AINEWS
Researchers building a framework to peer inside large language models want to know whether the models are actually reasoning or just producing answers that look like reasoning.
A field called mechanistic interpretability has organized around one question: when a large language model answers a question, is it computing an answer, or is it pattern-matching its way to a plausible one?
Frontier models ship into products, into regulators' inboxes, and into hospital and legal workflows before anyone can fully say what is happening inside them. Auditors want to know which behaviors are deliberate and which are artifacts of training. Regulators are being asked to write rules about systems whose internal logic is not yet understood. What gets deployed, audited, and trusted over the next few years will turn on whether the science catches up.
The field's working method is a framework called causal abstraction, laid out most rigorously in "Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability" by A. Geiger and colleagues, including Thomas Icard of Stanford, and now published in the Journal of Machine Learning Research. A preprint has been on arXiv since 2023. The framework gives researchers a way to ask whether a network's behavior at a high level is being produced by an interpretable computational structure at a lower level. If researchers can specify a candidate algorithm and show that the network's internal states implement it under intervention, the network is doing something close to that algorithm. If interventions on the candidate algorithm do not change the network's output, the analogy breaks.
Icard's working analogy, drawn from a CACM feature on the field, is the ideal gas law. Physicists do not track individual molecules; they group them into a small set of variables, pressure, temperature, volume, and write equations about the group. Mechanistic interpretability aims to do the same for neural networks, grouping many individual neurons into a small number of higher-level computational units that researchers can name and reason about.
The framework pulls several previously separate techniques under one roof. It formalizes polysemantic neurons, the linear representation hypothesis, modular features, and a notion of graded faithfulness, and it unifies activation patching, path patching, causal mediation, causal scrubbing, causal tracing, circuit analysis, concept erasure, and sparse-autoencoder-style methods in a single causal-abstraction language. In practice, a researcher studying one model can switch tools without switching theoretical commitments. A claim that a circuit in a model implements a particular operation can be checked under one framework rather than re-argued under five.
The honest limit is that real networks are messier than the gas law. A single neuron in a large model often contributes to many different functions at once. The field calls these polysemantic neurons, and they are the central reason the analogy is harder to apply in practice than in principle. Causal abstraction can still be used, but the mapping between neurons and higher-level units is no longer clean, and the same neuron can sit inside two different candidate algorithms at once. The framework accounts for this with its graded faithfulness notion, which lets researchers say how close the higher-level description comes to the underlying activity, rather than demanding a perfect match. That is progress, but it is not closure. The field has not yet produced a definitive account of when a model is reasoning and when it is sounding as if it is.
The gap is now being worked on with industrial resources. Icard is spending 2026-27 on leave at Goodfire working on the science of interpretability full-time, according to his Stanford page. The placement is a signal that the field's tools are considered load-bearing for the next round of model releases. A rigorous framework and a year of focused industrial effort are both in place. The next test is whether causal abstraction can produce concrete, reproducible claims about whether a frontier model is reasoning or only sounding as if it is, and whether regulators and auditors can be trained to read those claims.