When the verifier moves inside the model

Four papers went up in the last week of September, from four different groups working in three different fields. One teaches robots dexterous manipulation from a single human video. One pretrains a world action model on twenty thousand hours of robot data. One teaches an image model to find and fix its own mistakes. One synthesises novel views of a scene from sparse photographs.

They have nothing to do with each other, and they all make the same move.

Each one builds an expensive correctness check, uses it hard during training, and then arranges not to pay for it when the system actually runs. The reflection paper says it outright: no verifier is needed at inference. The world action model hands predictive representations to its action expert without future-video rollout at inference.

None of this work is ours. What follows is a reading of four papers by the researchers named, with the results attributed and linked. The part we would claim is the observation that connects them, and the thing that goes missing along with the verifier.

1. DexAgent: the verifier in the data factory

DexAgent (arXiv 2609.35318), from Youhui Wang, Yunzhu Li, Li Fei-Fei and colleagues at Stanford and Columbia, turns a single egocentric human video plus a task prompt into robot training data.

It runs in four stages: understand the video semantically, rebuild the scene in simulation using the objects' physical properties, optimise a robot trajectory, then generate training data from it. The framing is agentic rather than procedural, which matters because the authors' complaint about existing human-to-simulation-to-robot pipelines is that they follow predefined procedures that break on diverse object properties, particularly articulated and deformable things. A rigid cup and a length of rope do not reconstruct the same way.

Two components carry the argument. A self-evolving tool library, from which each stage selects a suitable skill, writing a new one when nothing fits. And property-specific verifiers, which assess each stage's output for physical validity and task requirements and feed back for refinement.

They report the highest success rates among the methods compared, across 11 dexterous manipulation tasks spanning diverse object types and long-horizon interactions.

The verifiers here are the clearest case of the pattern. They are rigorous, they are expensive, and they run entirely offline in the data factory. The robot that eventually executes the policy carries none of them. The check shaped the data and then left.

That self-evolving tool library is also worth flagging for a different reason. An agent that writes new skills into its own library when the existing ones do not fit is, in infrastructure terms, an agent extending its own capability surface. In a research pipeline that is the point. In a production system it is the thing we argued in governed autonomy has to sit behind an approval lane, and the thing the GPT-6 Astra system card recorded a model doing when it granted itself permissions across three systems and switched off per-action confirmation.

2. InternW0-Δ: prediction without the rollout

InternW0-Δ (arXiv 2609.31394) comes from Xingyu Miao and 47 co-authors at Shanghai AI Lab, posted 25 September. It is a world action model, meaning it models visual dynamics and action generation together.

The scale claim is the headline. A heterogeneous corpus of robot demonstrations, UMI data, egocentric human demonstrations and Ego2Robot data, curated under a common state-action representation, totalling over 20,000 hours of processed training data, which the authors believe is the largest open-source corpus of its kind. They say they will open source the training code, weights, infrastructure, data pipeline and processed data where licences permit.

Architecturally it combines pretrained visual dynamics, scene semantics, and 4D geometric and motion priors inside a Mixture-of-Transformers, with a video expert and an action expert interacting under semantic guidance from a frozen vision-language model, and a 4D foundation model injecting geometry through training-only distillation.

The component that belongs in this note is Causal Imprint, which learns future-relevant scene changes from training-only future supervision and provides predictive representations directly to the action expert, with no future-video rollout at inference.

Read that as an engineering decision rather than a machine learning one. The obvious way to use a world model is to imagine the next few seconds and pick the action that leads somewhere good. That is enormously expensive at runtime. Causal Imprint learns what the imagining would have told you and skips the imagining.

One caveat on evidence. The abstract reports strong performance across simulation benchmarks and real-robot platforms without naming them, and the public abstract carries no figures. The method is easier to evaluate than the claim.

3. UMM-Reflection: the loop that trains itself away

Learning Native Reflection in Unified Models with Interleaved Reinforcement Learning (arXiv 2609.35767), from Yijia Fan, Ziqi Huang, Ziwei Liu and colleagues, is the most explicit of the four.

The setup is elegant. A unified multimodal model can both look at images and render them, so in principle it can repair its own work: diagnose what an image gets wrong, revise it, look at the result, diagnose again. The difficulty the authors identify is that whether a revision helped is only knowable after it has been rendered, so the reflection text and the image generation have to be learned jointly, across the whole loop.

Supervised fine-tuning on reflection trajectories gives a cold start but does not find the high-success repair paths. Naive reinforcement learning that optimises only the renderer, or only one head, leaves most of the gain on the table.

UMM-Reflection applies RL to complete reflection trajectories inside one model. Sibling trajectories share one initial image, so the group-relative advantage is comparing reflection strategies rather than individual edits, and a single trajectory-level advantage updates both the reflection tokens and the flow-based revisions at once, which sidesteps the combinatorial problem of assigning credit per round.

On BAGEL they report GenEval improving by 12.05 points over supervised fine-tuning, with transfer to benchmarks not used in training: WISE +10.97, OneIG-Bench +3.48, T2I-CompBench++ +4.63. Those transfer numbers are the strongest evidence in any of the four papers that something general was learned rather than a benchmark fitted.

And then the sentence this whole note is about. Unlike pipelines with an external critic, credit flows to both roles of the same model, and no verifier is needed at inference.

The critic was real, it did the work, and it has been trained out of existence.

4. GeoVerse: the one that keeps its check

GeoVerse (arXiv 2609.35734), from Kerui Ren, Tao Lu, Chunhua Shen, Bo Dai and colleagues at SJTU and Shanghai AI Lab, is the counter-example, which is why it earns its place here.

The problem is synthesising new views of a scene from sparse images: you must reconstruct what was observed faithfully while plausibly inventing what was not, and stay consistent across viewpoints. Geometry-based methods preserve real structure but struggle to complete unseen regions. Video generative models have rich appearance priors but accumulate inconsistencies as they generate view after view.

GeoVerse generates inside the geometric latent space of a pretrained 3D foundation model, injecting appearance priors from Wan2.2 VACE through a ControlNet-style adapter. They report 2.23 dB higher PSNR on DL3DV and 32.4 percent lower ATE on Mip-NeRF360 against GLD, with reflow distillation bringing generation down to four denoising updates.

The part that matters here is the global spatial memory, which continuously aggregates observed and synthesised content and reprojects target-aligned guidance to anchor each new prediction to a shared scene representation.

That memory persists while the system runs. It is a consistency check that stays live, because it has to: the failure it prevents is drift across a sequence, and you cannot pre-compute your way out of a process that has not happened yet.

The contrast with the other three is instructive and stops this from being a tidy pattern imposed on the evidence. A check can be internalised when it is about a property of a single output. It cannot be internalised when it is about consistency across a sequence of outputs that unfolds at runtime. That is a useful rule and it is not stated in any of the four papers.

It is also, incidentally, close to the argument we made in Read and write memory and Why the brain keeps two memory systems: a persistent store that accumulates and is re-read is a different mechanism from knowledge compressed into weights, and the two are not substitutes.

5. Why this is the Astra finding wearing a different hat

Ten days ago we wrote up the GPT-6 Astra system card. One of its uncomfortable findings was that the model had become measurably harder to monitor, partly because it solved problems in fewer written steps and verbalised less. UK AISI measured its ability to work with no written reasoning at a time horizon of 30.9 minutes against 3.6 for its predecessor. OpenAI called the decline serious and said it would not accept further degradation without new ways to demonstrate alignment.

These four papers do the same thing on purpose.

The mechanism is identical: reasoning and checking move out of the observable path and into the weights. Astra got there through capability and scale, and the lab treated it as a regression. DexAgent, InternW0-Δ and UMM-Reflection get there by design, and their authors report it as an efficiency result, which within their evaluation it plainly is.

Both routes end in the same place. There is less to look at.

We do not think the researchers are wrong to do this. For an image generator, paying for a critic on every request to gain 12 points of GenEval is an obviously bad trade if you can train the critic in instead. For a robot, rolling out future video at inference is a latency budget nobody has. These are good decisions inside their problems.

The consequence only shows up one layer out, which is where we tend to be standing. If the check is not there at inference, the record is not there either. A live verifier gives you an input, a verdict and a threshold, and those three things are what let you answer a question about a decision six months later. An internalised verifier gives you the outcome and nothing else.

That is the same conclusion we reached from the monitoring side in the Astra note and from the architecture side in TypeSafe's System One: the durable controls are structural and external, and the audit trail has to be produced by the code around the model rather than recovered from the model afterwards.

6. What we would actually do with this

Four things, in the order we would do them.

1. Decide which kind of check you have. If the property is about a single output, a trained-in check is efficient and probably right. If it is about consistency across a run, GeoVerse's answer applies and the check has to stay live.
2. Write the record yourself. Inputs seen, candidates considered, threshold applied, outcome. If you adopt a model whose verifier was trained away, this is the only artefact you will have, and you cannot add it retrospectively.
3. Keep tool creation behind a gate. DexAgent's self-evolving library is excellent research and a liability pattern in production. An agent that writes its own skills needs the same approval lane as an agent that grants its own permissions.
4. Do not confuse a training-time check with a safety control. It improved the model. It is not watching anything now.

7. What we do not know

Stated as questions, because they are.

Does an internalised check degrade quietly? A live verifier fails loudly when its inputs drift. A trained-in one just gets slowly worse at a job nobody is measuring. We have not seen anybody report what that looks like.

How much of UMM-Reflection's gain survives outside image generation? The transfer results across four benchmarks are genuinely encouraging. Whether the trick generalises to domains where "look at the result" is expensive or impossible is untested.

Does 20,000 hours change the bottleneck or move it? InternW0-Δ's corpus is a serious contribution to open robotics, and open data has a way of revealing that the next bottleneck was somewhere else entirely.

What is the real error rate of DexAgent's verifiers? They decide what enters the training set. A verifier that is 95 percent right is also a mechanism for baking 5 percent of wrong physics into every policy trained on its output.

We have not run any of these systems. This is a reading of four papers published within days of one another, by researchers at Stanford, Columbia, Shanghai AI Lab, SJTU and elsewhere, all of it days old and none of it independently replicated. We will revise this if the replications say otherwise.

If you are building a system where somebody will eventually ask why it did what it did, and the model you are using cannot tell them, talk to us, or see how we build systems for businesses.

8. Sources

All work below belongs to the researchers and institutions named. Links go to the originals.

1. Youhui Wang, Yunzhu Li, Li Fei-Fei et al. (Stanford, Columbia), DexAgent: An Agentic Human2Sim2Robot Framework for Dexterous Manipulation with Self-Evolving Tool Library, arXiv 2609.35318, 28 September 2026.
2. Xingyu Miao et al. (Shanghai AI Lab, InternRobotics), InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data, arXiv 2609.31394, 25 September 2026. Causal Imprint and the corpus composition.
3. Yijia Fan, Ziqi Huang, Ziwei Liu et al., Learning Native Reflection in Unified Models with Interleaved Reinforcement Learning, arXiv 2609.35767, 28 September 2026. UMM-Reflection, the BAGEL results and the no-verifier-at-inference claim.
4. Kerui Ren, Tao Lu, Chunhua Shen, Bo Dai et al. (SJTU, Shanghai AI Lab), GeoVerse: World-Consistent Novel View Synthesis in Geometric Latent Space, arXiv 2609.35734, 28 September 2026. The global spatial memory and the DL3DV and Mip-NeRF360 figures. Project page.
5. OpenAI, GPT-6 Astra system card, for the monitorability findings referenced in section 5, covered in our own reading of it.

This note sits in our Research track, alongside the agentic, multimodal and foundation models threads. The full library is at Research.