- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies a precise form of robustness for autoregressive chain-of-thought generation: how many initial reasoning tokens can be guaranteed to remain exactly unchanged when a bounded local perturbation is applied to one input representation. The paper derives position-dependent finite-perturbation propagation bounds for causal Transformers and combines these bounds with clean next-token margins to obtain sufficient conditions for exact prefix preservation. The resulting certificate can be viewed in terms of certified prefix length, certified perturbation radius, or certifiable input positions (Sec. 3.2–3.3). Empirically, the paper evaluates the predicted position- and magnitude-dependent stability patterns on four reasoning models and three benchmarks, including matched-magnitude controls (Sec. 4.1–4.3, Appendix B). It also gives an end-to-end numerical realization on a controlled two-layer Transformer, including verified constants, certified radii, tightness analysis, and direct generation checks (Sec. 4.4, Appendix C). I find the formulation and position-aware analysis useful, although the gap between the formal certificate and practical certification of contemporary reasoning models remains the main limitation.
Critical strengths
- S1. The paper isolates a concrete and nontrivial robustness question at the trajectory level rather than reducing robustness to final-answer agreement. Defining robustness as exact preservation of the first K greedy-generated tokens under every perturbation in a local representation-space ball makes the object being certified precise and mathematically testable (Sec. 3.1, Eqs. 1–4). This is useful because autoregressive reasoning can diverge long before the final answer changes, so a prefix-level notion exposes a behavior that answer-level robustness can miss.
- S2. The main technical contribution is genuinely position-aware: the propagation analysis retains the source position of the perturbation instead of collapsing it into a position-agnostic worst case. Theorem 3.1 and Corollary 3.2 produce a logit-drift coefficient A(i,t,L) whose dependence on the perturbation position leads directly to the prefix-length, perturbation-radius, and input-position certificates in Sec. 3.3 (Theorem 3.3, Propositions 3.1–3.2). Appendix C.3 further shows that retaining source-position information improves the realized certified radius over a position-agnostic certificate, which makes the theoretical refinement consequential rather than merely cosmetic.
- S3. The large-model experiments are unusually well aligned with the structural predictions of the theory. Across four reasoning models and three benchmarks, the preserved-prefix length increases from early to late input locations under two independently drawn token substitutions, and the trend remains after matching absolute representation-space perturbation magnitudes across positions (Sec. 4.2, Fig. 2, Table 1; Appendix B.3–B.4). The matched-magnitude control is particularly important because it helps separate the claimed positional effect from the trivial possibility that later-token replacements simply induce smaller representation changes.
- S4. The magnitude-prefix experiments provide a second, distinct empirical check of the certificate geometry rather than reusing the same positional observation. Prefix retention decreases as perturbation magnitude increases across the tested target horizons, and longer target prefixes are systematically harder to retain (Sec. 4.3, Fig. 3; Appendix B.5). This supports the qualitative dependence on both perturbation budget and target prefix length that appears in the certified-radius formulation.
Critical weaknesses
-
W1: The strongest formal certification result is not instantiated on the contemporary reasoning models used in the main empirical study. Sections 4.2–4.3 test whether observed trajectory stability follows the predicted position/magnitude geometry, whereas the actual end-to-end certificate in Sec. 4.4 and Appendix C is realized only on a two-layer synthetic arithmetic Transformer. Moreover, the certified radii in that realization are extremely small (e.g., the minimum radius at K = 32 is 3.397 × 10^-12 in Table 10). This does not invalidate the theory, but it leaves unclear whether the certificate is practically non-vacuous at perturbation scales relevant to the four large reasoning models.
-
W2: The usefulness of the guarantee depends strongly on assumptions and uniform constants that may be conservative for realistic Transformers. In particular, the propagation result relies on A1–A4, including a diffuse-attention bound of the form c_alpha/(j-1), together with uniform Lipschitz/norm constants over the relevant perturbation domain and generation horizon (Sec. 3.2, Appendix A.1). Appendix C.3 already shows that the closed-form relaxations reduce certified radii by roughly a factor of 2.5 relative to the direct recurrence in the controlled model. I therefore remain unsure how much of the favorable position dependence survives numerically after obtaining valid constants for deep production-scale models.
Questions for authors
-
Q1: Have the authors attempted to compute even approximate or conservative certificate quantities (e.g., A(i,t,L), certified radii, or certified prefix lengths) for any of the four reasoning models in Sec. 4.1? If this is computationally infeasible or immediately vacuous, it would be helpful to quantify where the bottleneck occurs.
-
Q2: How restrictive is assumption A4 in the actual attention patterns of the four evaluated reasoning models? In particular, can the authors report empirical values or upper bounds for the effective diffuse-attention factor c_alpha over the positions/horizons used in Sec. 4, and indicate how sensitive the resulting certificate would be to this constant? This would help distinguish whether the current limitation is mostly numerical looseness or a more fundamental mismatch between the assumptions and realistic attention behavior.
Dear Reviewer Yi1Y,
Thank you again for your careful review and constructive feedback. In our point-by-point responses and revised manuscript, we have provided additional certification results, direct propagation measurements, and controls for generation-stage and context-dependent effects.
As the discussion period approaches its conclusion, we would appreciate any feedback you may have on whether these additions adequately address your concerns. If further clarification would be helpful, we would be happy to provide it.
Thank you for your time and consideration.
Best regards,
The Authors