Nobody is inventing new useful architectures right now. The important work this year is not another transformer variant. It is proving that the architectures we already have work perfectly well for almost every domain, if you stop trying to rebuild them from scratch for every task.

This week three unrelated papers landed on arXiv that make this point unambiguously. Each took a standard, widely deployed ML paradigm, dropped it into a highly specialized domain with minimal modification, and got results that beat every existing domain specific model. None of them added custom layers. None invented new attention mechanisms. All of them only adjusted training objective and data formatting.

The quiet shift in applied ML research

For the last seven years, every applied ML paper opened with "we propose a novel architecture for X". That era is over.

This new batch of work does the opposite. It starts with "we did not change the architecture. We tested if the existing one works here."

This is not incremental work. This is the phase where a technology stops being a research subject and becomes a tool. When you stop arguing about how to build the hammer and start testing what you can hit with it.

JEPA works on things that are not images

First paper: JA4-JEPA. Most people last heard about JEPA when Yann LeCun's group published V-JEPA for video back in January. Almost everyone wrote it off as another computer vision trick.

Nobody tested if the JEPA objective works on discrete, structured, non perceptual data. Until now.

JEPA does not reconstruct inputs. It trains a predictor encoder to guess the latent output of a frozen target encoder for masked portions of the input. There is no pixel loss. No token reconstruction loss. Only latent matching.

Everyone assumed this only worked because images have nice local spatial consistency. This paper proves that assumption was wrong.

JA4-JEPA implementation details

The authors took the exact I-JEPA transformer architecture, unchanged. They did not modify the attention pattern. They did not adjust the loss function weightings.

They only changed one thing: how input views are constructed. Instead of cropping image patches, they split JA4 network fingerprints into four independent subfields: JA4, JA4H, JA4S, JA4X. These are the standard four components of modern TLS network fingerprints.

Training data was 397,000 real fingerprints pulled from JA4DB and CIC-IDS-2017. Critically, 62% of samples were missing at least one subfield. No sample had all four views available.

This is exactly the messy incomplete data you deal with in production security systems. Not a cleaned research dataset.

JA4-JEPA benchmark results

After 120 epochs of self supervised training, they froze the entire encoder. No fine tuning. They ran a simple k=11 nearest neighbour probe for protocol classification across TLS, DNS and SSH.

Cosine similarity between matching fingerprints hit 0.9899 on held out data. This is good enough that you can replace exact JA4 hash matching with these embeddings today. You will catch more obfuscated clients with zero false positive increase.

Nobody expected this. JEPA was not designed for network traffic. It just worked.

Generative LM transfers directly to drug design

Second paper: DrugGen 2. This one is even more surprising.

Every drug generation model released in the last three years used custom graph transformers, 3D attention, or specialized molecular tokenizers. Everyone agreed general purpose language models were bad at this task.

These authors did not build a custom model. They took standard 124M parameter GPT-2. The exact same one released in 2019.

They fine tuned it on a dataset that just appended disease ontology terms, target sequence, and SMILES string. That was the entire input format. No special encoding. No graph layers. Just plain text tokens.

Fine tuning followed exactly the standard LLM alignment pipeline: supervised fine tuning, then GRPO reinforcement learning. Exactly the same pipeline people use to make chatbots.

Metric	DrugGPT	DrugGen 1	DrugGen 2
Valid molecules generated	72.1%	81.4%	96.3%
Novel unique molecules	41.2%	57.8%	89.1%
Mean predicted binding affinity	-6.72	-7.41	-9.18
Docking validation pass rate	12%	21%	47%

For the diabetic nephropathy ACE target, DrugGen 2 generated three candidates with predicted binding affinity better than enalapril, the approved standard of care. All three passed independent molecular docking validation.

This is not a minor improvement. This is doubling every relevant metric using a seven year old general purpose language model.

Texture perception: ViT matches human vision, CNN does not

Third paper: texture representation comparison. This is the one that should make every computer vision practitioner stop and re-evaluate their assumptions.

For 10 years we operated under the working belief that CNNs were the most human like vision model. That belief was built entirely on object recognition benchmarks.

This paper tested something else. Texture perception. Not identifying what an object is, but identifying what it feels like, what material it is made from.

All three ViT variants had almost identical alignment to human performance. All three were more than twice as well correlated as ResNet.

Most notably, ViT alignment did not change with texture complexity. CNN performance collapsed completely for fine grained natural textures.

This is not a small difference. This is a fundamental architectural split. CNNs encode edges and contours first. ViTs encode pattern and texture first. Humans do the latter.

If you are building any vision system that is not just classifying objects, you should be using a ViT. Not because it is newer. Because it works the same way human vision works.

The unifying pattern across all three results

None of these papers found a trick. All three found an absence of tricks.

There was no secret sauce. There was no clever modification. Every single one of these results was obtainable three years ago. Nobody bothered to test it.

We have spent the last decade operating under an unstated assumption: every domain requires a custom architecture. That network security needs special network models. That drug discovery needs special molecular models. That vision needs whatever architecture won ImageNet.

That assumption was wrong.

What this means for practitioners

Stop building custom architectures. Stop reading papers about new attention variants.

If you have a domain problem today:

1. Pick the standard established paradigm for that general class of data
2. Format your data correctly for that paradigm
3. Use the exact standard training objective
4. Test it before you change anything

Most of the time it will work better than anything you could build custom.

This is not a boring result. This is the best news ML has delivered in five years. It means we are finally done inventing the foundations. We can now start actually using them.

Limitations and open questions

None of this means architecture no longer matters. It means architecture only matters at the paradigm level. The difference between JEPA and contrastive learning is large. The difference between 16 and 17 attention heads is negligible.

We still do not know the boundary conditions. We do not know which paradigms will transfer to which domains. We do not know why JEPA works on network fingerprints but SimCLR does not.

We also do not know how far this goes. There will be domains where this does not hold. But there are far fewer of those than almost everyone thinks.

Closing

This cluster of papers is not an anomaly. It is the new normal.

The era of architecture innovation is over. The era of architecture application has just started.

Most of the best ML work for the next five years will not be inventing new models. It will be discovering all the things the models we already have can already do, that nobody ever bothered to test.