Thermodynamics-inspired OOD Detection

The authors present a method for out-of-distribution detection using pre-trained image classifiers. The way they've quantified performance is clear, and figure 1 suggests their method beats an energy-based baseline, which is exciting. The presentation could be clearer - I'm not sure how they define "blockwise energy" E_\ell(x) (I think it's based off projecting to logits for each layer but am not very certain). These results add to a growing body of work showing that intermediate-layer activations may be more useful for probing tasks than final layer activations (see e.g. https://arxiv.org/pdf/2412.09563v1). Confirming/refuting such theories seems like a nice goal for a hackathon.

The project investigates energy-based out-of-distribution detection by studying an extension of that method, in which the energy is computed at every residual block. The approach makes sense and the results are presented clearly. However, no mention is made of AI safety, and the project would have benefited from articulating a clear connection between the research performed and AI safety challenges.

Very interesting work! Seeing comparisons in OOD performance with other methods would indeed be interesting, as you note. The questions with new tools is always *whether they work*. In terms of detecting OOD samples better, this is really valuable and having a strong framework for doing this seems important for nearly all relevant real-world fields of NN use, such as medical and law, providing a heuristic for rejecting answers or at least providing human-readable warnings on the outputs.

However, I think there's another extension of all this work that would be really interesting: Whether we can evaluate the fuzzy boundaries of what is OOD to the models and what isn't. And whether we can see examples of algorithmic generalization (e.g. learning addition instead of memorizing results of single-digit addition tasks) through this method. I'm not certain how feasible it is, but having a proper evaluation of generalized OOD performance that transcends single tasks (i.e. through a benchmark-based method) seems incredibly important and like a feasible extension to the concepts and tools introduced in this work.

While the method is interesting and a reasonable extension, the experimental setup is unclear and limits reproducibility. For example, it isn’t specified how intermediate block activations are mapped to class logits per block (auxiliary heads? a shared classifier? pooling + linear map?), yet the energy definition requires logits.

Block‑wise energy pooling is a tidy proof‑of‑concept, but its benefit over the OOD detector from Liu et al. was unmotivated and unclear. A minimal result could have been to compare this with the reference, and to contextualize this a bit more within the literature (as it was the only reference used). I'm not convinced that this approach has any real implications for AI safety.

The authors could also have explained their experiments and plots more carefully.