## Does Hallucination in World Models Actually Break VLA Training — and Can HaWMPO Fix It?

A 15.0 percentage-point gain over a base [Vision-Language-Action Model](https://humanoidintel.ai/glossary/vision-language-action-model) on the LIBERO benchmark — without touching a single real robot during policy optimization — is the headline result from HaWMPO, a new pipeline out of a four-person academic team published to arXiv today. On a physical Unitree G1 humanoid, the same method pushes average manipulation success from 67.5% to 80.0% across two evaluated tasks. Those numbers are specific, benchmarked, and hardware-validated, which puts this paper several rungs above the average world-model proposal that never leaves simulation.

The core problem HaWMPO attacks is underappreciated in most VLA post-training discussions: world models used for imagined rollouts don't fail uniformly. They hallucinate — generating plausible-looking but physically incoherent state transitions — and those hallucinated transitions are fed back into policy gradient updates as if they were ground truth. In long-horizon [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) tasks, compounding hallucination error has been quietly sabotaging otherwise solid architectures. This paper names the failure mode explicitly and proposes a closed-loop mechanism to suppress it.

---

## What HaWMPO Actually Does

The architecture has two interlocking components. First, an action-conditioned hallucination-aware model estimates the reliability of each generated image sequence — essentially scoring how physically plausible the imagined rollout is at each step. Second, those hallucination scores are injected into the policy gradient update through what the authors call a **Reward-Soft mechanism**, which modulates group relative policy optimization (GRPO) to down-weight action chunks that originated from low-reliability imagined transitions.

The framing matters: this is not a world model paper per se. HaWMPO doesn't propose a new generative architecture for imagining future states. It proposes a quality-control layer *on top of* existing world model rollouts, making the post-training pipeline aware of its own unreliability. That's a more tractable and immediately deployable contribution than yet another novel world model design.

The pipeline is explicitly positioned as a VLA *post-training* method — meaning it assumes you already have a pretrained generalist policy and want to improve it via synthetic experience without the cost and safety exposure of online real-robot RL. This is a design choice with significant commercial implications. Hardware-in-the-loop RL at scale requires either large robot fleets or significant downtime tolerance; a world-model-based alternative that can run on GPU clusters is far easier to operationalize.

---

## Benchmark Results and Why They Hold Up

On LIBERO — a widely used manipulation benchmark covering object interaction across diverse scenarios — HaWMPO achieves the best reported average success rate among compared methods, with a **+15.0% gain over the base model** and a **+2.8% gain over the strongest baseline**. The authors don't identify which specific baseline represents the +2.8% ceiling, so that margin should be read as competitive-but-not-dominant against the existing literature rather than a rout.

The more credible signal is the G1 real-world transfer. Moving from 67.5% to 80.0% on two physical manipulation tasks validates that the hallucination-suppression mechanism isn't simply overfitting to LIBERO's simulated dynamics. [Sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) for VLA-based manipulation is still fragile enough that any paper showing consistent real-hardware gains on a full-sized humanoid deserves attention, even when the evaluated task count is small.

**What to be skeptical about:** Two real-world tasks is a thin validation set. The paper doesn't specify task complexity, object diversity, or the number of evaluation trials — all variables that would significantly affect how much weight to put on the 67.5% → 80.0% lift. The LIBERO gains are more statistically robust given the benchmark's standardized structure, but replication on more varied physical platforms remains an open question.

---

## Why This Matters for the VLA Post-Training Race

The dominant approach to improving generalist robot policies in 2026 has bifurcated: either scale demonstration data (the [Physical Intelligence](https://humanoidintel.ai/companies/physical-intelligence) and [Skild AI](https://humanoidintel.ai/companies/skild-ai) direction) or use online RL to fine-tune deployed robots (the approach several hardware-first teams are pursuing). HaWMPO represents a third path — synthetic post-training with explicit reliability filtering — that could sit between those two paradigms in both cost and performance ceiling.

The hallucination problem it targets is not HaWMPO-specific. Any team using a video diffusion model or neural world model as a training environment for VLA policies will encounter the same compounding error dynamic in long-horizon rollouts. The Reward-Soft GRPO mechanism is modular enough that it could, in principle, be layered onto other world model backends — which is either a strength (broad applicability) or a risk (the paper's results are tied to their specific implementation, and generalization to other world model architectures is unvalidated).

For hardware companies deploying humanoids in unstructured environments — where long-horizon task sequencing is the unsolved problem — the 15-point LIBERO gain on complex scenarios is the number to watch. Success rates in the 60–75% range are commercially marginal; pushing past 80% on difficult multi-step tasks is where sustained deployment becomes operationally viable.

---

## Key Takeaways

- **HaWMPO delivers +15.0% on LIBERO** over its base VLA model and +2.8% over the strongest evaluated baseline, entirely through world-model-based post-training — no real-robot RL required.
- **Physical validation on the Unitree G1** shows manipulation success climbing from 67.5% to 80.0% across two tasks, a meaningful but narrow hardware confirmation.
- **The core innovation is a hallucination-aware quality filter** on imagined rollouts, not a new world model architecture — making it potentially compatible with other world model backends.
- **The Reward-Soft GRPO mechanism** suppresses unreliable action chunks during training, directly addressing how hallucinated state transitions corrupt policy gradient updates.
- **Skeptics should note** that two real-world tasks and unspecified trial counts leave the hardware results underspecified; broader multi-platform replication is the next required validation step.
- **Commercial relevance:** Synthetic VLA post-training that bypasses hardware-in-the-loop RL could dramatically compress the cost and timeline of policy improvement for teams operating robot fleets.

---

## Frequently Asked Questions

**What is HaWMPO and what problem does it solve?**
HaWMPO (Hallucination-Aware World Model-based Policy Optimization) is a closed-loop reinforcement learning pipeline for post-training Vision-Language-Action models. It addresses the problem of world models generating physically incoherent "hallucinated" state transitions during imagined rollouts, which corrupt policy gradient updates — a failure mode that particularly affects long-horizon manipulation tasks.

**What results did HaWMPO achieve on the LIBERO benchmark?**
According to the paper, HaWMPO achieves the best average success rate on LIBERO, with gains of 15.0% over the base model and 2.8% over the strongest evaluated baseline.

**Was HaWMPO tested on a real humanoid robot?**
Yes. The authors validated HaWMPO on a Unitree G1 robot, raising average success rate on two physical manipulation tasks from 67.5% to 80.0%.

**How is HaWMPO different from standard world model training approaches?**
Standard world model pipelines use imagined rollouts directly for policy optimization without filtering for rollout quality. HaWMPO adds an action-conditioned hallucination-aware model that scores the reliability of generated image sequences and suppresses low-reliability action chunks via its Reward-Soft GRPO mechanism.

**Does HaWMPO require real-robot data collection during training?**
No. The method is designed for VLA post-training using imagined rollouts from world models, explicitly avoiding the cost, sample inefficiency, and safety risks of online reinforcement learning on physical hardware. Real-robot experiments are used for evaluation only.

**Why does hallucination matter more in long-horizon tasks?**
In long-horizon scenarios, prediction errors compound across steps. A world model that produces a slightly incoherent state transition early in a rollout will generate increasingly divergent future states, and policy updates based on those states will learn incorrect behaviors. Short-horizon tasks are more forgiving because the rollout has fewer steps over which errors can accumulate.