acceptodds
Under review as a conference paper at ICLR 2027

On the Stability of Causal Attribution Across Language Model Continuations

Abstract

An internal state’s measured influence on a language model’s behavior may depend on the continuation that follows it. We test this dependence using error acknowledgment in mathematical reasoning. Using matched clean–error traces, we replace regional key–value cache states after prefill across eight checkpoints in thinking and non-thinking modes. Effects on acknowledgment concentrate in the localized error span and post-error states, with thinking generally increasing relative error-span sensitivity. Regional effects on an early error-related projection differ systematically from those on acknowledgment: error-span replacement has a larger relative effect on acknowledgment, whereas post-error replacement shows the reverse pattern. Across six checkpoints, replaying natural thinking prefixes partially reproduces the thinking-associated increase in error-span sensitivity under non-thinking generation, while controlled artificial prefixes show that selected continuation content changes the balance between error-span and post-error sensitivity. These prefix interventions hold the checkpoint, original prompt prefill, and target decoding mode fixed. Together, these findings show that, in this setting, regional causal attributions obtained through cache replacement depend on subsequent continuation content and need not transfer unchanged across continuations.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic interpretability, Internal mode monitoring, Language model reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9cPtOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies whether regional KV-cache replacement effects depend on subsequent continuation content. Using matched correct–incorrect GSM8K solutions across eight checkpoints, it compares thinking modes, early error-related projections, and later error acknowledgment. Natural and controlled prefix interventions show that continuation content can change relative regional sensitivity.

Critical strengths
  1. The matched examples and fixed-mode prefix interventions provide a controlled setting for testing whether continuation content changes cache-replacement effects.
  2. Evaluation across multiple model families, together with decoding and judge robustness checks, strengthens the empirical findings within the studied setting.
Critical weaknesses
  1. The framing exceeds the demonstrated scope. All experiments concern explicit acknowledgment of localized arithmetic errors, combining error recognition, information use, and reporting behavior. Although acknowledged as a limitation, the connection to general causal attribution and CoT monitoring remains untested. The paper should narrow its framing or validate the finding on a distinct reasoning task without an error-acknowledgment requirement.

  2. The incremental contribution and mechanism remain unclear. Prior work studies how intermediate sentences influence subsequent reasoning [1], and Li [2] already shows that CoT changes sensitivity to source-cache edits. The authors should clarify what fixed-mode prefix interventions establish beyond these findings. Regional KV replacement measures intervention sensitivity but does not explain how continuation changes information use. Targeted attention or computational-path interventions would strengthen this interpretation. The early-projection versus behavior comparison also needs a clearer connection to the central question.

[1] Paul C. Bogdan, Uzay Macar, Neel Nanda, and Arthur Conmy. Thought Anchors: Which LLM Reasoning Steps Matter? 2025. https://arxiv.org/abs/2506.19143

Questions for authors
  1. Which finding goes beyond the reasoning-dependent cache use demonstrated in [2]?
  2. What causal quantity does “causal relevance” denote? Which controls distinguish effects of error-related information from other changes introduced by replacing an entire cache region?
  3. What evidence connects regional sensitivity changes to CoT monitoring reliability, given that predictive monitoring need not rely on causal signals?

[2] Bojie Li. Models Take Notes at Prefill: KV Cache Can Be Editable and Composable. 2026. https://arxiv.org/abs/2606.17107

10 comments
AuthorsOpenReview ↗

Response to Reviewer 9cPt (1/10)

Thank you for taking the time to review our work. We appreciate your emphasis on the contribution and broader relevance of the findings. Your comments motivated the additional experiments below and helped us make the connection to monitoring more concrete.

W1. Generalization beyond arithmetic error acknowledgment (1/2)

Reviewer concern: Evidence limited to arithmetic error acknowledgment does not establish the broader causal-attribution and CoT-monitoring framing; the finding should be tested on a distinct reasoning task without requiring error acknowledgment.

Summary: We extended mode comparisons and fixed-mode prefix interventions to ShARC-derived rule reasoning and XSTest-derived safety responses, measuring expressed information limitations and refusal. Both tasks show regional-profile differences across modes and at least one significant prefix contrast in every tested model–mode profile (8/8 per task, BH q < .05), extending the finding beyond error acknowledgment.

Thank you for highlighting the need for evidence beyond arithmetic error acknowledgment. To test the scope of our findings, we extended the Sec. 5.2 mode comparisons and Sec. 6.2 fixed-mode prefix interventions to ShARC-derived rule reasoning [1] and XSTest-derived safety responses [2].

Settings. Each task has 200 token-aligned pairs. From 100 ShARC rules, we built two pairs per rule differing in one Yes/No answer token, yielding decidable versus information-insufficient cases; the outcome is requesting information or expressing an information limitation. XSTest requests were rewritten so a localized edit establishes safe/unsafe status; the outcome is expressed refusal or an intention to refuse. Paired inputs match in token length and all unedited tokens. PchangeP_{\mathrm{change}} denotes the edited span; Pafter changeP_{\text{after change}} denotes the post-change region, comprising all subsequent prompt tokens through the generation boundary.

We tested Gemma4 26B-A4B, Qwen3.6 27B, Qwen3 8B and GPT-OSS 20B in thinking/non-thinking modes (T/NT), giving eight profiles (GPT-OSS: native/empty analysis). We use the paper's GPT-OSS-120B judge with task-specific definitions and no perception filtering. Both donor directions were evaluated. Matched cohorts use unpatched = 0% and joint replacement of the edited span and post-change region = 100% per checkpoint/mode; Δ = NT − T (normalized pp, before rounding).

Prefix comparisons hold the checkpoint, prefill, and mode fixed. ShARC uses neutral/restatement/inference prefixes with three wordings each. We average across wordings and donor directions, with inference clustered by source rule. XSTest uses four 18-token prefixes (wording/evaluation score/skip review/first conclusion) under a shared evaluation context, with 152 held-out cases, safe donors into unsafe bases and suffix-only judging.

L is the donor-oriented regional-effect difference. Raw contrasts use sign-flip tests with BH correction across all pairwise prefix comparisons: 3 contrasts × 8 profiles = 24 for ShARC, and 6 × 8 = 48 for XSTest.

(Continued in Part 2/2 below.)

AuthorsOpenReview ↗

Response to Reviewer 9cPt (2/10)

W1. Generalization beyond arithmetic error acknowledgment (2/2)

Results. Both tasks reproduce the paper's central findings: regional attribution varies across modes and with prefix content within a fixed mode. With an information-insufficient ShARC base and sufficient donor, Gemma reverses from edited-span to downstream dominance. Qwen3.6 shifts similarly but retains stronger edited-span sensitivity; Qwen3 8B shows a smaller, opposite shift. Direction and magnitude vary by checkpoint.

CheckpointPchangeP_{\mathrm{change}}: TPchangeP_{\mathrm{change}}: NTΔPafter changeP_{\text{after change}}: TPafter changeP_{\text{after change}}: NTΔ
Gemma4 26B-A4B84.00.0−84.09.089.7+80.7
Qwen3.6 27B102.870.2−32.64.99.2+4.3
Qwen3 8B97.398.0+0.7−6.2−12.0−5.8
GPT-OSS 20B94.4Unstable†Unstable†5.6Unstable†Unstable†

†GPT-OSS NT's joint reference is +0.5 pp [95% CI: −3.5, +5.0], precluding stable normalized NT effects and mode differences.

For XSTest's safe bases with unsafe donors, all four checkpoints show lower normalized edited-span effects and higher downstream effects in non-thinking mode.

CheckpointPchangeP_{\mathrm{change}}: TPchangeP_{\mathrm{change}}: NTΔPafter changeP_{\text{after change}}: TPafter changeP_{\text{after change}}: NTΔ
Gemma4 26B-A4B61.20.0−61.235.961.5+25.7
Qwen3.6 27B67.13.8−63.210.419.2+8.8
Qwen3 8B56.746.5−10.246.566.3+19.9
GPT-OSS 20B88.110.1−78.051.992.4+40.6

Within each fixed mode, prefix content changes regional sensitivity in every profile (BH q < .05):

TaskSignificant raw-L contrastsProfiles with a significant contrast
ShARC18/248/8
XSTest31/488/8

Overall, these results support the broader task scope of our finding: regional attribution depends on decoding mode and continuation content across distinct tasks and behavioral targets.

AuthorsOpenReview ↗

Response to Reviewer 9cPt (3/10)

Q1 + W2.1. Distinction from Prior Work (1/2)

Reviewer concern: Given prior work on intermediate-sentence influence and reasoning-dependent cache use, it is unclear what the fixed-mode prefix interventions add.

Summary: We show that continuation content changes the behavioral effect of the same earlier cache replacement and shifts the relative effects of error-span and post-error replacement. We compare each region’s patched–unpatched effect across prefix conditions, testing changes in intervention sensitivity beyond changes in unpatched behavior. A controlled 2×2 manipulation of error restatement and a correct proposition changes this regional balance while the checkpoint, original prompt prefill, target mode, and endpoint remain fixed; natural-prefix replay provides complementary evidence. This establishes continuation dependence of relative regional intervention effects, extending the CoT-dependent source-edit sensitivity demonstrated in [3].

Our contribution concerns instability in regional causal attribution across generation modes, continuation content, and measured endpoints. For each prefix, we measure the behavioral difference between replacing a particular earlier cache region and leaving it unpatched. We then compare these differences across prefixes. The question is whether changing the prefix changes the behavioral effect of the same cache replacement and shifts the balance between error-span and post-error replacement effects.

Cache replacement precedes prefix processing. Within each prefix condition, identical text is processed separately against patched and unpatched caches, with its internal states recomputed in each case. The checkpoint, original prompt prefill, target mode, and behavioral endpoint—error acknowledgment—remain fixed.

Distinction from [4]. [4] measures sentence importance through changes in final-answer distributions after sentence resampling and in subsequent token distributions after attention suppression (§§2–3, 5). The manipulated sentence is itself the object whose influence is measured. In our experiments, the prefix instead provides the conditions under which separate interventions on earlier cached regions are evaluated. A sentence’s influence on subsequent outcomes does not establish that it changes the behavioral effect of intervening on an earlier cached state. Its reasoning/base comparison also concerns receiver-head attention concentration (Appendix K), whereas our mode comparison measures regional KV replacement effects within the same checkpoint.

Distinction from [3] and the central finding. Li [3] studies how cached conclusions support decisions and how caches can be edited and reused, including CoT-dependent source-edit recovery and decision recovery through appended errata (§4; Fig. 2). Our additional finding is that controlled prefix content shifts the balance between error-span and post-error replacement effects, established by comparing the two regions’ patched–unpatched effects across content conditions.

Our controlled prefixes manipulate two content factors—restatement of the erroneous relation and provision of a correct proposition—using short declarative sentences, without explicit correctness judgments or checking/repair instructions. These are controlled propositional manipulations rather than generated or replayed natural CoT. Under each content condition, we separately measure error-span and post-error replacement effects. Their relative effects change across content conditions, with interactions between the two factors in some profiles (§6.2; Fig. 7). This establishes continuation-dependent changes in regional attribution even when the target mode remains fixed.

(Continued in Part 2/2 below.)

Reviewer_LBd7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors patch regional key-value cache states between matched clean and erroneous GSM8K traces across eight checkpoints and both reasoning modes then vary the continuation with replayed thinking and controlled prefixes. Which cached region drives error acknowledgment turns out to depend on what is generated afterward.

Critical strengths
  • Solid experimental setup across eight checkpoints with mode-shared cohorts, cross-fitted readouts and cross-judge robustness
  • I trust the numbers (am questioning what they teach)
Critical weaknesses
  • Section 6.2 is motivated by "qualitative observations of restatement and independent recomputation" (namely the mechanism) yet there doesn't appear to be an experiment that conditions on it
  • The crossover (Section 5.3 and Figure 5b) pits a one-token readout at the first generated token against behavior over 1024 tokens. A first-token readout is bound to lean on the post-error key-value states (where the direction was fit) so I'd predict it. Can the authors read the projection at later tokens (this would allow for a claim about when the cache is consulted)
Questions for authors
  • Within a fixed checkpoint and mode does the profile differ between continuations the targeted-control judge labels targeted_check or independent_resolve versus continue_or_answer?
4 comments
AuthorsOpenReview ↗

Response to Reviewer LBd7 (1/4)

Thank you for the time you devoted to reviewing our paper. We appreciate your suggestion to examine the projection beyond the first generated token. This prompted a more direct test of whether the representation–behavior dissociation persists at later generated positions.

W1. Connection between the motivating observations and the Section 6.2 experiment

Reviewer concern: Section 6.2 invokes restatement and independent recomputation as motivation, but there doesn't appear to be an experiment that conditions on it.

Summary: Section 6.2 uses the motivating observations to test whether continuation content changes regional cache sensitivity. Its 2×2 design independently manipulates restatement of the erroneous proposition and provision of a correct proposition—the information that recomputation could introduce. Annotations of original continuations ground this choice of content factors; the experiment tests their content effects rather than the recomputation process itself.

Section 6.2 operationalizes the informational content of the motivating observations through two independently manipulated factors: restatement of the erroneous proposition (error anchor) and provision of a correct proposition. The correct-proposition factor manipulates the information that recomputation could introduce, rather than the model's recomputation process itself. These observations motivate a hypothesis about content-dependent regional sensitivity; they are not offered as a mechanistic explanation of the earlier profiles.

An additional annotation of 7,112 original unpatched continuations found error-proposition restatement in 57.1% and explicit correct-target derivation in 64.9% of the first-256-token windows. These frequencies provide empirical grounding for the choice of content factors.

The 2×2 design varies the presence of each factor while holding the checkpoint, original prompt prefill, and reasoning mode fixed. We compare the effects of error-span and post-error cache replacement across these content conditions. These controlled contrasts directly test whether changing continuation content changes the relative influence of earlier cache regions on subsequent error acknowledgment.

AuthorsOpenReview ↗

Response to Reviewer LBd7 (2/4)

W2. Projection readouts at later generated tokens

Reviewer concern: The first-token projection may preferentially reflect nearby post-error states, making its crossover with behavior predictable; later-token measurements could clarify whether the pattern persists.

Summary: Appendix Fig. 21 tests dependence on the fitting position. We extended projections through g128 across all seven fitting positions and 16 checkpoint–mode profiles. With the original boundary-fitted direction, crossover recurs in 14/15 eligible profiles at g4 and 5/6 at g128, supporting a dissociation beyond g1 within the eligible profiles.

We understand the concern that fitting the direction near the end of the prompt and measuring it at g1 could favor post-error states. The original g1 comparison fixes the first token across patching conditions to measure the initial response before continuations diverge.

Appendix Fig. 21 already addresses dependence on the fitting location. It varies where the direction is fitted while evaluating the directions on the same downstream g1 states. The origins include the error-span end, the end of the supplied reasoning before the filler, and an independently constructed no-filler boundary. The crossover also occurs with these alternative origins. Thus, the separation does not require a direction fitted at the final boundary after the post-error context, weakening an explanation based solely on that fitting choice.

To directly test whether the dissociation is confined to measurement at the first token, we extended the analysis to g1, g2, g4, g8, g16, g32, g64, and g128 across all eight checkpoints in both reasoning modes. Here, g_n denotes the post-block residual state after processing the nth generated continuation token. We retained all seven original cross-fitted directions, standardization, patching conditions, normalization, and behavioral labels, and replayed each condition's exact original continuation tokens. The g1 point estimates and sample counts reproduce the original analysis.

Table 1. Crossover profiles / eligible profiles across fitting positions and measurement times. Sixteen checkpoint–mode profiles are assessed in each cell.

Fitting positiong1g2g4g8g16g32g64g128
Boundary (pbndp_{\mathrm{bnd}}; original primary)16/1613/1614/1510/137/96/84/55/6
Error-span end (perrp_{\mathrm{err}})6/104/84/63/52/43/32/35/5
Next token after error position (perrNp_{\mathrm{errN}})10/1111/139/106/104/54/46/78/11
Reasoning end (pTp_T)11/1111/1311/128/125/61/27/76/8
Filler midpoint (pfmidp_{\mathrm{fmid}})16/1613/1512/1511/1413/158/1112/1313/14
Filler end (pfTp_{\mathrm{fT}})14/1413/1413/149/118/104/46/74/5
Boundary without filler (pbndno fillerp_{\mathrm{bnd}}^{\mathrm{no\ filler}})16/1613/1614/147/99/113/54/46/8

Counts describe crossover directions in point estimates, not statistical significance. Eligibility uses the same readout-validation and reference-effect criteria as the paper’s Sec. 5.3 analysis, applied at each measurement position to both regions and donor directions; ineligible profiles are excluded from the denominator. Later comparisons use sources reaching the measured token. perrNp_{\mathrm{errN}} is capped at the reasoning end.

With the original boundary-fitted direction, 6 of 16 profiles are eligible at g128, and 5 of these 6 show crossover. Of the ten ineligible profiles, nine have an uncertain reference projection effect and one fails readout validation.

The later-token dissociation is also supported by the magnitudes of the normalized projection–behavior gaps, reported in Table 2 below. Each endpoint is normalized by its own reference effect, so these gaps quantify how far the regional transfer profiles differ, complementing Table 1's crossover counts.

For example, at g128 the boundary-fitted direction in Gemma-4 12B thinking has gaps of +40.2 / +39.8 pp. Error-span replacement therefore transfers more of the reference's behavioral effect than its projection effect, whereas post-error replacement shows the opposite relationship.

Together, Fig. 21 and the new analysis test fitting position and measurement time separately. They support Section 5.3's central point: a region's influence on an error-related readout need not track its influence on subsequent behavior, and this dissociation is not confined to the first token. Later-state effects can include propagation through preceding generated states, so these measurements characterize the projection response over time rather than pinpoint direct cache access.

AuthorsOpenReview ↗

Response to Reviewer LBd7 (3/4)

W2. Additional results: magnitude at g128

Expression transfer (E) and projection transfer (R) are normalized separately:

  • E = 100 × regional expression effect / reference expression effect.
  • R = 100 × regional projection effect / reference projection effect.

The reference intervention is the same, but the denominator is specific to each endpoint. We normalize within each donor direction, compute error gap = E − R for error-span replacement and post-error gap = R − E for post-error replacement, then average each gap equally over the two directions. Both gaps being positive defines the crossover; their magnitudes measure the separation between the normalized regional transfer profiles.

“Boundary” and “Filler midpoint” denote readout directions fitted at the prompt boundary and the midpoint of the appended filler, respectively; both are evaluated at g128.

Table 2. Error gap / post-error gap at g128 (pp).

CheckpointModeN rangeBoundary: error / post-error gapsFiller midpoint: error / post-error gaps
Gemma-4 12BT267–275+40.2 / +39.8+25.5 / +25.0
Gemma-4 12BNT164–167—+5.4 / +7.9
Gemma-4 26B-A4BT366–384+6.5 / +6.3+29.4 / +29.3
Gemma-4 26B-A4BNT377–398—+13.8 / +17.7
Gemma-4 31BT312–320—+29.1 / +28.7
Gemma-4 31BNT179–289—+31.2 / +21.0
Qwen3.5 27BT463+83.2 / +83.4+69.3 / +45.1
Qwen3.5 27BNT513–527—+32.2 / -0.3
Qwen3.5 122B-A10BT514–515—+76.9 / +49.0
Qwen3.5 122B-A10BNT443–477—+27.3 / +6.1
Qwen3.6 27BT471+23.0 / +6.8+41.4 / +28.6
Qwen3.6 27BNT430–486+16.0 / +15.9+23.1 / +19.5
Qwen3 8BT416–417——
Qwen3 8BNT71–84——
GPT-OSS 20BT726–729—+9.9 / +9.3
GPT-OSS 20BNT103–107+13.2 / -11.2+26.4 / +15.6

T/NT: thinking/non-thinking. N is the source-count range across regions and donor directions at g128. — denotes exclusion from the normalized comparison under the original readout-validation or reference-effect criteria. These are differences between normalized transfer percentages, not raw expression-probability changes.

In a separate source-composition check, we restricted sources to those reaching g128 under all core conditions and held this cohort fixed across measurement times. With the direction fitted at the filler midpoint, crossover at g128 was observed in 14/14 eligible profiles, compared with 13/14 in Table 1’s position-specific source cohorts.

AuthorsOpenReview ↗

Response to Reviewer LBd7 (4/4)

Q1. Regional intervention profiles for different behavioral outcomes

Reviewer question: Within a fixed checkpoint and mode, do regional intervention profiles differ for targeted checking, independent re-solving, and continuing/answering?

Summary: Treating the three labels as separate, potentially co-occurring outcomes, we compared regional effects on matched source cohorts. Raw effects differ in direction and magnitude, while normalized profiles can remain similar: in Qwen3 8B thinking, error-span replacement reduces targeted checking and increases independent re-solving, yet their normalized regional profiles closely align.

Because cache replacement can change behavioral labels, we compare regional effects on each behavior’s occurrence probability while holding the checkpoint, mode, and source cohort fixed, rather than grouping continuations by post-intervention labels. The labels can co-occur: each continuation is scored for targeted checking, independent re-solving, and continuing/answering as separate binary outcomes.

We evaluated all eight checkpoints, both reasoning modes, and both donor directions, retaining the original judge labels and identical source cohorts across outcomes and conditions within each setting. The raw intervention effects differ across behaviors in direction and magnitude, while their normalized regional profiles can be similar.

Under clean-donor error-span replacement, mean checking rates decrease in all 16 checkpoint–mode profiles, independent re-solving increases in 15 (one unchanged), and continuing/answering decreases in 13 but increases in three. Table 1 illustrates these behavioral responses for Qwen3 8B thinking on 394 sources.

Table 1. Raw behavior occurrence rates (%), Qwen3 8B thinking, clean donor.

OutcomeUnpatchedError spanPost-errorBoundaryJoint
Targeted check98.018.588.697.714.7
Independent resolve2.377.210.92.081.2
Continue/answer73.634.873.671.832.2

To compare relative regional profiles, we normalize each outcome's regional change by its own joint error-to-boundary effect:

Normalized effect = 100 × (patched rate − unpatched rate) / (joint-patched rate − unpatched rate).

This sets the unpatched reference to 0% and joint replacement to 100%, aligning each profile with the direction of its outcome's joint effect.

Table 2. Reference-normalized regional effects (% [95% CI]) on the same sources.

OutcomeError spanPost-errorBoundaryJoint
Targeted check95.4 [90.2, 100.6]11.3 [7.6, 15.2]0.3 [−1.5, 2.1]100.0
Independent resolve94.9 [89.3, 100.6]10.9 [7.0, 15.1]−0.3 [−2.3, 1.6]100.0
Continue/answer93.9 [80.8, 108.1]0.0 [−15.4, 12.8]4.3 [−8.7, 16.3]100.0

Pointwise 95% intervals use 10,000 paired source-bootstrap draws, recomputing numerator and denominator together. Joint replacement is fixed at 100% by definition.

Tables 1–2 show that, within this fixed checkpoint and mode, checking decreases while independent re-solving increases, yet their normalized regional profiles align closely. Continuing/answering also has its largest localized effect at the error span.

As an additional robustness check on the paper's mode-dependent finding, we hold each behavioral outcome fixed and compare reasoning modes on shared sources. Table 3 shows Gemma-4 31B under clean-donor replacement, with the same 211 sources in both modes.

Table 3. Reference-normalized regional effects by reasoning mode (%).

OutcomeNT: error spanNT: post-errorT: error spanT: post-errorError-span difference, T − NT (pp) [95% CI]
Targeted check5.142.350.016.4+44.9 [+32.6, +58.2]
Independent resolve2.639.543.117.6+40.5 [+23.0, +63.0]

T/NT: thinking/non-thinking. Each mode uses its own joint reference; included outcomes have reference-effect intervals excluding zero in both modes. Differences are pp of normalized effect. Intervals are pointwise 95% paired source-bootstrap intervals (10,000 draws).

In this additional comparison, both targeted checking and independent re-solving show greater normalized error-span sensitivity under thinking, shifting the larger regional response from post-error to error-span replacement.

Reviewer_Zf9fOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how regional K/V cache replacement affects models’ error acknowledgement and correction. It examines three main regions: the error span, the span after the error, and the position immediately before generation. Experiments across multiple models and reasoning modes show that thinking generally increases relative sensitivity to error-span replacement. Regional effects on an early error-related signal also differ from those on later acknowledgement. Furthermore, changing continuation prefixes can reshape regional sensitivity while keeping the model, original prompt prefill, and decoding mode fixed.

Critical strengths
  1. The matched cache-replacement and controlled prefix experiments provide good support for the paper’s main claims within the tested setting.

  2. The differences between thinking and non-thinking modes are insightful, particularly the finding that thinking generally increases relative sensitivity to error-span replacement.

Critical weaknesses
  1. The main text could better explain why the error-related direction fitted at the prompt boundary provides a meaningful readout of the hidden state after processing the first generated token, and how the "error-related projection" should be interpreted.

  2. The experiments show that continuation content can change regional sensitivity, but the specific trends still vary across models and settings. For instance, in Fig. 6, with a 64-token prefix, post-error sensitivity increases for Gemma-4 12B but decreases for Qwen3 8B and GPT-OSS 20B. Also, in Fig. 7, the interaction between erroneous and correct propositions is negative in thinking mode but positive in non-thinking mode for both Gemma-4 31B and Qwen3.5 27B. The paper does not fully discover what determines these differences. This makes it difficult to predict the direction and magnitude of sensitivity changes and apply the findings to other models, tasks, or experimental settings.

Questions for authors
  1. In Fig 4, thinking mode often shows larger absolute reductions in error expression after error-span replacement. Does this indicate greater sensitivity to cache manipulation? The normalised comparison in Fig 5a focuse on relative regional effects and does not retain their absolute scale. How should these two comparisons be interpreted together when assessing differences between thinking and non-thinking modes?

  2. Beyond the cross-judge comparison in Appendix K, were the LLM judgments of error expression and targeted control validated against human annotations? Reporting agreement with human judgments would strengthen the confidence in the evaluation.

7 comments
AuthorsOpenReview ↗

Response to Reviewer Zf9f (1/5)

Thank you for your time and careful review. We especially appreciate your questions about the interpretation of the error-related readout and the reliability of our behavioral judgments. Your feedback helped us clarify these measurements and add direct validation against human annotations.

W1. Meaning and validation of the downstream error-related projection

Reviewer concern: Why is a boundary-fitted direction meaningful at g1, and what does its projection represent?

Summary: The projection is a standardized clean–error score. We fit the direction at the boundary after the full trace is processed and validate its transfer to g1 on held-out sources. Measuring g1 captures the initial response to cache replacement before continuations diverge; the representation–behavior dissociation also recurs with directions fitted at other prompt positions (Appendix Fig. 21).

For each model and layer, we define an error-related direction from the mean hidden-state difference between error-present and error-free traces. We use the prompt boundary because the entire supplied trace has been processed before continuation begins. The projection is a standardized score along this direction: higher scores place the state further toward the error-present side of the contrast.

Cache replacement occurs after prefill and leaves the already-computed prompt hidden states unchanged. We therefore measure g1, the hidden state after processing the first continuation token with the modified cache, while predicting token 2. Fixing that token across patch conditions lets us compare the initial internal response before the generated text diverges.

We explicitly validate that the boundary-fitted readout generalizes to unpatched g1 states on held-out sources (Appendix C). We use five-fold source-grouped cross-fitting, excluding both twins of each evaluated source from direction fitting and score standardization. Qualification requires statistically supported error–clean separation in both the mean score gap and within-pair ordering.

Appendix Fig. 21 provides a complementary robustness check. We repeat the analysis at the same g1 state with directions fitted at other prompt positions, applying the same held-out g1 validation separately to each readout. The representation–behavior dissociation recurs across multiple validated readouts, showing that the finding does not depend on the boundary-fitted direction alone.

Overall, the validated readout lets us measure how cache replacement shifts the initial downstream state along the clean–error contrast. Section 5.3 uses this measure to compare each region's effect on the early error-related readout with its effect on subsequent error acknowledgment, under the same cache interventions and examples.

We have revised Sec. 5.3 to clarify the direction’s definition, the projection’s interpretation, and the rationale for fitting at the prompt boundary.

AuthorsOpenReview ↗

Response to Reviewer Zf9f (2/5)

W2. Heterogeneous effects and predictive generalization (1/2)

Reviewer concern: Unexplained model- and mode-specific trends in Figs. 6 and 7 limit prediction of sensitivity changes and generalization to other settings.

Summary: Additional analyses identify partial predictive associations. Prefix features reduced prediction error for Fig. 6 by 13.7% when target checkpoints were held out. For Fig. 7, response features had limited predictive value, and mode-dependent sign reversals persisted after matching problems and styles. These analyses aid interpretation without establishing a general mechanism for the heterogeneity.

We agree that the current results do not establish a general rule for predicting the direction or magnitude of sensitivity changes. Our focus is on demonstrating continuation dependence while holding the model, mode, and original input fixed. We agree that understanding the heterogeneity would aid interpretation and generalization, and therefore analyzed existing data.

The prediction analyses below use linear regression: an ordinary least-squares baseline is compared with extensions that apply ridge (L2) regularization to added features. Added features are standardized using training data, and the ridge penalty is selected by inner grouped cross-validation.

W2.1. Fig. 6: Variation in natural-prefix effects across models

For Fig. 6, we predicted how inserting a natural prefix (32, 64 or 96 tokens) changed post-error suppression of error acknowledgment. The baseline used prefix length, mode, and measured pre-insertion regional effects. We then added prefix features describing discourse, arithmetic, wording overlap, expression ordering and style. With both modes of each target checkpoint held out, these features reduced mean squared error (MSE) by 13.7%, with improvements in 5 of 6 models.

The table gives changes in post-error suppression at 64 tokens (after minus before prefix insertion, pp). Both predictors use the target model's measured pre-insertion effects.

ModelProblemsObserved [95% CI]Baseline+ prefix features
Gemma-4 12B54+3.7 [−9.3, +16.7]−8.3+2.8
Gemma-4 31B134+4.5 [−4.5, +14.2]−24.5−7.6
Qwen3.6 27B2710.0 [−1.8, +1.8]+11.9+11.0
Qwen3 8B51−9.8 [−27.5, +5.9]−19.9−5.1
Qwen3.5 27B128+1.6 [−1.6, +4.7]+3.6−3.6
GPT-OSS 20B106−26.4 [−38.7, −15.1]−6.6−10.8

The 13.7% MSE reduction covers all three prefix lengths; this table gives model-level means at 64 tokens, with CIs from the original problem-level resampling.

Predictions matched the reviewer's three point-estimate signs, but the observed CIs for Gemma-4 12B and Qwen3 8B include zero, and GPT-OSS's decrease was underestimated. Predictions for other models sometimes had incorrect signs. The gain persisted after shuffling features across problems within each model/length, suggesting prediction based on average reasoning styles rather than an explanation of individual sentence effects.

(Continued in Part 2/2 below.)

AuthorsOpenReview ↗

Response to Reviewer Zf9f (3/5)

W2. Heterogeneous effects and predictive generalization (2/2)

W2.2. Fig. 7: Variation in content-factor interactions across modes

Among unpatched responses that mentioned the erroneous computation and presented the correct one within their first 256 tokens, we compared error-evaluative and other openings. The table reports error-span minus post-error suppression of error acknowledgment (pp) in two Qwen non-thinking models.

ModelOther openingsError-evaluative openings
Qwen3.5 27B61.1 (n=54)1.4 (n=295)
Qwen3.5 122B-A10B60.6 (n=99)3.6 (n=195)

The association persisted across problem subsets and after accounting for model, problem, and response length; it remains observational, with adequate samples limited to these two models.

Separately, we predicted Fig. 7’s error-anchor × correct-proposition interaction in regional contrast L. The baseline controlled for response length, reference availability, prefix-style coverage, and model/mode. We tested adding either discourse, arithmetic and style features of unpatched responses together with problem/error properties, or the proportion of short continuations under each prefix condition (at most 20 words and numbers in total). Response features came from native responses and continuations generated after each prefix.

Entries are percentage reductions in MSE relative to the corresponding baseline; negative values mean worse prediction.

Added informationNew problems, same modelsNew modelsBoth new
Response and problem/error features3.8%−6.5%−6.9%
Short-response proportion2.0%2.6%1.2%†

† The 95% CI for error reduction includes zero; CIs use problem-level resampling.

Predictive gains depended on the validation condition and short-response threshold. Matching problems and wording styles did not remove the sign reversals: thinking/non-thinking effects remained −17.5/+20.1 pp for Gemma-4 31B and −15.8/+22.9 pp for Qwen3.5 27B. These analyses identify some predictive associations, but do not yet provide a general mechanism for the direction and magnitude of changes across models.

Comments on the paper

2 comments

We have uploaded a revised manuscript incorporating the updates discussed in our responses. Revised and added text is highlighted in blue. Thank you for your time and consideration.

AuthorsOpenReview ↗

Summary of Revisions and Responses

We thank all three reviewers for their time and careful consideration of our work. We appreciate their recognition of the strengths of the controlled experimental design and their suggestions for clarifying the interpretation, scope, and implications of our findings. Below, we summarize the manuscript revisions and additional analyses provided in response to their comments.

Extensions beyond arithmetic error acknowledgment. To address Reviewer 9cPt’s concern about whether our findings generalize beyond arithmetic error acknowledgment, we extended the mode comparisons and fixed-mode prefix interventions to rule-based information seeking and safety-related refusal, using ShARC- and XSTest-derived tasks. Both extensions reproduce the central findings beyond error acknowledgment: regional attribution varies across modes, and continuation content can alter regional effects within a fixed mode. Each task shows significant prefix-dependent changes in every tested model–mode profile. These results are included in Section 7 and Appendix M.

Measurement interpretation and validation. In response to the measurement questions raised by Reviewers Zf9f and LBd7, Section 5.3 now explains the interpretation and held-out validation of the error-related projection, and distinguishes effects on this internal readout from effects on subsequent behavior. Additional later-token measurements show that the readout–behavior dissociation can persist beyond the first generated token (Appendix H). We also added validation of the behavioral judgments against human annotations (Appendix K).

Additional controls and analyses. Our individual responses provide additional evidence concerning the interpretation of cache replacement. Calculation-preserving donor replacements produce much smaller behavioral effects than error-status-changing replacements despite substantial cache changes. Timing and attention-path experiments further constrain possible explanations for the prefix-dependent effects, although a specific mediating path remains unresolved. We also explain how Section 6.2’s content manipulations connect to the motivating observations and compare effects on different behavioral outcomes within fixed checkpoints and modes. These additional analyses are reported in the responses.

Contribution and monitoring relevance. Finally, we clarified the contribution beyond prior work and its relevance to monitoring. When intervention evidence is used to interpret why a behavior occurred or to select a corrective intervention, its validity should be checked under the conditions in which it will be applied. Our findings show that regional causal attribution need not remain stable across reasoning modes or changes in continuation content.

The revised manuscript incorporates the task extensions, measurement clarifications, and monitoring discussion, with changes highlighted in blue. We hope that these revisions and the accompanying responses help clarify the evidence and the scope of our conclusions. We thank you again for your time and consideration and would welcome further discussion during the remainder of the rebuttal period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.