If I read this batch of papers together yesterday, I think there is a very clear collinearity:

AI research is becoming less and less than the "capability cap", but rather the hard constraints that the system will encounter after it actually enters the real environment more seriously.

What are these constraints?

  • The training budget is not unlimited, and the Scaling Law itself has to talk about the cost
  • Agent not only generates text, but also has to be able to predict the environment, understand feedback, and correct the world model
  • Multi-picture visual understanding is not to feed a few pictures together, even if it is completed, what is really difficult is the consistency of cross-picture and object constancy.
  • Real-world 3D reconstruction not every scene has a lot of high-quality pictures like popular landmarks, more sparse, noise, long tail
  • After the model-generated content enters the real distribution link, the issue of source tracking and copyright ownership has begun to become more and more difficult to get through

So the 5 articles I picked yesterday were not simply picking the "headline", but as much as possible to pick those jobs that can reflect how the system bottleneck is changing.

from:

  • Agent / World Model Review
  • scaling law and experimental budget
  • Multi-picture MLLM training
  • 3D Reconstruction
  • LLM watermarking

They look scattered, but they just see a trend together:

In the next stage of AI competition, it is not only the stronger model, but who knows more about the realization of budget, environment, evidence, consistency and traceability into the system design together.


1) Agentic World Modeling: What the Agent really lacks may not be more tools, but a more decent World model

  • arXiv:2604.22748
  • Title: Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond
  • Direction: Agent/World Model/Environment Modeling/Survey

The first thing worth noting in this paper is that it does not continue to say the world model into a vague word, but tries to reorganize the understanding of the "world model" by different communities into a unified framework.

The author's judgment is very clear:

When the AI system moves from "generating a text" to "continuously accomplishing goals", environmental dynamics modeling becomes the core bottleneck.

This is true in many scenarios:

  • Operating the GUI's Agent
  • Agents for browsing the web and calling tools
  • multi-agent collaboration system
  • Scientific Discovery and Experimental Design System
  • Task Assistant that requires long-term interaction and feedback correction

The paper proposes a very useful "Levels × Laws" Taxonomy.

It divides the ability into three layers:

  • L1 Predictor: learn one step local transfer
  • L2 Simulator: do multistep, action conditions Rollout
  • L3 Evolver: After the prediction fails, you can also revise your own model according to the new evidence

At the same time, the constraints are divided into four categories:

  • physical
  • digital
  • social
  • scientific

The value of this framework is not just a beautiful classification, but it is reminding a real problem:

The reasons for the failure of different agent scenarios are different, because the "regular types" of their environment are different.

For example:

  • Robots must obey the laws of physics
  • GUI Agent to obey the software interface status transfer
  • Social simulation to face social norms and multi-subject behavior
  • Scientific research agents must comply with scientific constraints and experimental feedback

The paper also systematically organizes more than 400 related work and more than 100 representative systems, which makes it more like a roadmap, not just ordinary Survey.

I think the most important inspiration for this article is:

Whether the Agent can do well is not just "will call the tool", but whether it has a good enough environment prediction structure, can form effective expectations before action, and correct expectations after action.

If you are looking at Browser Agent, GUI Agent, and Multi-Agent Simulation recently, this article is worth adding to the reading list.


2) Spend Less, Fit Better: scaling law is not only by "burning a lot of money before fitting"

  • arXiv:2604.22753
  • Title: Spend Less, Fit Better: Budget-Efficient Scaling Law Fitting via Active Experiment Selection
  • Direction: Scaling Law / Experimental Design / Budget-Aware Training

This paper captures an often overlooked but actually very expensive problem:

Everyone likes to use Scaling Law to plan large-model training, but in order to fit Scaling Law, you have to do a bunch of expensive experiments first.

In other words, Scaling Law seems to be helping you save money, but its own "tuition fees" may already be very high.

The author reiterates this question as:

  • There are a bunch of candidate experiments
  • The cost of each experiment is different
  • Limited budget
  • The goal is not to run all points, but to allow you to extrapolate the most accurate in the high-cost target range

This reconstruction is very important.

Because it changed the Scaling Law fitting from a "data analysis problem" to a more realistic one:

Sequential experimental design problems under budget constraints.

The method they propose is the active selection strategy of Uncertainty-aware.

The results are also interesting:

  • On a diverse set of Scaling-Law tasks
  • It usually only uses about 10% total training budget
  • Approximately the effect of fitting with full experiments

I think the biggest value of this paper is not just "another better fitting algorithm", but it is a reminder of an increasingly critical fact in large model engineering:

As the training cost is getting higher and higher, even "small experiments to make judgments" must start to talk about ROI.

In the future, many training decisions may be viewed like this:

  • Don't do it first and then analyze it
  • Instead, decide where to spend the next budget

This kind of thinking is actually very valuable for formula search, distillation design, and post-training experiments.


3) CGC: It is really difficult to understand more than one picture. It is not to see multiple pictures, but not to read the object relationship.

  • arXiv:2604.22498
  • Title: Compositional Grounded Contrast for Fine-Grained Multi-Image Understanding
  • Orientation: MLLM / Multi-Image Understanding / Grounding / RL-based Reward

This paper is a problem that I have always felt very critical, but not fully solved:

MLLM seems to be quite strong on a single picture, but it is still easy to make mistakes when it is a multi-picture and fine-grained understanding.

Typical questions include:

  • spatial hallucination
  • attention leakage
  • object constancy fails

That is:

  • The attributes of Figure A run to Figure B
  • The model's attention is biased by a similar background
  • After the same object changes the perspective, the model will not recognize whether it is the same

The CGC (Compositional Grounded Contrast) proposed in the paper has two points worthy of attention.

First, it tries to avoid expensive artificial labeling and large-scale CoT synthesis

The authors build multi-image training examples by combining existing grounding annotations for single images. This spares them the work of annotating a new multi-image dataset at scale.

This is very pragmatic.

Because if multi-picture data is completely manually marked, the cost will be very high, and MLLM training is already very burning resources.

Second, it explicitly does "Cross-Picture Contrast" and "Intra-Graphic Contrast" explicitly

CGC uses two types of constructs:

  • Inter-Image Contrast: Strengthen cross-map identification to avoid semantic confusion in different diagrams
  • Intra-Image Contrast: Enhances object constancy across perspectives or related samples

In addition, it also adds a Rule-Based Spatial Reward, to strengthen:

  • source-image attribution
  • spatial alignment
  • structured output validity

This is very important.

Because many of the problems with MLLM are not "not at all", but:

It does not have strong structural constraints when it is output, so it is easy to organize the things that are seen wrong.

In the experiment, CGC got good results on the multi-picture Benchmark, and also has a migration gain for the wider multimodal tasks.

I think the most worthy judgment of this paper is:

Multi-graph understanding is not simply lengthening the context length, but must explicitly address cross-graph attribution, spatial consistency, and object constancy.

If these points are not individually modeled, the model is easy to "every picture is understood, but it starts to mess together".


4) MegaDepth-X: 3D Reconstruct the next difficulty, not a popular landmark, but a long tail part of the Internet image distribution

  • arXiv:2604.22714
  • Title: Long-tail Internet photo reconstruction
  • Direction: 3D Vision/Reconstruction/Sparse-View/Foundation Model Finetuning

This paper discusses a particularly realistic 3D issue.

In the past, we have seen many 3D Reconstruction systems that work very well, and the default data conditions are often:

  • a famous landmark
  • There are a lot of photos
  • Rich shooting angle
  • Relatively stable image quality

But most of the scenes in the real world do not meet these conditions.

Really more common situations are actually:

  • image sparse
  • quality varies
  • The shooting angle is very uneven
  • Repeated textures, symmetrical structures
  • No reliable geometric oversight available

The author believes that this "Long-Tail Regim" is the area that the 3D Foundation Model must be dealt with seriously next.

Their core idea is:

Clean supervision is hard to come by for sparsely observed scenes in the long tail. The authors instead start with popular scenes that already have good reconstructions and deliberately sample subsets that mimic the sparsity of long-tail scenes.

Based on this idea, the paper does two things:

  • Put forward MegaDepth-X dataset
  • Design a sampling strategy to bring the training image set closer to the camera distribution in the long tail scene

Then use this data to go to fine-tuning 3D Foundation models.

The result is:

  • Reconstructing more stable under extreme sparse conditions
  • Also, more reliable in symmetrical and repetitive structure scenarios
  • It also maintains the generalization of the standard Dense Benchmark

What really appealed to me in this paper is that it embodies a trend very typical:

Many "models are not strong enough" problem, and finally found that it is not that the backbone is not large enough, but that the training distribution does not cover the most difficult and common tail situation in the real world.

This is the same for 3D directions, and it is actually similar to autonomous driving, robots, document understanding, and long tail OCR.

In other words:

It's not the most eye-catching part on the Benchmark that is the most worrying, but the long-tail conditions that were ignored by default in the past, so it is most likely to decide whether the system can land.


5) SSG: Why is the LLM watermark easy to fail in low entropy generation? The problem may be in the way of dividing the vocabulary

  • arXiv:2604.22438
  • Title: Logit-Balanced Vocabulary Partitioning for LLM Watermarking
  • Direction: LLM Watermarking/Code Generation/Math Reasoning/Security

The LLM Watermarking line has recently become more and more worthy of attention.

Because as long as the model generation content starts to really enter:

  • content platform
  • Educational Scene
  • code generation link
  • Enterprise Writing and Auditing System

Issues such as "whether this paragraph is generated by the model, which model system comes from" will become more and more realistic.

This paper discusses the problem that the classic KGW watermarking scheme fails in low entropy scenarios.

The most typical of the low entropy scenario is:

  • code generation
  • Mathematical Reasoning

The token distribution of such tasks is usually sharper, and the model has less space for optional space at each step.

That leaves the watermark less room to influence token selection, making it harder to detect.

The authors refine this phenomenon into a very critical concept:

Watermark strength is strongly correlated with the next-token probability distribution.

This means that the question is not just "no watermarking," but:

  • Under the current token distribution, is there enough space for the watermark to operate?

The SSG (Sort-then-Split by Groups) practice they proposed is essentially redesigning the vocabulary to make the two candidate subsets more balanced on logit.

The goal of this is:

Raise the lower bound of the Watermark Strength every time the token prediction is raised.

To put it bluntly, it is to ensure that:

  • even in low entropy generation
  • It can also leave a little controllable space to give a watermark signal

Experiments show that it can improve watermark detectability in code generation and mathematical reasoning tasks.

One thing that makes me think this work is very interesting is:

It's not doing a particularly grand "AI security narrative", but it's a very specific focus on an engineering fact:

Whether you can put a stable watermark depends on whether the sampling space is too narrow for the task itself.

This kind of work may become more and more important in the future, because if watermarking technology is only valid in open chat, but when it comes to code and reasoning scenarios, the actual availability will be very limited.


Final summary

If I read these 5 papers together yesterday, I think the most noteworthy is not the small direction they are in, but a systematic tendency to which they all point to:

  • Agent is no longer just a "will the tool be adjusted", but whether there is an environment model
  • Scaling Law is no longer just "fitting a beautiful curve", but how the budget is more valuable
  • MLLM is no longer just "can you see multiple pictures", but whether the cross-picture attribution is stable
  • 3D reconstruction is no longer just "how high the head of the benchmark", but whether the long tail scene can hold it
  • LLM watermarking is no longer just "theoretical feasible", but whether it is still valid in low entropy tasks

These questions have one thing in common:

They are all pushing AI research from "capability display" to "real availability under system conditions".

This is also an observation that I have become more and more concerned about recently:

In the future, it may not only be the model headline that is bigger, but who is more able to make the cost, environment, consistency, interpretability and traceability more solid together.

If you are more concerned about the implementation of the project recently, rather than just looking at the list, I think yesterday's batch of papers is worth reading.