LangForce: Bayesian Decomposition of Vision-Language-Action Models via Latent Action Queries

Conference: ICML 2026 arXiv: 2601.15197
Code: The paper notes "Code and videos are available," but the actual repository URL is not provided.
Area: Robotics / Embodied AI / VLA Keywords: VLA, visual shortcuts, Bayesian decomposition, latent action queries, pointwise mutual information

TL;DR

LangForce formulates the VLA policy as a Bayesian decomposition \(\pi(a\mid v,\ell)=p(\ell\mid a,v)\,p(a\mid v)/p(\ell\mid v)\). By introducing learnable Latent Action Queries, it executes both "vision-only" and "vision+language" branches using a single set of VLM weights. It explicitly penalizes "visual shortcuts" by maximizing the log-likelihood ratio between actions and instructions, achieving an 11.3 percentage point absolute improvement over the QwenGR00T baseline on SimplerEnv.

Background & Motivation

Background: Current mainstream VLAs (such as OpenVLA, π0, GR00T, and StarVLA series) attach pre-trained VLMs to diffusion action heads and perform imitation learning on large-scale human demonstrations. These models aim to ground natural language instructions into continuous actions using the world knowledge inherent in VLMs.

Limitations of Prior Work: The authors observe that these models frequently fail in OOD (Out-of-Distribution) scenarios and multi-task ambiguous settings. In LIBERO Goal (where a single tabletop setup corresponds to multiple tasks), a "vision-only" model that ignores language instructions achieves success rates nearly identical to the full model (e.g., 44.6 vs 47.8 on RoboCasa; action loss of 0.13 vs 0.08 on BridgeData+Fractal). This suggests that instructions play a minimal role during training.

Key Challenge: Goal-driven data collection often results in a mapping where \(v\to\ell\) is nearly injective—seeing a cabinet implies "open the cabinet," and seeing a bottle implies "pick up the bottle." This leads to a conditional entropy of \(H(\ell\mid v)\approx 0\), which in turn drives the conditional mutual information \(I(\ell;a\mid v)\le H(\ell\mid v)\) toward zero. Consequently, the model only learns \(\pi(a\mid v,\ell)\approx p(a\mid v)\), a phenomenon the authors define as "information collapse."

Goal: To force the strategy's true dependence on language to emerge from the training objective without re-collecting data or increasing inference computation.

Key Insight: Utilize the Bayesian formula \(\pi(a\mid v,\ell)=p(\ell\mid a,v)\,p(a\mid v)/p(\ell\mid v)\) to decompose the posterior into "visual prior \(\times\) likelihood / language marginal." The log-likelihood ratio \(\log p(\ell\mid a,v)-\log p(\ell\mid v)\) (i.e., conditional PMI) is then used as a regularizer to reward strategies where instructions can be accurately inferred from the actions.

Core Idea: Use a shared set of VLM weights combined with learnable latent action query tokens. By leveraging decoder-only causal masking, the model simultaneously simulates a "visual prior branch" and a "vision+language posterior branch," explicitly optimizing the difference between the language log-probabilities of these two branches as a PMI reward.

Method

Overall Architecture

LangForce augments a native VLA (specifically QwenGR00T based on StarVLA, using Qwen3-VL-4B and a DiT action head) with three components: (1) \(K=64\) new tokens added to the vocabulary \(\mathcal{Q}=\{\langle\text{action}_1\rangle,\dots,\langle\text{action}_K\rangle\}\) as latent action queries; (2) dual branches sharing weights within the same batch, constructed using two token sequences: \([v,\mathcal{Q},\ell]\) and \([v,\ell,\mathcal{Q}]\); (3) a total loss combining the flow-matching action loss from both branches with a language log-likelihood ratio (LLR) term. During inference, only the posterior branch is executed, maintaining the same computational cost as a standard VLA.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["Input: vision v + instruction ℓ"] --> Q["Latent Action Queries 𝒬<br/>64 learnable tokens as info bottleneck"]
    subgraph DUAL["Dual-Branch Bayesian Training (Shared VLM Weights)"]
        direction TB
        Q --> PR["Prior Branch [v,𝒬,ℓ]<br/>𝒬 sees vision only → H𝒬prior"]
        Q --> PO["Posterior Branch [v,ℓ,𝒬]<br/>𝒬 sees vision+lang → H𝒬post"]
    end
    PR --> LLR["LLR / PMI Language Reward<br/>log p(ℓ|v,H𝒬prior) − sg log p(ℓ|v)"]
    PR --> FM["Flow-matching Action Loss<br/>Prior + Posterior branches"]
    PO --> FM
    FM --> TOT["Total Loss<br/>(1−λ)L_FM_post + λL_FM_prior − βL_LLR"]
    LLR --> TOT
    PO -->|Inference uses posterior only| OUT["DiT generates action a"]

Key Designs

1. Latent Action Queries as Information Bottleneck: Compressing Action Semantics into 64 Conditional Tokens Models like π0/GR00T feed all vision-language hidden states into the action head, resulting in \(O(N^2)\) DiT attention complexity and lacking a clean interface for condition switching. LangForce adds \(K=64\) learnable tokens \(\mathcal{Q}=\{\langle\text{action}_1\rangle,\dots,\langle\text{action}_K\rangle\}\) to the VLM vocabulary. These tokens condense action-related semantics into fixed-length hidden states \(\mathbf{H}_{\mathcal{Q}}\in\mathbb{R}^{K\times D}\). The DiT only processes these \(K\) queries, reducing attention complexity to \(O(K^2)\). Crucially, by using decoder-only causal masking, placing \(\mathcal{Q}\) before \(\ell\) restricts it to vision only, while placing it after \(\ell\) allows it to see both vision and language. This architectural choice allows the same weights to instantiate both "prior" and "posterior" conditions, providing the structural basis for Bayesian decomposition.

2. Dual-Branch Bayesian Training: Explicit Contrast Between Visual Prior and Full Posterior Under goal-driven data, training \(\pi(a\mid v,\ell)\) directly often collapses into \(p(a\mid v)\). To fix this, both distributions must be estimated and compared. LangForce runs two branches with shared weights: the Prior Branch takes \([v,\mathcal{Q},\ell]\) as input, where \(\mathcal{Q}\) only sees \(v\) to produce \(\mathbf{H}_{\mathcal{Q}}^{\text{prior}}\). A stop-gradient is applied to \(\mathbf{H}_{\mathcal{Q}}^{\text{prior}}\) so that the prior loss only updates the DiT, preventing the VLM backbone from internalizing visual shortcuts. The Posterior Branch takes \([v,\ell,\mathcal{Q}]\) as input, where \(\mathcal{Q}\) sees both vision and language to produce \(\mathbf{H}_{\mathcal{Q}}^{\text{post}}\) for standard flow-matching. The Rectified Flow Matching loss is applied to both:

\[\mathcal{L}_{\text{FM}}(\psi;\mathbf{C})=\mathbb{E}\|v_\psi(\mathbf{a}_t,t,\mathbf{C})-(\mathbf{a}_1-\mathbf{a}_0)\|^2\]

This contrastive setup introduces minimal overhead due to weight sharing and prefix prefilling, while explicitly exposing the difference between prior and posterior for PMI optimization.

3. LLR / PMI Language-side Reward: Penalizing Actions that Lack Instruction Information To force the policy to depend on language, the authors maximize the pointwise mutual information \(\log[\pi(a\mid v,\ell)/p(a\mid v)]\). This is implemented as a log-likelihood ratio, approximated using the VLM's own LM loss:

\[\mathcal{L}_{\text{LLR}}=\log p(\ell\mid v,\mathbf{H}_{\mathcal{Q}}^{\text{prior}})-\mathrm{sg}\big(\log p(\ell\mid v)\big)\]

The numerator encourages \(\ell\) to attend back to \(\mathcal{Q}\) (which contains action tokens), while the denominator represents the vision-only marginal, which is frozen with a stop-gradient. This freezing is critical; otherwise, the model might "cheat" by degrading the denominator, which would destroy the VLM's general language capabilities. The final loss is:

\[\mathcal{L}_{\text{total}}=(1-\lambda)\mathcal{L}_{\text{FM}}^{\text{post}}+\lambda\mathcal{L}_{\text{FM}}^{\text{prior}}-\beta\mathcal{L}_{\text{LLR}}\]

The authors set \(\lambda=0.3\) and \(\beta=0.1\). By rewarding policies where instructions can be inferred from actions, the model is forced to utilize instructions, all while maintaining standard VLA inference costs.

Loss & Training

The model was trained on 8 H100 GPUs using AdamW (lr=1e-5 + cosine), DeepSpeed ZeRO-2, and a gradient clip of 1.0. Fine-tuning on BridgeData V2 + Fractal was performed for 50k steps (batch size 16 per card) for SimplerEnv. Inference uses only the posterior branch, matching baseline costs.

Key Experimental Results

Main Results

SimplerEnv (WidowX, Avg@480 Success Rate, %):

Method	Put Spoon	Put Carrot	Stack Block	Eggplant	Average
OpenVLA-OFT	34.2	30.0	30.0	72.5	41.8
CogACT	71.7	50.8	15.0	67.5	51.3
π0	29.2	62.5	29.2	91.6	53.1
π0.5	49.3	64.7	44.7	69.7	57.1
Isaac-GR00T-N1.6	64.5	65.5	5.5	93.0	57.1
QwenGR00T (baseline)	87.5	50.0	29.2	54.2	55.2
LangForce	89.6	63.8	33.3	79.2	66.5

RoboCasa GR1 Tabletop 24 task average success rate: LangForce 52.6 vs QwenGR00T 47.8 vs Isaac-GR00T-N1.5 48.2 vs VisionOnly 44.7.

Real-world Franka Pick-and-Place (OOD 1 red cube): LangForce 9/30 vs QwenGR00T 2/30 vs π0.5 7/30. Vegetable sorting total: LangForce 97/120 (80.8%) vs QwenGR00T 71/120 (59.2%).

Ablation Study

Configuration	Spoon	Carrot	Stack	Eggplant	Avg
QwenGR00T (Baseline)	87.5	50.0	29.2	54.2	55.2
+ Latent Action Queries	74.6	58.3	29.2	67.9	57.5
Full LangForce	89.6	63.8	33.3	79.2	66.5

Key Findings

• The dual-branch + LLR setup is the primary driver of performance: Adding queries alone only improved the result by 2.3 points, whereas the PMI loss added another 9.0 points. This confirms that the "Bayesian decomposition + contrastive loss" is the core innovation, not just the head architecture.
• LangForce preserves the VLM's general dialogue and reasoning capabilities: While QwenGR00T fine-tuned for VLA outputted repetitive gibberish on text-based math problems, LangForce could still solve them, indicating the LLR term protects language representations.
• Improvements are concentrated in tasks requiring language-based target differentiation (Eggplant +15, Carrot +13.6). Limited gains were seen in fine manipulation tasks (Stack Block), consistent with the goal of improving grounding rather than low-level control.

Highlights & Insights

• Formalizes the engineering phenomenon of "ignored instructions" into information theory: \(H(\ell\mid v)\approx 0\Rightarrow I(\ell;a\mid v)=0\), making PMI a principled target for correction.
• Employs a clever engineering trick using decoder-only causal masking and token reordering to switch conditional distributions "at no cost."
• The use of stop-gradients to pin the language baseline \(p(\ell\mid v)\) is a vital technique seen in other contrastive losses (DPO, PPO ratio); it prevents the model from gaming the reward by degrading the denominator.

Limitations & Future Work

• Training requires running two branches. Although prefix prefilling mitigates redundant visual computation, there is still an increased memory and compute cost compared to single-branch baselines.
• Real-world experiments focused on pick-and-place and did not address high-contact or dexterous manipulation. The authors admit the advantage may be less pronounced in tasks dominated by low-level control rather than language grounding.
• The LLR loss relies on the VLM's own LM head as an approximation for \(\log p(\ell\mid\cdot)\); quality is thus sensitive to the VLM tokenizer and instruction style, potentially requiring \(\beta\) retuning for different templates.

Related Work & Insights

• vs BayesVLA: Both use Bayesian decomposition, but BayesVLA requires a two-stage process (training the visual prior, then freezing it to train the posterior). LangForce enables single-stage end-to-end training via shared weights and token reordering.
• vs π0 / GR00T: These models feed all VLM hidden states to the action head (\(O(N^2)\)). LangForce compresses information into 64 queries and uses prior/posterior contrast as regulation, emphasizing "signal refinement" over "context volume."
• vs ChatVLA: While both care about preserving language ability post-VLA training, ChatVLA uses task routing/decoupling, whereas LangForce embeds the language dependency directly into the loss via the LLR term, providing inherent protection against forgetting.

Rating

• Novelty: ⭐⭐⭐⭐ Maps information collapse to PMI via a dual-branch VLA implementation; clear logic, though individual tricks (stop-gradient, shared weights) are known.
• Experimental Thoroughness: ⭐⭐⭐⭐ Comprehensive coverage across SimplerEnv, RoboCasa, LIBERO, and real-world Franka tasks, including general capability visualization.
• Writing Quality: ⭐⭐⭐⭐ Motivation is well-supported by pilot experiments, and the Method section follows the Bayesian derivation strictly.
• Value: ⭐⭐⭐⭐ Provides a practical training-side fix for the common "VLA instruction neglect" issue with low replication costs for engineering teams.

Related Papers

• [ICML 2026] Latent Reasoning VLA: Latent Thinking and Prediction for Vision-Language-Action Models
• [ICML 2026] Neural Implicit Action Fields: From Discrete Waypoints to Continuous Functions for Vision-Language-Action Models
• [ICML 2026] StableVLA: Towards Robust Vision-Language-Action Models without Extra Data
• [ICML 2026] Spatial Memory for Out-of-Vision Manipulation in Vision-Language-Action
• [ICML 2026] Contrastive Representation Regularization for Vision-Language-Action Models