acceptodds
Under review as a conference paper at ICLR 2027

Why Safety Alignment Can Fail: A Geometric Theory of Structural Hallucination in LLMs

Abstract

Safety alignment can fail under jailbreak even measured task capability remains intact, but a change in a safety projection alone does not reveal whether the representation moves farther or moves in a different direction. We give an exact geometric decomposition of every normalized inter-layer update into direction-independent retention and a direction-specific tangent contribution, derive the sharp reachable interval of a single update, and prove a finite-depth directional-budget bound with a closed-form characterization of equality. Exponential projection decay follows as the constant-retention, zero-budget special case. Across four Llama and Qwen models on HarmBench, refused trajectories receive systematically larger safety-directed tangent contributions than jailbroken trajectories, with refused-minus-jailbroken effect sizes from 1.3 to 2.3. The retention contrast has the opposite sign and partly offsets the tangent contrast. The tangent ordering remains after adjustment for layer and current projection. A separate Llama-3.1-8B GSM8K experiment observes no statistically distinguishable accuracy difference between clean prompting and the tested jailbreak prefix. These results support analyzing safety maintenance through both retention and direction-specific writes, clarifying how task computation can persist while constraint transmission weakens.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
mechanism analysis, Safety Alignment, geometrical analysis, jailbreak

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6pNBOpenReview ↗
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
AuthorsOpenReview ↗

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%3.3\% to 28.1%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%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.

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' gg 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 R:=GL(v)−DL(v)2/(1−Kc(0,L)2)R := \mathcal{G}_L(v) - D_L(v)^2 / (1 - K_c(0, L)^2) (Theorem 3.5), we evaluated RR across the 1,600 archived middle-layer windows (Appendix C.3.4) and regressed RR on outcome (JB−Ref\mathrm{JB} - \mathrm{Ref}) adjusting for (ρ0,ρL,Kc(0,L))(\rho_0, \rho_L, K_c(0, L)):

ModelRaw Mean RR (Ref / JB)Adjusted γ^rem\hat{\gamma}_{\mathrm{rem}} (JB − Ref)Permutation pp
Llama-2-7B0.13297 / 0.04106+0.022950.00030
Llama-3.1-8B0.16969 / 0.03867+0.021680.00080
Qwen2.5-7B0.04265 / 0.01384−0.001100.70053
Qwen2.5-14B0.05106 / 0.01790+0.009390.00070

Unadjusted RR is larger for refusal in all four models (p=1/10001p = 1/10001, Hedges' g=0.99g = 0.99–1.601.60). After adjusting for (ρ0,ρL,Kc(0,L))(\rho_0, \rho_L, K_c(0, L)), the full-sample estimate reverses sign in Llama-2-7B, Llama-3.1-8B, and Qwen2.5-14B (p≤0.0008p \le 0.0008, remaining positive on the log scale), while for Qwen2.5-7B the original-scale interval crosses zero (p=0.70053p = 0.70053, with a negative contrast on the log scale). Restricting the regression to the overlapping endpoint region also shifts the magnitude—for example, from +0.00939+0.00939 to +0.00018+0.00018 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 G+\mathcal{G}^+ (Table 2). By contrast, our paper does not claim a uniform outcome-group separation in the endpoint-conditioned remainder RR, 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.

Reviewer_d7EuOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper investigates how LLM models can be tricked into giving harmful responses while still retaining their ability to perform useful tasks. The authors develop a mathematical framework for tracking how refusal-related signals change throughout the model layers, and find that successful jailbreaks are associated with weaker reinforcement of these signals in several Llama and Qwen models. The paper also examines whether task performance is preserved, but only through one model, one math benchmark, and one fixed jailbreak prefix. Although that experiment finds no statistically significant accuracy difference, it does not establish that performance is unaffected in general, or that task capabilities are preserved. The broader interpretation in the abstract therefore seems overstated: “These results support analyzing safety maintenance through both retention and direction-specific writes, clarifying how task computation can persist while constraint transmission weakens.” (L24).

Critical strengths

I believe the main research question the authors pose, as I understood it, is important and timely: why can a jailbreak cause a model to violate safety constraints while retaining its ability to perform the task? And more specifically, does this happen because safety-related information is lost across layers, or because later layers fail to reinforce it? However, I do believe the paper as it is currently written obscures these questions and its main contributions behind dense, jargon-heavy, and difficult-to-follow prose (see below).

Critical weaknesses

Severe clarity and presentation issues

The paper presentation substantially hinders being able to properly assess the main claims and contributions. The abstract and introduction introduce numerous specialized terms that are poorly defined, undefined, or lack sufficient context for the paper narrative to read like a clear and logical progression of ideas. The mathematical results are also too dense, and could be synthesized into a few main takeaways, or a "proof sketch" of the approach without loss of impact, while the lengthy derivations could be offloaded to the appendix. As a result, it is difficult to discern the paper's core contribution and its significance, follow the arguments, and determine how the theoretical analysis supports or complements the empirical claims.

Questions for authors

Can’t really tell what the authors did here: “We give an exact geometric decomposition of every normalized inter-layer update into directionindependent retention and a direction-specific tangent contribution, derive the sharp reachable interval of a single update, and prove a finite-depth directional-budget bound with a closed-form characterization of equality. Exponential projection decay follows as the constant-retention, zero-budget special case” (L13-14).

L20: “refused-minus-jailbroken” feels hectic and AI-ism; can you state simpler? You mean the effect size of the observed gap between refused trajectories, vs. trajectories under jailbreak?

Unclear; “The retention contrast has the opposite sign and partly offsets the tangent contrast. The tangent ordering remains after adjustment for layer and current projection.” (L20-21).

L25: “constraint transmission” feels ill-defined, what does this mean?

L31: what does the citation bracket do here? Are these works referenced with regards to scaling LLM capacity, or public access deployment?? “As Large Language Models (LLMs) scale in capacity and are deployed with public access (Brown et al., 2020; Kaplan et al., 2020; Hoffmann et al., 2022),” — seems like the citations are a bit premature here, and if not, please specify what specifically stands out in these works.

L47: unclear language and difficult to follow logic: “Paired reference and jailbreak representations also separate across depth (Figure 1), but this descriptive separation does not distinguish two possibilities: successive layers may retain less of the preceding state, or their new contributions may point less strongly along the safety direction.”

L66-73: “This gives a sharp interval for one-step projection updates and a finite-depth lower bound on the directional write budget required to depart from passive propagation. The familiar exponential decay of a fixed-direction projection is recovered when this budget is zero and retention is constant, a passive reference rather than a prescribed shape for observed trajectories. This geometry supplies a measurable signature for structural hallucination, the semantic-level failure in which constraint transmission weakens while stable behavior violates the constraint.” — this entire passage reads very contrived and unclear; I understood essentially none of it or how it relates to the rest of the introduction

L77: what is “the safety-coordinate gap”? A lot of terminology is being introduced in this hyphenated composition of several known terms, but undefined — you need to define terms your results presentation or core claims rely on upfront.

L108 (and throughout): can you define “the residual stream” for the reader where first mentioned?

3 comments
AuthorsOpenReview ↗

Response to Reviewer d7Eu

Thank you for calling our research question "important and timely" and highlighting the distinction between losing safety-related information and failing to reinforce it across layers.

W1. Presentation and proofs.

We must point that w are no proofs in the main text. All proofs are already in the appendix. The main text contains the key equations, result statements, and their interpretations, with explicit references to the proofs.

Q1, Q6, Q7. What the geometry explains.

Jailbroken trajectories have higher mean retention, yet weaker refusal-directed contributions, in all four models over the prespecified middle-layer window (Table 1 and the appendix table "Measured Middle-Window Retention by Group"). This shows why weaker refusal signals cannot simply be attributed to lower retention. The decomposition separates what is retained from what is added along the refusal direction, making the two explanations testable. The identity itself does not prescribe the observed signs.

Figure 1 compares paired representations at the same layer. Equation (5) follows each trajectory across adjacent layers, separating the retained projection from the projection of the component orthogonal to the preceding state.

The theory connects these individual updates to the whole path. The sharp interval gives the possible next projections at fixed current state and retention. Theorem 3.5 separates a lower bound fixed by endpoint projections and retentions from the path's additional budget, and Proposition 3.6 states when that bound can be attained. This distinguishes paths with the same endpoint projections and retentions. Exponential decay follows when tangent contributions vanish and retention is constant in (0,1)(0,1). The empirical analysis allows both to vary.

Q2, Q3. Group comparison and opposite signs.

The groups are defined by response outcome: completions providing substantive harmful assistance are labeled jailbroken, and the rest refused. "Refused-minus-jailbroken" is a difference in group means. Hedges' gg is reported separately.

Table 1 reports positive net and tangent contrasts and a negative retention contrast. Thus retention partly offsets the tangent difference, whose Hedges' gg ranges from 1.331.33 to 2.332.33. The jailbroken group still has a lower fitted tangent contribution after adjustment for layer and current projection on overlapping projection ranges. This is the adjusted ordering stated in the abstract.

Review summary. Capability scope.

The GSM8K result is limited to Llama-3.1-8B and the tested prefix (0.800.80 clean versus 0.790.79 prefixed accuracy). A statistically indistinguishable difference does not establish general capability preservation. We will state that scope consistently. The geometric results and M2 comparison require no assumption of general capability preservation.

Q4, Q8. Measured quantities.

The safety coordinate ρl\rho_l is the normalized state's projection onto the measured refusal-associated direction. Table 1 compares mean per-layer changes Δρl\Delta\rho_l. We will call this the "gap in projection change." Empirically, "constraint transmission" describes how this coordinate is carried and updated across layers.

Q5. Opening citations.

We have rechecked these references and consider them appropriate for the scaling claim. Brown et al. study few-shot performance with scale, while Kaplan et al. and Hoffmann et al. study language-model scaling and compute allocation.

Q9. Residual stream.

The term is used correctly in its standard Transformer sense. It denotes the token representation carried across blocks and updated by attention and MLP branches. Related Work cites studies using this term, including Zhao et al. (2024).

Thank you again for recognizing the research question as "important and timely."

  • W1 / Q1 / Q6 / Q7: We added an explanation of the budget decomposition after Theorem 3.5 and clarified the measured-coordinate interpretation in the Introduction. The proofs were already in the appendix.
  • Q2 / Q3 / Q4 / Q8: We clarified the response groups and the gap in projection change in the Introduction, Section 4.2.1, and Table 1. We also aligned the safety-direction and tangent-contribution terminology across the main text and appendix.
  • Q5: We moved the scaling citations directly after "scale in capacity" in the Introduction.

We would appreciate your considering a improve the rating in light of these revisions and our response.

Dear Reviewer d7Eu,

As the discussion period is drawing to a close, we wanted to check in to see if anything in our responses or revised manuscript remains unclear. In the revision, we clarified the definitions of the response groups, projection-change gaps, and safety-coordinate terminology throughout the Introduction, Section 4.2.1, and Table 1, added an explicit explanation of the budget decomposition after Theorem 3.5, adjusted the scaling citations in the Introduction, and stated the exact scope of the GSM8K capability check.

Please let us know if there are any remaining questions we can address before the deadline. If these revisions have helped clarify the paper’s presentation and contributions, we would greatly appreciate it if you could consider updating your assessment and scores.

Reviewer_znLhOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates why llms can still complete their original tasks under jailbreak attacks while ceasing to follow safety constraints. The authors argue that understanding this phenomenon requires more than examining whether safety information exists in hidden states or whether the projection onto the safety direction decreases. Instead, one should further analyze, across the model’s layer-wise computation, how much of the safety-related state from the previous layer is retained and how much of the current layer’s update aligns with the safety direction. The paper proposes a geometric decomposition that splits the layer-to-layer evolution of hidden states into two components. Evaluations on four Llama and Qwen models verify that refusal trajectories and jailbreak trajectories exhibit systematic differences in tangential updates along the safety direction.

Critical strengths

The core observation of this paper is that the preservation of a model’s task capability does not guarantee sustained enforcement of safety constraints with greater theoretical depth. It seeks to explain how the same model, within a single forward pass, can retain task-relevant computational capacity while potentially generating content that violates safety constraints.

Critical weaknesses

The authors first split trajectories into refusal and jailbreak groups based on final outputs, then compare their tangential contributions along the safety direction. Even if the difference is statistically significant, the final behavioral label and hidden states may both reflect confounding factors such as prompt structure, other internal model features, or distinct computational paths. This does not establish that insufficient tangential contribution causes jailbreak.

Questions for authors

Why do geometric dynamics on the unit sphere serve as a valid proxy for the propagation of safety information within a Transformer?

3 comments

I am sorry for the oversight where I swapped the reviews of the two papers. I have updated it to the correct version. My sincere apologies again.

Thank you for updating the review and for emphasizing that preserving task capability "does not guarantee sustained enforcement of safety constraints."

W1. Confounding and causal interpretation.

This has been an important concern throughout our study. The original analyses address three specific issues:

  1. Data reuse: cross-fitting estimates each direction on separate training prompts. Evaluation prompts and their outcome labels are excluded from that fit.
  2. Differences in layer and current projection: the regression adjusts for both. The common-support estimate uses only projection ranges represented in both outcome groups, avoiding extrapolation beyond their overlap.
  3. Direction specificity: we compare the adjusted contrast along the refusal-associated direction with the same statistic computed using label-free random directions. The observed contrast lies outside the sampled reference range.

The lower fitted tangent contribution in jailbroken trajectories persists across all four tested Llama and Qwen models. Averaging the retention and tangent terms in Equation (5) within each group shows that the tangent term accounts for the refusal group's larger mean projection gain, while retention partly offsets it (Table 1). These controls do not match the outcome groups on prompt structure, other hidden-state features, or computational paths. Those factors may still affect both updates and outcomes, leaving the causal effect of changing the tangent contribution unidentified.

We will make this boundary explicit alongside Table 1, consistent with §\S 3.1.

Q1. Why use unit-sphere geometry to study safety-related propagation?

It is appropriate for the directional component we measure: the decomposition is exact after normalization, and the chosen direction's association with refusal is evaluated on held-out prompts. These modeling choices follow established practice. Prior work uses unit-sphere geometry in Transformer models [1] and studies refusal through a linear residual-stream direction [2]. For each nonzero recorded state h~l\widetilde h_l, we use hl=h~l/∥h~l∥2h_l=\widetilde h_l/\|\widetilde h_l\|_2. Equation (5) then separates changes in its projection onto the fixed refusal-associated direction into retention and tangent contributions.

The theory supplies quantitative constraints on this coordinate. The sharp projection interval bounds the next projection at fixed current state and retention. Across layers, Theorem 3.5 decomposes the directional budget into a lower bound fixed by endpoint projections and retentions plus a nonnegative remainder. Proposition 3.6 states when the bound is attainable. These results characterize the measured normalized trajectories without requiring a stochastic model for layer updates. The appendix random-direction control supports the choice of direction: its adjusted group contrast lies outside the label-free random-direction null in all four models (pdir=0.002p_{\mathrm{dir}}=0.002). Normalization discards magnitude, and a single coordinate need not capture all safety information.

[1] A mathematical perspective on Transformers

[2] Refusal in Language Models Is Mediated by a Single Direction


In fact, our contribution is to connect layer-wise representation geometry with the maintenance of safety-related signals. We hope this connection offers a complementary perspective for understanding model safety and informs efforts to build more reliable models. Thank you for updating the review again!

If these clarifications resolve your concerns, we would appreciate your considering a higher rating.

Thank you again for emphasizing that preserving task capability "does not guarantee sustained enforcement of safety constraints."

  • W1: Section 4.2.1 now explicitly states that trajectories are grouped by final response outcomes and that the adjusted contrasts describe observational associations.
  • Q1: We added the explicit normalization of nonzero residual states in Section 3.1 and explained the endpoint lower bound and path remainder alongside Theorem 3.5 in $\S$3.2.

If these clarifications address your concerns, we would appreciate your considering a higher rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Response and Revisions

We thank all reviewers and the AC for their thoughtful feedback. Below we summarize our paper's core contributions, the revisions already incorporated into the manuscript, and the additional verification completed during the discussion.

1. Our Contributions

  • Exact Spherical Geometry and Attainable Budget: We formalize normalized residual stream evolution through an exact single layer retention and tangent decomposition and finite depth directional write-budget decomposition with closed-form attainability conditions (Theorem 3.5, Proposition 3.6). The attainment condition is strictly verified across all 1,600 archived middle-layer windows.
  • Consistent Cross-Model Empirical result: Across all models, our decomposition uncovers an empirical sign pattern not prescribed by the identity itself: jailbroken trajectories exhibit higher mean inter-layer retention (Appendix Table 4), ruling out faster passive decay. Instead, refusal projection gains are driven specifically by stronger safety directed tangent contributions, persisting after adjusting for layer and current projection (Table 1).

2. Revisions and Resolved during Discussion

  • Self-Contained Presentation and Sharpened Scope We expanded the intuition and diagnostic interpretation of the write-budget decomposition directly in the main text, keeping full proofs cleanly in the appendices already, clarified the exact link between our measured safety coordinate ρl\rho_l and Definition 3.2, and formalized the observational identifiability boundary for outcome-grouped forcing (Proposition C.10). During the discussion, Reviewer 6pNB confirmed that these revisions address the concerns regarding Definition 3.2 scoping (W1) and budget presentation (W3), and recognized that the retention/tangent analysis provides information beyond endpoint projections alone.
  • Supplementary Cross-Layer Remainder Check (to be added in the final revision): In response to Reviewer 6pNB's follow-up on Theorem 3.5, we evaluated the path remainder R:=GL(v)−DL(v)2/(1−Kc(0,L)2)R := \mathcal{G}_L(v) - D_L(v)^2 / (1 - K_c(0, L)^2) across the 1,600 archived windows. We will include this supplementary analysis to document this boundary explicitly, as our primary claims rest on the uniform 4/4 results above rather than universal separation in the conditioned remainder.

Overall, our core theoretical framework and consistent all model empirical results remain fully intact, with all major presentation and scoping questions directly addressed. We thank the reviewers and Area Chair again for their careful consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.