If we look at these papers together today, I think the most obvious line is not "the model has grown again", but:
Everyone is more and more serious about the "middle layer problem" in the AI system.
What is the middle layer problem?
It is not the lowest-level Backbone, nor the most superficial BenchMark score, but the connection layer that determines whether the system can work stably:
- Extend the strategy during the test, whether it should be manually adjusted by the person, or let the agent discover it for itself
- Is it possible to unify a set of reusable infrastructure for so many heterogeneous data for autonomous driving?
- Vision-Language Model If you want to truly understand the 3D world, the visual token means how to compress, align and inject
- After training after image generation, how to avoid multi-target rewards and fight each other as soon as they come up
- If you want to be credible, how to really connect "with explanation path" and "with statistical calibration"
They seem to be in different directions:
- LLM Reasoning
- Autonomous Driving Data Engineering
- 3D Vision-Language Model
- Vincent post training
- Trusted Q&A
But put together, you will see a trend that is getting clearer and clearer:
AI research is shifting from "making ability" to "connecting ability, saving, and credible".
Today, I will pick 5 articles to represent the five lines of strategy search, data infrastructure, 3D representation, multi-target image alignment and trusted path reasoning during the test.
If you are more concerned about "what directions are more likely to directly affect the real system", I think this group of papers is worth reading today.
1) AutoTTS: Extension during testing, you don't have to rely on people to design strategies
- arXiv:2605.08083
- Title: LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling
- Orientation: Test-Time Scaling / Agentic Discovery / Reasoning Control
Test-time scaling has become one of the most common keywords in the inference model in the past year.
One way that everyone accepts by default is:
- Sampling several inference paths
- Appropriate branch, look back, prune
- Add some Probe, score, vote, select strategy
The problem is that these strategies are mostly labor design yes.
That is to say, although we are talking about "let the model think for a while", but "how to think, when to stop, when to fork, and when to retreat", it is often the researchers who have come together from experience.
It is here that AutoTTS wants to change this paper.
Its core proposition is:
Instead of designing a set of test-time scaling rules manually, it is better to turn the rule search itself into a task that the agent can do automatically.
There are two of the most interesting things about this paper
First, it is not a direct violent search, but a "searchable environment" first
The paper emphasizes the Environment Construction.
This is the key point.
Because the action space of the expansion strategy is not sorted out first, the so-called "automatic discovery" can easily become directly:
- Too much search space
- LLM call cost is too high
- It is too expensive to verify a strategy at a time
- The last method itself is more expensive than the gain
So the author made a very engineering compromise:
- Based on pre-collected Reasoning Trajectories and Probe Signals
- Let the controller decide when to branch, continue, probe, prune, stop
- Inexpensively evaluate the strategy without repeating a large number of the bottom modulus
This turns the problem from "online expensive search" to "controlled policy discovery in offline environments".
Second, it represents a direction worthy of long-term attention: turning researcher handwritten Heuristics into a learnable controller
I think the importance of this matter is not just test-time scaling.
It's more like saying:
Many reasoning processes that still rely on manual adjustment today may be re-expressed as "controller learning problems".
Include:
- when to retrieve
- when multisampling
- When to stop
- when to call critic
- When to roll back to a more conservative strategy
The results given in the paper are also very eye-catching:
- Mathematical Reasoning BenchMark Can Improve Accuracy-Cost TradeOff
- It is found that the cost is not high, and the Discovery process even emphasized that it only cost tens of dollars
It is not "another test-time scaling skill" that this paper is worth taking away, but it is seriously advancing a bigger direction:
The reasoning strategy itself can also be agentized.
2) 123d: Autonomous driving researches a long-term pain point, and finally someone starts with the "unification layer of data"
- arXiv:2605.08084
- Title: 123D: Unifying Multi-Modal Autonomous Driving Data at Scale
- Orientation: Autonomous driving / multimodal data infrastructure / dataset unification
Autonomous driving for so many years, the data has actually been very rich.
But an embarrassing fact is:
"Abundance" does not mean "can be used together".
These problems often occur between different datasets:
- The modality is different: Camera, Lidar, Ego State, Traffic Light, and HD Map are organized in different ways
- The sampling rate is different, the timing alignment is very troublesome
- Inconsistent labeling system
- File format, toolchain, dependency stack split
- You can reproduce experiments on one dataset, it does not mean that you can move to another
123D does not seem to be so flashy what this paper does, but I think it's particularly important:
It attempts to turn the self-driving multimodal data into a set of "event flow infrastructure" under the unified API.
Why is this idea correct?
Because many people talk about the self-driving Foundation model, they are watching by default:
- Model structure
- policy learning
- world model
- planning
But in reality, a lot of time is actually spent on the lower level:
- How to read data
- How to align the modality
- How to unify the semantics of different datasets
- Training and evaluating whether it can be migrated across sets
If this layer is not solved, many "unified models" on the upper floor will be dragged by the infrastructure.
A key design of 123D: each modality is seen as a standalone timestamped event stream
This design is very practical.
Because the driving sensor in reality is not synchronized in a neat manner.
If the system is forced to:
- All modals at the same frequency
- Hard alignment of all data
- All data is organized in a rigid format
Then it will be very painful to integrate across datasets.
123d in reverse do:
- Each modality is its own event flow
- Synchronous or asynchronous access can be supported
- Data access logic across datasets is more unified
In terms of system design, this is actually more long-term value than "making a new Benchmark".
Why do I think this is worth paying attention to?
Because it represents a mature signal:
Autonomous driving research has begun to regard "unification layer of data" as the first problem, rather than default preconditions.
As long as the data layer is truly unified, many things will be smoother after:
- cross-dataset transfer
- Planning and reinforcement learning reuse
- world model pre-training
- Data quality and labeling deviation analysis
This kind of work may not be as eye-catching as the strongest model headline in the short term, but in the long run, it can often determine whether a direction can really expand.
3) Proxy3D: If VLM wants to really understand the 3D world, the key may not be more pixels, but a better proxy representation
- arXiv:2605.08064
- Title: Proxy3D: Efficient 3D Representations for Vision-Language Models via Semantic Clustering and Alignment
- Orientation: VLM / 3D Spatial Intelligence / Representation Alignment
This paper is a direction worthy of continued attention:
How the Vision-Language Model can truly gain a stable 3D space comprehension.
Many VLMs are still typical 2D pipelines:
- images come in
- become a visual token
- re-access language model
This method has no problem with the general picture question and answer, but once the task becomes:
- 3D grounding
- Spatial Consistency across Perspectives
- Spatial Relational Reasoning in Video
- Object layout understanding in real environment
2D Token means that it is easy to start to show shortcomings.
The core idea of Proxy3D is very clear:
Instead of keeping all the visual information at the pixel alignment level, first extract the more compact and geometric semantic 3D Proxies.
The essay is:
- Extract semantic and geometric features from video frames
- Do semantic perception clustering in 3D space
- Get the proxy representation of a set of Compact
- Then align this set of expressions with VLM through multi-stage training
The value of this paper is not just "more token"
More importantly, it is answering a system question:
What kind of representation VLM should use to enter the language side can not only keep the spatial consistency, but also do not let the visual sequence length explode.
This is a particularly realistic contradiction in the Multimodal system:
- Keep more pixel details, token will explode soon
- Too much pressure, 3D relationship is unstable
Proxy3D is a proxy-based that is trying to find an intermediate point.
Why do I think this line is very important?
Because in the future, many multimodal tasks are not as simple as "looking at a picture and say a sentence", but more biased:
- spatial intelligence
- embodied reasoning
- video-grounded QA
- GUI / environment interaction
In these tasks, the 2D token directly pours the LLM route, which is likely to be less and less enough.
So I will see this paper as a clear signal:
The key to the next competition of the Multimodal system is not only the size of the bottom mode, but who first presses the visual world into a structured representation that is more suitable for reasoning.
4) Flow-OPD: training after image generation, more and more like "multi-teacher knowledge integration" rather than a single reward reinforcement
- arXiv:2605.08063
- Title: Flow-OPD: On-Policy Distillation for Flow Matching Models
- Direction: text-to-image/post-training/reward alignment/distillation
Training after image generation has also become more and more like the LLM community recently.
Everyone started talking:
- RL alignment
- reward hacking
- multi-objective balance
- on-policy sampling
- Cold start and distillation
I think the entry point of this paper is very accurate.
It points out that there are two common pits for the Flow Matching Vincent graph model when doing multitasking:
- scalar reward too sparse
- When multiple heterogeneous targets are trained together, the gradients interfere with each other
The end of the typical seesaw effect will appear:
- increase an indicator
- Another indicator fell
- The model begins to learn to speculate and score points, instead of really getting better overall
The core practice of Flow-OPD: first train different goals into their own teachers, and then slowly distillate back to a student
This idea is very similar to "divided into subjects first, and then the unified examination".
The two-stage design in the paper is roughly:
- First use a single reward to push different teachers to their respective ceilings
- Then, through the on-policy distillation, put multiple specialties back into a student
Also added:
- Flow-based cold start
- task routing labeling
- trajectory-level dense supervision
- Manifold Anchor Regularization as Universal Quality Anchor
Why is this important?
Because it reflects an increasingly clear reality:
The difficulty of training after the generative model is not only "what is the reward", but "how to integrate multiple rewards without pulling each other".
This is more and more similar to LLM post-training.
The model is not only chasing one indicator, but at the same time:
- command follow
- Aesthetic quality
- OCR Accuracy
- structural consistency
- Human preference
If it's all blurred into a reward to be hard-trained, it's usually messed up quickly.
My judgment on this paper
It is worth noting that it is not necessarily how many points have been improved, but it represents a methodology that will become more and more common:
The post-training of the generic generative model is more like a multi-expert distillation and strategy integration in the future, rather than just doing RL directly on a single model.
If this approach matures, the idea of specialized teachers passing what they learn on to a single student could extend beyond text-to-image generation to video, multimodal generation, and even agent policy alignment.
5) CPR: If KGQA wants to be truly credible, it is not enough to explain the path, and the coverage calibration must be added.
- arXiv:2605.08077
- Title: Conformal Path Reasoning: Trustworthy Knowledge Graph Question Answering via Path-Level Calibration
- Direction: KGQA / Conformal Prediction / Trustworthy reasoning
There has always been a very attractive point in the knowledge graph question and answer:
It is naturally more "evidence to check" reasoning than plain text generation.
Because the model often finds the answer along the relationship path in the figure.
But the problem is, "with path" does not mean "trusted".
In reality, there will still be:
- The coverage is unstable
- Prediction set is too large
- calibration failure
- The model seems to give the inference path, but it is not reliable in the statistical sense
It is this layer that is to be supplemented by this paper.
The core idea of CPR:
The approach applies conformal prediction to the reasoning paths themselves, going beyond a loose check on the final answer.
It does two key things:
First, conformal calibration is performed at the query level, using scores assigned to individual reasoning paths
This is very important.
In knowledge graph question answering, the basic unit of reasoning is a candidate path leading to an answer, not just the final answer token.
If you only do Conformal in the final answer layer, a lot of intermediate information has actually been lost.
CPR directly incorporates the Path Score into the calibration object, which is equivalent to the reasoning unit of KGQA itself.
Second, it adds a Residual Conformal Value Network to learn nonconformity scores that better distinguish between paths
This is also a critical engineering patch.
Because the Conformal method is often beautiful in theory, but the common problems after really entering complex tasks are:
- Score too thick
- set becomes big
- Although the coverage rate has gone up, the availability has decreased
Therefore, the author specially designed light modules to improve the separability of the path-level score.
What is the most interesting thing about this paper?
I don't think it's the specific task of KGQA, but the more common trend behind it:
The AI system is trying to really connect "structured inference" and "statistical calibration guarantee".
This will become more and more important in the future.
Because whenever the system enters a high-risk or high-value scenario, everyone will ask:
- Not just what the answer is
- Also know what the credible boundary of this answer is
- How likely is the correct answer
- Is the Prediction Set size still available?
This combination of "structured explanation + statistical guarantee" is likely to have a longer-term value than a simple and stronger generation ability.
Final summary
If I read these 5 papers together today, I think they are pointing to one thing together:
AI research is earnestly supplementing the "system intermediate layer".
Specifically:
- Autott's Strategic Discovery Layer in Supplemental Reasoning Stage
- 123d in the unified layer of data supplementing autonomous driving
- Proxy3D in the 3D presentation layer of VLM
- Flow-OPD trains the layer after multi-objective generation
- CPR in the trusted calibration layer of complementary structured Q&A
They're not the kind of work that "make the model bigger".
But it is precisely this kind of work that often determines whether the system is finally better or not:
- Can really expand
- can really migrate
- Can really align
- Can really be trusted
If I summed up today's papers in one sentence, I would write:
In the next stage of AI, the gap between AI is more and more is not just the ability gap of the main model, but who first makes the intermediate layers of connection ability, representation ability, training ability and credibility ability solid.
If you've been chasing your papers recently, it's not just to look at the headline, but to find those directions that are more likely to affect the design of the real system, I think this group is worth keeping.