- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper derives exact decompositions and directional-budget bounds for normalized hidden-state trajectories, then compares refusal and jailbreak trajectories across four models.
Critical strengths
The exact decomposition provides a transparent description of representation dynamics, supported by cross-fitted measurements across multiple models.
Critical weaknesses
1 The central conceptual connection remains unvalidated. Definition 3.2 characterizes structural hallucination through information-theoretic and behavioral conditions, whereas the experiments measure directional projections. The paper does not establish that these geometric quantities diagnose the stated phenomenon. 2 Theorem 3.5 applies to arbitrary normalized trajectories. Its additional explanatory value for alignment failure is not demonstrated: separating outcomes along a refusal-trained direction does not establish that the proposed budget reveals anything beyond simpler refusal-projection measurements. 3.Too many details essential to assessing the central claims are deferred to the appendix, leaving the main paper insufficiently self-contained.
Questions for authors
What result connects the directional budget to Definition 3.2? Does the budget provide information about behavioral failures beyond the underlying refusal projection?
4 comments
Thank you again for recognizing our "transparent description of representation dynamics."
- W1 / Q1: We replaced the Introduction's "measurable signature" wording with a description of the measured safety coordinate. Section 3.2 now states explicitly that these measurements do not establish all conditions of Definition 3.2.
- W2 / Q2 / W3: We added an explanation of the endpoint lower bound and path remainder after Theorem 3.5, and connected the theory to Table 1 and the budget analysis in $\S$4.2.1.
In light of these revisions, we respectfully ask you to consider raising your rating.
Thank you for recognizing our "cross-fitted measurements across multiple models" and for asking what the geometry adds beyond the refusal projection (W2/Q2). The full projection and retention sequences determine the directional budget. In our early-layer prediction check, the paired 95% confidence interval for the AUROC gain from adding budget summaries to those features included zero in each model. Rather than serving as a stronger predictive feature, the geometry isolates two structural properties along the trajectory:
- At a single layer (Eq. 5). Measured retention separates the projection carried forward from what the next layer adds along the refusal direction. Over the middle layers, jailbroken trajectories had higher mean retention in all four models (Appendix Table 4)—an empirical sign pattern not prescribed by the identity itself. Table 1 shows that larger tangent contributions accounted for the refusal group's greater mean projection gain, while retention partly offset that difference. The tangent ordering persisted after adjustment for layer and current projection.
- Across layers (Table 2, Theorem 3.5 & Proposition 3.6). The refusal group also accumulated a larger positive directional budget in each model (Table 2, Hedges' from 1.00 to 1.84). For the total directional budget, Theorem 3.5 gives a lower bound fixed by the endpoint projections and retentions, while its remainder measures how much additional directional writing the observed path used. We verified the attainment condition in Proposition 3.6 across all 1,600 archived middle-layer windows. Across models, the median fraction of the observed budget needed to reach the same endpoint projection with the same retentions ranged from 3.3% to 28.1% (Appendix C.3.4).
In the revised manuscript, the Introduction and $\S$3.2 clarify how the measured coordinate relates to Definition 3.2 (W1) and explain the budget decomposition directly beside Theorem 3.5 (W3). Is there a specific concern about the geometric result or its relevance to alignment that still weighs on your recommendation? In light of these clarifications, we hope you will consider raising your rating.
Thank you for confirming that our revisions address W1 and W3, and that the single-layer retention/tangent analysis adds information beyond endpoint projections.
To directly test your remaining question on the endpoint-conditioned path remainder (Theorem 3.5), we evaluated across the 1,600 archived middle-layer windows (Appendix C.3.4) and regressed on outcome () adjusting for :
| Model | Raw Mean (Ref / JB) | Adjusted (JB − Ref) | Permutation |
|---|---|---|---|
| Llama-2-7B | 0.13297 / 0.04106 | +0.02295 | 0.00030 |
| Llama-3.1-8B | 0.16969 / 0.03867 | +0.02168 | 0.00080 |
| Qwen2.5-7B | 0.04265 / 0.01384 | −0.00110 | 0.70053 |
| Qwen2.5-14B | 0.05106 / 0.01790 | +0.00939 | 0.00070 |
Unadjusted is larger for refusal in all four models (, Hedges' –). After adjusting for , the full-sample estimate reverses sign in Llama-2-7B, Llama-3.1-8B, and Qwen2.5-14B (, remaining positive on the log scale), while for Qwen2.5-7B the original-scale interval crosses zero (, with a negative contrast on the log scale). Restricting the regression to the overlapping endpoint region also shifts the magnitude—for example, from to in Qwen2.5-14B.
This additional check tests a further question beyond the claims made in our paper. Our contributions match the results that hold consistently across all four models—namely, the exact geometric decomposition and attainable budget framework (Eq. 5–6, Theorem 3.5, Proposition 3.6), together with the uniform cross-model separation in single-layer retention/tangent updates (Table 1, Appendix Table 4) and accumulated positive directional writing (Table 2). By contrast, our paper does not claim a uniform outcome-group separation in the endpoint-conditioned remainder , and the check above confirms that any such difference varies with model and endpoint overlap. As the discussion period nears its close, we will include this analysis in the revision to state that scope explicitly.
We genuinely appreciate your rigorous reading and the time you have spent sharpening our analysis. Since W1 and W3 are resolved, this additional cross-layer boundary is now directly tested and delineated, and our stated core contributions remain intact, we respectfully ask whether you might consider raising your rating accordingly. Thank you again for your thoughtful feedback.
Response to Reviewer 6pNB
Thank you for recognizing the "transparent description of representation dynamics" and the "cross-fitted measurements across multiple models".
W2 / Q2. What the geometry adds.
Thank you for your question. The budget can distinguish paths with the same endpoint projections and retentions. It is determined once the full projection and retention trajectories are given. Our additional prediction experiment did not establish a gain in area under the receiver operating characteristic curve (AUROC) over projection-and-retention features.
The orthogonal decomposition and the Cauchy-Schwarz inequality are classical. Theorem 3.5 expresses the budget as a lower bound fixed by endpoint projections and retentions plus a weighted sum of squares measuring deviation from the minimizing contributions. Proposition 3.6 determines when this bound is attainable. The reported middle-window analysis verifies attainability and gives model-wise median minimum-to-observed budget ratios from 3.3% to 28.1%.
The theorem's generality makes these quantities available without assuming a specific stochastic law for layer updates. The alignment-specific result is the measured comparison along the refusal-associated direction.
Retention and tangent contributions. Table 1 addresses which part of the update accounts for the observed group difference. The refusal group's larger mean projection gain over the prespecified middle-layer window is carried by the tangent term. Retention partly offsets that difference in all four models. With cross-fitted directions and adjustment for layer and current projection on common projection support, the jailbroken group still has a lower fitted tangent contribution. The comparison therefore examines how the groups differ in their next-layer contributions at comparable current projections.
Prediction check. The test used the first half of the network. All four paired 95% confidence intervals for the AUROC gain from adding budget summaries include zero.
W1 / Q1. Connection to Definition 3.2.
Thank you for asking us to make this connection precise. The directional budget does not establish all conditions of Definition 3.2. Section 3.1 uses the measured projection as a coordinate of the internal constraint state. Theorem 3.5 gives an exact geometric identity for its trajectory, and M2 compares its tangent contributions across response groups after adjustment for layer and current projection. These conclusions do not require the definition's information-theoretic conditions.
We will replace "measurable signature of structural hallucination" with a description of the measured coordinate dynamics. This will make clear that the empirical analysis studies how a refusal-associated coordinate propagates, without claiming to verify the full definition.
W3. Main-text self-containment.
Yes. We will explain the endpoint lower bound and the path remainder alongside Theorem 3.5, and state explicitly that these measurements do not verify all conditions of Definition 3.2. We will also make the link to Table 1's empirical comparison clear in the main text. Detailed proofs will remain in the appendix.