Researchers Turn AI Interpretability Into a Surgical Tool to Fix LLMs

Mechanistic interpretability has spent years as a mostly observational science, peering inside transformers to explain what neurons and attention heads do without doing much about it. A sprawling new survey from researchers at Hong Kong, Fudan, LMU Munich, Tsinghua and more than a dozen other institutions wants to change that. It argues that interpretability tools are mature enough to actively steer, edit, and optimize large language models, and it organizes over 200 papers into a single actionable pipeline: locate, steer, and improve.

Scaling is hitting familiar walls. Progress in LLMs has been driven mostly by external factors like larger models, more data, and more compute, and that paradigm is running into bottlenecks. The authors argue the next round of gains has to come from surgically modifying models from the inside rather than piling on more hardware.

Why past surveys fell short

Prior interpretability reviews tend to blur diagnostic tools with intervention techniques, and their application coverage is often too abstract to translate into concrete engineering work. This survey fixes that by first defining what it calls interpretable objects, the specific pieces of a decoder-only transformer you can actually poke at, and then splitting methods cleanly into diagnosis (localizing) and intervention (steering).

Those objects are the usual suspects, formalized with consistent notation: token embeddings, the residual stream, multi-head attention with its QK and OV circuits, feed-forward neurons, and sparse autoencoder (SAE) features. SAEs get particular attention because a single neuron may activate for multiple unrelated concepts, making direct interpretation difficult due to superposition. SAEs disentangle these dense polysemantic representations into monosemantic features you can actually target.