Every week 120+ ML preprints hit arXiv. Most are incremental. Most will never be cited. A handful contain work that will quietly end up inside every production system you use two years from now.

This is the good batch from June 19 2026.

PP-OCRv6: Specialized models still beat general VLMs

PP-OCRv6 dropped this week on Hugging Face. No fanfare. No demo thread with 100k impressions. It is the best general purpose OCR you can run today.

The largest variant is 34.5 million parameters. That is 0.02% the size of GPT-4o. It beats GPT-4o on structured document OCR accuracy by 7 percentage points on Paddle's internal benchmark. It runs on a Raspberry Pi.

There are three correctly calibrated tiers for actual deployment: 1.5M for edge devices, 7.7M for mobile, 34.5M for server. Detection Hmean improved 4.6 points over v5, recognition up 5.1 points. 50 languages run out of the same model weights.

You will not read this on any AI news site. Everyone is talking about VLMs. Everyone is building RAG pipelines that feed blurry screenshots to 1T parameter models. Every single one of those pipelines would get better results, 100x faster, 1000x cheaper, by putting this 34M parameter model first.

This is not an opinion. This is measured. The entire industry has forgotten that specialized narrow models still win, by very large margins, on every single well defined task that existed before 2023.

BIM-Edit: LLMs cannot edit structured engineering models

This benchmark is brutal. And it is important.

Everyone demos LLMs generating CAD models from text. No one demos them editing an existing model correctly. No one tests if they preserve the thousands of implicit semantic and topological relationships that make engineering models useful.

BIM-Edit tests exactly that. 324 editing tasks on real building IFC models. Tasks include things like "move all fire extinguishers 1.2m above floor level", "delete every wall that is adjacent to an exterior window", "increase the rating of all circuit breakers on level 3 by 10 amps".

No model passed more than 3.4% of tasks. The best performing closed model available today scored 49.5% average across geometric, semantic and topological metrics.

This is the gap no one talks about. LLMs are very good at generating disposable demo artifacts. They are terrible at modifying existing, structured, correct systems. Until this changes, all the talk about AI doing engineering is marketing.

You do not hire an engineer that can draw a pretty house once. You hire an engineer that can modify the same house 47 times over three years without breaking the plumbing.

Effective dimension explains quantum ML's weird results

Quantum ML papers have been reporting the same two weird results for 18 months now. More entanglement gives better generalization. Adding noise makes test accuracy go up.

Everyone treated this as magic. No one had an explanation. People just grid searched ansatzes and noise levels and published whatever worked.

This paper resolves both observations completely. Both effects are just adjustments to the kernel effective dimension. Entanglement spreads the kernel spectrum. Noise contracts it. You are just tuning implicit regularization.

When you are in overfitting regime, contracting effective dimension improves test accuracy. When you are in underfitting regime, the exact same noise will make accuracy worse. There is a clean inverted U curve. The authors measured a 13% accuracy lift from correctly calibrated amplitude damping noise.

This is not a small result. This moves quantum ML from empirical alchemy to something you can actually design. There is no magic anymore. Just kernel machines running on weird hardware.

Thermodynamic measure of intelligence

This is the most important paper in this batch. It will probably be ignored.

The authors define intelligence as lawful amplification of rare valid futures. An intelligent system increases the probability of outcomes that would almost never happen by chance, but are allowed by the hard constraints of the world.

They prove two formal results. First: you cannot get high amplification unless the system runs a recursive self simulation that includes itself in the modeled future. Second: if you do have that self simulation and a working policy, you get very close to the theoretical maximum possible amplification.

This is a universal scale. You can measure the intelligence of a thermostat, an LLM, a human, a Maxwell demon on the exact same axis. There are no subjective scores. No benchmarks. No trick questions. Just a number you calculate from observed behaviour.

We have spent 70 years arguing about what intelligence is. This paper gives a working, testable definition. We will see if anyone uses it.

Time series modeling needs to abandon transformers

This position paper dropped at ICML. It has already started fights on every technical forum.

The core argument is very simple. Every real world time series comes from a dynamical system. Dynamical systems are defined by recursion. Transformers do not do recursion. They approximate local windows. They will never capture the global structure of the system.

The authors make five concrete recommendations. Stop benchmarking only on 12 step forecasting. Train for dynamical system reconstruction, not next token prediction. Pretrain on simulated chaotic systems, not synthetic sine waves. Stop using transformers. Go back to modern RNNs.

They have the numbers. Every transformer time series model fails catastrophically once you go past 50 steps. They get the next step right. They get the long term statistical behaviour completely wrong. They will never predict tipping points. They will never generalize out of domain.

This is a direct attack on every time series foundation model released in the last two years. And they are right.

RL for game AI is stuck at deployment

Everyone knows RL can beat humans at any game. No one is using it in shipped commercial games.

This paper explains why. All published RL game AI work optimizes for win rate. Commercial game AI does not exist to win. It exists to lose in interesting, believable, consistent ways. It exists to make the player have fun.

There is no research framework for this. There are no benchmarks. There are no standard training objectives. Every studio that has shipped RL agents built everything from scratch, and will not talk about how they did it.

The paper proposes a minimal set of requirements for deployable game RL. It also lists the hard unsolved problems: consistent personality, bounded stupidity, graceful failure, memory that does not drift over 10 hours of play.

This is one of the largest open gaps in applied ML right now. 3 billion people play video games. Almost none of them have ever interacted with an ML powered NPC.

DataMagic: end to end data video generation

Everyone hates dashboards. No one reads static charts. People watch videos.

Making good data videos takes 10 hours minimum. You need to do analysis, design narrative, animate every transition, record voiceover.

DataMagic takes raw tabular data and a natural language query and outputs a complete narrated animated data video. It does not generate pixels. It generates a declarative specification bound directly to the source data. Every element in the video traces back to an exact cell in the input table. There is zero hallucination.

The architecture uses parallel agents to generate candidate scenes, then a global orchestrator to arrange them into a coherent narrative. You can also click any point in the finished video to ask questions about the underlying data.

This is not another AI presentation generator. This system changes the entire workflow of how people communicate data. It will be inside every BI tool by 2027.

What unites all this work

None of these papers are about scaling. None of them are about making models bigger.

All of them are about fixing the things that are broken right now. All of them are about moving past demos to things that actually work in the real world.

We have just finished a three year period where almost all research effort went into scaling foundation models. That era is ending. The good work now is going into understanding what these models can actually do, what they cannot do, and building specialized systems that do the boring important jobs correctly.

No one will make a billion dollar startup from any one of these papers. But every system you build over the next five years will use ideas from at least one of them.

Full source references

1. DataMagic: Transforming Tabular Data into Data Insight Video - http://arxiv.org/abs/2606.20388v1
2. Thermodynamic Measure of Intelligence - http://arxiv.org/abs/2606.20231v1
3. Augmenting Game AI with Deep Reinforcement Learning - http://arxiv.org/abs/2606.20210v1
4. Effective Dimension Governs Generalization in Quantum Kernel Vision Models - http://arxiv.org/abs/2606.20183v1
5. BIM-Edit: Benchmarking Large Language Models for IFC-Based Building Information Modeling - http://arxiv.org/abs/2606.20146v1
6. Time Series Modeling Needs a Dynamical Systems Perspective - https://arxiv.org/abs/2602.16864
7. PP-OCRv6 on Hugging Face - https://huggingface.co/blog/PaddlePaddle/pp-ocrv6