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

Conference: ICML 2026
arXiv: 2601.15197
Code: The paper mentions "Code and videos are available," but no repository URL is provided.
Area: Robotics / Embodied AI / VLA
Keywords: VLA, Visual shortcuts, Bayesian decomposition, Latent Action Queries, Pointwise Mutual Information

TL;DR

LangForce reformulates the VLA policy as a Bayesian decomposition \(\pi(a\mid v,\ell)=p(\ell\mid a,v)\,p(a\mid v)/p(\ell\mid v)\). It introduces learnable Latent Action Queries to execute "vision-only" and "vision+language" branches simultaneously on shared VLM weights. By maximizing the log-likelihood ratio between actions and instructions to penalize "visual shortcuts," it achieves an 11.3 absolute point improvement over the QwenGR00T baseline on SimplerEnv.

Background & Motivation

Background: Current mainstream VLAs (OpenVLA, π0, GR00T, StarVLA series) attach diffusion action heads to pretrained VLMs and perform imitation learning on large-scale human demonstrations. They aim to ground natural language instructions into continuous actions using the world knowledge within the VLM.

Limitations of Prior Work: The authors find that these models frequently collapse in OOD scenarios and multi-task ambiguous settings. In the LIBERO Goal benchmark (multiple tasks for the same tabletop), a "vision-only" model—fed only visual input without instructions—achieves success rates close 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 makes the mapping \(v\to\ell\) almost injective—seeing a cabinet implies "open the cabinet." This results in a conditional entropy \(H(\ell\mid v)\approx 0\), which further drives the conditional mutual information \(I(\ell;a\mid v)\le H(\ell\mid v)\) towards zero. The model learns \(\pi(a\mid v,\ell)\approx p(a\mid v)\), a phenomenon the authors call "information collapse."

Goal: Force the policy to truly depend on language during training without re-collecting data or increasing inference computation.

Key Insight: Rewrite the posterior using Bayes' theorem \(\pi(a\mid v,\ell)=p(\ell\mid a,v)\,p(a\mid v)/p(\ell\mid v)\) and use the log-likelihood ratio \(\log p(\ell\mid a,v)-\log p(\ell\mid v)\) (i.e., conditional PMI) as a regularizer to reward policies where instructions can be inferred from actions.

Core Idea: Utilize a shared set of VLM weights and learnable Latent Action Query tokens. Using decoder-only causal masking, the model simultaneously simulates the "vision prior branch" and the "vision+language posterior branch." The difference between the language log-probabilities of these two branches is explicitly optimized as a PMI reward.

Method

Overall Architecture

LangForce adds three components to a native VLA (specifically QwenGR00T based on StarVLA, using Qwen3-VL-4B and a DiT action head): (1) \(K=64\) new tokens \(\mathcal{Q}=\{\langle\text{action}_1\rangle,\dots,\langle\text{action}_K\rangle\}\) in the vocabulary as Latent Action Queries; (2) A dual-branch setup within the same batch using \([v, \mathcal{Q}, \ell]\) and \([v, \ell, \mathcal{Q}]\) token sequences; (3) A total loss combining flow-matching action loss from both branches and a Language Log-Likelihood Ratio (LLR) term. Only the posterior branch is used during inference, maintaining the same overhead as a standard VLA.

Key Designs

1. Latent Action Queries as Information Bottleneck:

   • Function: Condenses action-related semantics from VLM outputs into \(K\) fixed-length latent states \(\mathbf{H}_{\mathcal{Q}}\in\mathbb{R}^{K\times D}\) before feeding them to the DiT head.
   • Mechanism: Adds \(K=64\) learnable tokens to the VLM vocabulary and embedding table. The DiT iterates only on \(\mathbf{H}_{\mathcal{Q}}\) rather than the full hidden states. By placing \(\mathcal{Q}\) before \(\ell\) (prior) or after \(\ell\) (posterior) with causal masking, the model instantiates two conditional distributions using the same weights.
   • Design Motivation: Reduces DiT attention complexity from \(O(N^2)\) to \(O(K^2)\) and provides a structural basis for the switchable conditions in Bayesian decomposition.

2. Dual-branch Bayesian Training:

   • Function: Estimates the visual prior \(p(a\mid v)\) and the full conditional posterior \(\pi(a\mid v,\ell)\) simultaneously.
   • Mechanism: The prior branch takes \(\text{Input}_{\text{prior}}=[v,\mathcal{Q},\ell]\), where \(\mathcal{Q}\) only sees \(v\) due to causal masking, resulting in \(\mathbf{H}_{\mathcal{Q}}^{\text{prior}}\). To prevent the VLM backbone from internalizing visual shortcuts, a stop-gradient is applied to \(\mathbf{H}_{\mathcal{Q}}^{\text{prior}}\) so the prior loss only updates the DiT. The posterior branch takes \(\text{Input}_{\text{post}}=[v,\ell,\mathcal{Q}]\), resulting in \(\mathbf{H}_{\mathcal{Q}}^{\text{post}}\).
   • Design Motivation: Direct training on goal-driven data collapses \(\pi\) into \(p(a\mid v)\). Placing these distributions in explicit contrast allows the use of PMI to widen the gap. Shared weights and prefix prefilling minimize training overhead.

3. LLR / PMI Language Reward:

   • Function: Penalizes actions that contain no language information, forcing \(\mathbf{H}_{\mathcal{Q}}\) to carry features capable of inferring \(\ell\).
   • Mechanism: Uses the VLM's own LM loss as an approximation for \(\log p(\ell\mid\cdot)\). The objective is \(\mathcal{L}_{\text{LLR}}=\log p(\ell\mid v,\mathbf{H}_{\mathcal{Q}}^{\text{prior}})-\mathrm{sg}(\log p(\ell\mid v))\). The final loss is \(\mathcal{L}_{\text{total}}=(1-\lambda)\mathcal{L}_{\text{FM}}^{\text{post}}+\lambda\mathcal{L}_{\text{FM}}^{\text{prior}}-\beta\mathcal{L}_{\text{LLR}}\), with \(\lambda=0.3\) and \(\beta=0.1\).
   • Design Motivation: Maximizing PMI \(\log[\pi(a\mid v,\ell)/p(a\mid v)]\) is implementationally equivalent to this log-likelihood difference. The stop-gradient prevents the model from "gaming" the reward by degrading the denominator.

Loss & Training

Trained on 8 H100 GPUs using AdamW (lr=1e-5 + cosine), DeepSpeed ZeRO-2, and grad clip 1.0. Fine-tuned for 50k steps on BridgeData V2 + Fractal on SimplerEnv (batch 16/GPU).

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
Ours (LangForce)	89.6	63.8	33.3	79.2	66.5

RoboCasa GR1 Tabletop (24 tasks): LangForce 52.6% vs. QwenGR00T 47.8%.

Real-world Franka Pick-and-Place (OOD red cube): LangForce 9/30 vs. QwenGR00T 2/30. 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 provides the primary gain: Adding queries alone improves the average by 2.3 points, but PMI adds another 9.0 points.
• LangForce preserves the VLM's general capabilities: While QwenGR00T outputs gibberish on text math problems after fine-tuning, LangForce remains communicative, as the LLR term protects language representations.
• Improvements are concentrated in tasks requiring language-based target disambiguation (Eggplant +15, Carrot +13.6), rather than fine manipulation (Stack Block).

Highlights & Insights

• Formulates the engineering issue of "ignored instructions" through information theory: \(H(\ell\mid v)\approx 0\Rightarrow I(\ell;a\mid v)=0\), identifying PMI as a mathematically sound target.
• Uses decoder-only causal masking and token reordering to switch conditional distributions "for free" during training.
• Utilizing stop-gradients to fix the language baseline \(p(\ell\mid v)\) is a critical trick to prevent the model from cheating the contrastive reward.

Limitations & Future Work

• Dual-branch training increases memory and compute costs, though mitigated by prefix prefilling.
• Real-world experiments focus on pick-and-place; the method's advantages might be less pronounced in high-contact or dexterous manipulation tasks that rely more on low-level control than language grounding.
• The LLR loss depends on the VLM's LM head; \(\beta\) might need re-tuning if the instruction style or tokenizer changes significantly.

Related Work & Insights

• vs BayesVLA: Both use Bayesian decomposition, but BayesVLA requires a two-stage process (train prior, freeze, train posterior). LangForce is single-stage end-to-end.
• vs π0 / GR00T: Instead of feeding all hidden states (\(O(N^2)\)), LangForce compresses info into 64 queries and emphasizes "signal distillation" through contrastive regularization.
• vs ChatVLA: While ChatVLA uses task-routing for decoupling, LangForce embeds language dependency directly into the loss, providing inherent protection against forgetting.

Rating

• Novelty: ⭐⭐⭐⭐ Applies information collapse and PMI to VLA via dual-branch training; solid logic though individual tricks (stop-gradient, shared weights) are known.
• Experimental Thoroughness: ⭐⭐⭐⭐ Extensive coverage spanning SimplerEnv, RoboCasa, LIBERO, and real-world Franka tasks.
• Writing Quality: ⭐⭐⭐⭐ Clear motivation supported by pilot experiments; methodology is logically organized.
• Value: ⭐⭐⭐⭐ Provides a practical training-side fix for the common "VLA ignoring language" problem with low implementation cost 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