- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper repurposes the log-alignment ratio (LAR), a prior training-time diagnostic, as a per-input monitor. At selected layers, the authors track the log input norm and log output norm of a linear map, reparameterize these as "gain" and "radius", standardize these values against in-distribution data, and score each input with a class-conditional Mahalanobis distance (U_LAP). The claims are that this score (1) detects structural shifts (contrast, noise, blur, Python code in language models, corrupted LLaVA images, corrupted robot observations) which output-confidence scores miss, and (2) adds information to entropy for error prediction (ResNet-50 0.632 to 0.811, ViT-B/16 0.706 to 0.849), and (3) raises alarms earlier than action entropy on LIBERO, and adds failure information beyond ensemble disagreement on ManiSkill. A separate section shows that alignment regularizers modify grokking delay in small transformers. Evidence is ImageNet and CIFAR detection tables, a random-basis control, a partial-correlation "structural vs. semantic" analysis, an ImageNet-200 factorial, LM and VLM pilots, and four robotics testbeds.
Critical strengths
- The monitor is cheap and was tested widely. It needs no SVD, gradients, or shifted data. Overhead is about 9-10%, and calibration storage is about 28 KB versus 0.8-1.2 GB for the Gram baseline. It is evaluated on CNNs, ViTs, LMs, a VLM and several policy classes. Several controls are good practice: the exact random-direction null in Appendix C, the confidence-matched CIFAR/SVHN split, and the 2x2 factorial. Together these make the empirical contribution useful to people deploying monitors.
- The authors are open about where the method does not win. Feature-density methods win on semantic and far-OOD sets (Table 5, Fig. 10), entropy wins on ImageNet-200 class novelty, and the appendix reports no incremental failure information on LIBERO (p=0.94). This makes the “structural, not semantic” story much more believable than a uniform-win story would be.
Critical weaknesses
- The paper isn't fully isolating 'alignment' as an informative signal, since by Prop. E.1 the joint [gain, radius] score is equivalent to a Mahalanobis score on raw per-layer [log||Wh||, log||h||], hence a re-parameterization of the same norms. Table 5 is consistent with this in that radius-only is close to the joint score on ResNet-50 (0.958/0.963/0.992/0.963 vs. 0.983/1.000/1.000/0.998), and gain-only outperforms the joint score on ViT noise (0.984 vs. 0.971). The control on random-V shows that the true ||Wh|| beats a rotated surrogate, which is useful but doesn't show that alignment specifically matters, or has per-layer output norms at all, as opposed to having them. This is fixable with additional controls.
- Some of the headline gains exploit protocols or baselines that favor a shift detector. The error-prediction result (0.632 to 0.811) fits the predictor on a clean+shifted mixture, split by example. Because shifted inputs have higher error, that partly measures shift detection rather than added information about correctness. Leave-corruption-out splits or within-condition AUROC would be more convincing.
- For language models, perplexity is a weak baseline for domain detection, and Python vs. enwik8 is an easy distinction. There is one reference corpus and one seed per model. In Table 3b, the disagreement baseline is near chance on 7 of 8 tasks (0.45-0.54), so the 0.14-0.24 gains are against a weak reference. Some of the framing is stronger than the evidence. Fig. 1's 0.880 comes from a split constructed so that MSP is at chance and has a 10-model ensemble, whereas single-model results are noisier (0.59 +/- 0.24 at M=1 for a single layer). The structural and semantic distances are author-designed, and the factorial still has a nonzero novelty response (0.37). The depth "archetypes" are an interpretive clustering of 24 sets. These are reasons to tone down the claims, not to doubt the core results.
Questions for authors
- Given Prop. E.1 plus the radius-only results in Table 5, what is evidence that alignment contributes beyond a per-layer log-norm detector? Can you add controls using different per-layer norms or random-weight layers?
- Were the probe positions, pooling choices, L values or alarm thresholds selected using any evaluation data, or the seeds reported in Table 4?
- Is the 0.880 in Fig. 1 from the joint profile or an alignment-only variant? What are the single-model numbers on that split?
- What are the results with leave-corruption-out fitting and within-condition AUROC for error prediction? How does a baseline built from simple image statistics (such as your D_s features) compare?
- How do hidden-state Mahalanobis and simple token-frequency or compression features perform on the same shifts? Does the profile track LM correctness anywhere?