performance in robotic control (Liu et al., 30 Dec 2025).

5. Empirical Performance and Comparative Analysis

GenieReasoner achieves 82.72% average ERIQ accuracy versus the 58.64% baseline from Qwen2.5-VL-3B. Substantial gains are reported in specialized areas such as action understanding (96.7% vs 65.5%) and dual-view matching (+31%). On open VLM reference benchmarks, GenieReasoner meets or exceeds state-of-the-art results for 7–8B parameter models, suggesting generalization despite integrating embodied reasoning.

Ablation studies establish that:

• Models lacking embodied VQA and action data reach high ERIQ but 0% real-world success.
• Models with action alignment alone score lower on reasoning, supporting the necessity of joint pre-training.
• The combined regime achieves both high diagnostic accuracy (ERIQ) and 25–35% success in multi-arm pick-and-place simulation.

For real-world robot tasks, GenieReasoner outperforms continuous-only and prior discrete-only baselines in instruction following, grasp-and-place success rates, and a composite aggregate score, establishing closure of the reasoning-precision gap (Liu et al., 30 Dec 2025).

6. Addressing the Reasoning-Precision Bottleneck

Traditional discrete policies (e.g., VQ-VAE, FAST) either require massive vocabularies to support fine resolution or incur loss of trajectory fidelity, while continuous diffusion-based heads can disrupt semantic grounding in LLMs. GenieReasoner factors the problem such that:

• All reasoning and planning occur in a small, discrete space co-trained with vision-language representations.
• The precision requirement is offloaded to a separately pre-trained flow-matching decoder (FACT) without gradient entanglement.

This configuration enables the VLM to learn tight semantic grounding for actions without compromising trajectory resolution, eliminating the need for insulation between reasoning and control heads.

7. Summary and Broader Context

GenieReasoner introduces an architecture for joint multi-modal and motor control optimization, grounded in formal embodied reasoning diagnostics (ERIQ) and technical advances in action discretization (FACT). The approach enables robust, instruction-driven performance in realistic robotic manipulation tasks. It stands as a template for future vision-language-action systems that require co-optimization of reasoning and control, and its diagnostic and modular design provides a principled basis for further advances in embodied intelligence (Liu et al., 30 Dec 2025).