What does Google's "Learning without training" paper reveal about how LLMs work internally?

The paper shows that passing input through a Transformer's feed-forward network, combined with the context provided by attention, mathematically behaves like creating a temporary, input-specific weight matrix. This means a large language model can adapt dynamically to each prompt even though its trained weights stay completely fixed, revealing a form of implicit in-context plasticity behind its apparent contextual intelligence.

What is in-context learning in a large language model?

In-context learning is a large language model's ability to adjust its behavior based solely on the examples and instructions given in a prompt, without any additional training or weight updates. Google's research suggests this happens because attention and the feed-forward layers jointly act as if they briefly rebuilt a specialized weight matrix for that specific input, rather than because the model literally learns something new.

How do Mixtures-of-Experts (MoE) and Hyper-Networks relate to dynamic weight adaptation in Transformers?

Mixtures-of-Experts activate only a few specialized sub-models for each input through a gating mechanism, cutting compute cost while allowing contextual specialization, while Hyper-Networks generate the weights of another network on the fly instead of predicting outputs directly. Both approaches align with the idea that Transformers already perform implicit, input-specific weight adaptation, suggesting future architectures could make that adaptation explicit and more controllable.

Are large language models fully understood by researchers today?

No. The mathematical description of dynamic weight-like behavior in Transformers is a starting point for explainability and targeted optimization, not proof that LLMs are fully understood. The underlying paper had not yet been peer-reviewed at publication, and turning this formal insight into diagnostic tools or a unified theory of how LLMs work still requires substantial further research.

Google Opens the Hood of LLMs

In a recent paper (“Learning without training: The implicit dynamics of in-context learning”), Google’s teams propose a mathematical dive into the inner workings of Transformers.

Quick technical reminder

An LLM consists of stacked Transformer layers. In each layer, we find two main blocks:

Multi-Head/Self-Attention: This mechanism allows the model to evaluate the relative importance of each word compared to others, to capture context.
Feed-Forward Network (FFN): A denser processing step where information is deeply transformed through activation functions, which modulate the signal intensity of each neuron.

But what’s fascinating is that the interaction between these layers isn’t limited to simple sequential processing.

The discovery: dynamic plasticity

The paper demonstrates that passing through the FFN, combined with the context brought by attention, mathematically amounts to temporarily creating a weight matrix specific to the input. In other words, the architecture allows dynamic adaptation to each situation, a form of temporary plasticity, only triggered by the prompt.

Even with fixed weights, the model seems to adapt in real-time to each new input. A form of contextual intelligence!

The demonstration (not yet peer-reviewed) aims to better understand this black box, to eventually develop diagnostic techniques, “explicability,” and even targeted optimization.

The opened perspectives

This formal framework opens several concrete paths:

Designing conditioned inference mechanisms, for adaptability without fine-tuning.
Testing hybrid and modular architectures, in connection with approaches like Hyper-Networks or Mixtures-of-Experts.
And, why not, advance toward a unified theory of LLMs – which would finally explain why it works so well.

However, be careful not to confuse mathematical description with deep understanding: these observations should be the starting point, not the conclusion, of our understanding of LLMs.

Conclusion

In short, the models are powerful, but it’s still human intelligence that will need to extract their full potential.

The experts’ corner 🤓

Mixtures-of-Experts (MoE)

Technical definition: An architecture in which multiple sub-models (“experts”) are available, but only a few are activated for each input, according to a selection mechanism (gating), allowing specialization of calculations while controlling costs.

Advantages:

Significant reduction in computational costs
Adaptive specialization according to context
Improved scalability

Concrete example: A MoE model could have experts specialized in different domains (medicine, code, literature) and only activate the most relevant ones according to the task.

Hyper-Networks

Technical definition: A network that doesn’t make predictions directly, but dynamically generates the weights of another main network, allowing the latter to adapt to each task or context without additional training.

Potential applications:

Rapid adaptation to new domains
Personalization without retraining
Efficient meta-learning

Analogy: A chef who adapts their recipe in real-time according to available ingredients, without needing to relearn how to cook.

References

Learning without training: The implicit dynamics of in-context learning