analiza aiqbench
This paper argues that a common evaluation method for Natural Language Autoencoders, a round-trip test where text explanations must reconstruct the original model activation, only shows that the explanation contains enough information to identify the activation. It does not show the explanation is accurate about what the activation represents. For readers tracking interpretability, the key point is that a high reconstruction score alone shouldn't be taken as evidence that a model's internals have been correctly explained, particularly if these tools are meant to support alignment work.
Natural Language Autoencoders Identify Without Explaining
arXiv:2610.32691v1 Announce Type: new Abstract: Natural Language Autoencoders (NLAs) express a model's internal activations in the form of readable text. The verbalizer turns an activation into an English explanation; the reconstructor then converts that explanation back into an activation, and the explanation is considered successful if the round trip brings back the original activation. We claim that this criterion evaluates whether the information is sufficient and does not address descriptive faithfulness. A sentence can contain sufficient information to make it possible to identify an activation even if it wrongly explains what the activation does: in the case of reconstruction, the primary aim is to accurately identify the activations and not to describe them. This issue has the same structure as the chain-of-thought faithfulness problem, in which verbal output that has been trained to go with a process is taken as a report on that process, and it parallels the dissociation between an individual's introspective report and the actual governing process as documented in human psychology. Although proposed as interpretability tools for the purpose of alignment, NLAs reintroduce, through unverified explanations, the epistemic problem that they were intended to resolve.
Abstract:Natural Language Autoencoders (NLAs) express a model's internal activations in the form of readable text. The verbalizer turns an activation into an English explanation; the reconstructor then converts that explanation back into an activation, and the explanation is considered successful if the round trip brings back the original activation. We claim that this criterion evaluates whether the information is sufficient and does not address descriptive faithfulness. A sentence can contain sufficient information to make it possible to identify an activation even if it wrongly explains what the activation does: in the case of reconstruction, the primary aim is to accurately identify the activations and not to describe them. This issue has the same structure as the chain-of-thought faithfulness problem, in which verbal output that has been trained to go with a process is taken as a report on that process, and it parallels the dissociation between an individual's introspective report and the actual governing process as documented in human psychology. Although proposed as interpretability tools for the purpose of alignment, NLAs reintroduce, through unverified explanations, the epistemic problem that they were intended to resolve.
Computation and Language (cs.CL)
arXiv:2610.32691 [cs.CL]
(or arXiv:2610.32691v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.32691 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 279 cuvinte
Subiecte conexe
XAInomaly: Explainable and Interpretable Deep Contractive Autoencoder for O-RAN Traffic Anomaly Detection
arXiv · 9 iun. 2026
EXPOSE: Explainable and Domain-Robust Embeddings from Pathology Vision Foundation Models using Sparse Autoencoders
arXiv · 31 aug. 2026
The Router Already Knows
arXiv · 21 oct. 2026
VERS: Steering Reasoning Models by Editing Their Activation Explanations
arXiv · 3 nov. 2026