acceptodds
Under review as a conference paper at ICLR 2027

A Wrong Turn in Addition: Generation, Routing, and Correction of Arithmetic Errors in LLMs

Abstract

Despite rapid progress in mathematical reasoning, large language models remain surprisingly brittle on elementary arithmetic. We ask how a local arithmetic error becomes a persistent autoregressive trajectory, and whether evidence of that error survives after the model has committed to it. Using matched digit-preserved carry counterfactuals and causal interventions, we decompose multi-operand addition errors into generation, read-back routing, and downstream compatibility. Natural off-by-one carry errors are characterized by weakened carry-sensitive influence, while the emitted digit subsequently selects between local lower- and higher-carry continuation routes through a localized query–key interaction. Crucially, error propagation and error representation coexist: the model can continue along a wrong route while downstream states remain sensitive to whether the read-back digit is arithmetically compatible with its context. This delayed compatibility signal is visible in next-token behavior, entropy, and representation geometry, and enables a training-free mechanism-guided correction method that substantially improves exact-answer accuracy and outperforms most training-required baselines.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability, Large Language Models, Multi-Operand Addition, Error Propagation, Inference-Time Correction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_uH1rOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how local arithmetic errors arise, propagate, and can be corrected in multi-operand addition by LLMs. Using carry-preserving counterfactuals and causal interventions, the authors identify three stages of an error. First, weakened carry-sensitive computation can lead the model to generate a wrong digit. Second, once that digit is fed back into the model, Query-Key interactions in specific attention heads route the subsequent computation along an incorrect carry path. Third, even after following the wrong path, the model's downstream hidden states retain signals of arithmetic inconsistency.

Based on these findings, the authors propose Route+Entropy, a training-free correction method that uses internal carry routing to select candidate digits and late-layer predictive entropy to decide whether to replace the original output. On Qwen3-4B, the method improves exact-answer accuracy from 60.9% to 81.7%. The authors also examine the mechanisms and correction performance on other Qwen and Gemma models.

Critical strengths

1. Strong causal analysis of the error mechanism. The paper goes beyond observing correlations in hidden states and predictive entropy. The authors construct controlled counterfactual examples and use activation patching, attention-head analysis, and Query-Key edge interventions to identify components involved in error propagation. In particular, intervening on just two QK edges nearly reproduces the effect of replacing the full attention patterns. This provides convincing causal evidence for the proposed carry-routing mechanism.

2. A useful delayed error signal. An interesting finding is that the model retains information about whether a generated digit is correct, even after it has committed to an incorrect continuation. This helps explain why errors can propagate despite the presence of internal signals that distinguish correct and incorrect computation paths. The finding is also practically useful: the proposed training-free correction method improves exact-answer accuracy from 60.9% to 81.7% on the main model.

Critical weaknesses

1. Limited generality of the mechanism. The analysis focuses mainly on off-by-one carry errors in multi-operand addition, with the most detailed causal experiments conducted on specific attention heads in Qwen3-4B. Although the authors include experiments on other models and identify analogous mechanisms across architectures, it remains unclear whether the findings extend to reasoning errors beyond arithmetic. There is also a potential confounding factor in the construction of Case C. The authors require the model to naturally generate the correct digit on the modified problem. This selection criterion may introduce differences in the model's original prediction confidence, making it harder to determine whether the observed compatibility signals are independent of confidence differences.

2. Limited additional benefit and unclear computational efficiency of the correction method. Although Route+Entropy substantially improves over the original model, its advantage over Entropy-only is less consistent. The gain in exact-answer accuracy is 3.8 percentage points on Qwen3-4B and 7.9 points on Gemma-3-4B, but only 0.4 points on Qwen3-30B and 0.1 points on Qwen3.5, with no improvement on Qwen3.8. These results raise questions about how much the routing mechanism contributes to correction performance across different models. The method also requires additional forward computation for multiple candidate digits. Without comparisons under matched inference budgets, it is difficult to assess its efficiency relative to other inference-time correction methods.

Questions for authors

1. Case C requires the model to naturally generate the correct digit on the modified problem. Could the authors control for prediction confidence or logit margins at the target position to check whether the downstream entropy differences are partly explained by this selection criterion?

2. Route+Entropy provides only small improvements over Entropy-only on several models. Could the authors explain why the benefit of routing varies across models? A comparison under matched inference time or compute budgets would also help clarify the practical contribution of the routing mechanism.

5 comments

Thank you for your time and constructive feedback! We have carefully addressed your comments below, and are incorporating some of these new results into our manuscript, which we look forward to sharing soon.

All newly added figures are prefixed with "R" (e.g., Fig. R1) for distinction, which can be viewed in the Supplementary Material.

Q1. Generalization Beyond Arithmetic Reasoning

Our study focuses on arithmetic because its intermediate carry states are precisely defined and causally testable. Nevertheless, we further embed the original addition task within a GSM8K-style reasoning context (Fig. R1(f)). The same L34H24/H27 lower-/higher-carry routing pattern largely persists, suggesting that the identified mechanism is not restricted to isolated arithmetic prompts. While this supports robustness within a broader reasoning context, extending the mechanism to non-arithmetic computations remains an open question.

Q2. Target-Position Confidence Confound

We agree that Case C\mathcal{C} is typically more confident at the target-position than A\mathcal{A}. We therefore test whether the downstream separation in Sec. 5 persists under target-confidence matching.

Fig. R5 performs two single-criterion sensitivity analyses on the original A/C\mathcal{A}/\mathcal{C} pairs, progressively relaxing the matching tolerance for either the target top-two logit gap (a) or target-position entropy (b), until all 800 pairs are included.

Across both sweeps, the paired A−C\mathcal{A}-\mathcal{C} differences in L34 next-position entropy and next-digit GT margin remain stable and do not reverse. Thus, while C\mathcal{C} is indeed more confident at the target position on average, the downstream separation in Fig. 6 is robust under either confidence-matching criterion and cannot be explained solely by target-position confidence.

This complements the causal QK recombination in Sec. 5 (Table 3), where changing context-side keys strongly alters the next-position margin. Together, these results support a context-dependent arithmetic-compatibility effect beyond target-position confidence differences.

Q3. Model-Dependent Correction Gains

Entropy-only already shares the mechanism-guided entropy readout layers (Sec. 6 and Appendix E.4), and routing/recurrent proposals additionally restrict the candidate set. We analyze this contribution on matched decoding states from five 1,000-question test cohorts.

Table R8. Proposal quality and incremental gains over Entropy-only.

ModelΔ\Delta Ques AccPGTP_{\mathrm{GT}}EGTE_{\mathrm{GT}}ΔR\Delta RΔP\Delta P
Qwen3-4B+3.8%95.74%2.62%+23+25
Qwen3-30B-A3B+0.4%80.41%0.76%−2+3
Gemma-3-4B+7.9%94.77%0.87%+63+20
Qwen3.5-4B+0.1%88.73%11.27%0−1
Qwen3.8-27B0.0%90.18%9.82%−12+11

Δ\Delta Ques Acc is the end-to-end question-accuracy gain from Table 11. On aligned natural off-by-one first errors, PGTP_{\mathrm{GT}} is the fraction proposing the GT neighbor (unresolved counted as misses), while EGTE_{\mathrm{GT}} is the fraction where candidate restriction excludes GT. At matched decoding states, ΔR\Delta R counts additional first-error repairs (Route minus Entropy-only), and ΔP\Delta P counts harmful edits avoided on originally correct digits (Entropy-only minus Route).

The key factor is the reliability and coverage of candidate restriction. Reliable routing can help entropy-based correction both recover errors and preserve correct digits. However, incorrect restrictions may exclude GT, while unresolved routing limits how often the mechanism can contribute.

Gemma and Qwen3-4B improve error recovery and correct-digit preservation simultaneously (Table R8). Qwen3-30B is mainly limited by unresolved routing, whereas Qwen3.5 shows nearly offsetting gains and losses, and Qwen3.8 trades fewer repairs for better preservation.

Architecturally, Qwen3-30B's MoE pathways may make fixed-head routing preferences less consistently resolved, suggesting a coverage bottleneck. In contrast, Gated DeltaNet derives directional proposals indirectly from recurrent innovation energy rather than explicit QK preferences; our implementation always issues a direction, so unreliable proposals may exclude GT. These architectural explanations remain hypotheses, but suggest that routing-signal reliability and coverage govern its incremental value over entropy selection.

Q4. Computational Overhead

As stated in Appendix E, our correction experiments serve as functional validation of the identified mechanism, rather than a practical arithmetic-solving method. For real-world usage, using an external calculator is clearly preferable. We nevertheless report latency and FLOPs in the following table for completeness, without claiming computational efficiency or deployment advantages.

Table R9. Inference-time accuracy and computational cost.

MethodLatency (ms)GFLOPsQues Acc (%)
Original450.08747.9460.9
Re-prompting490.321436.8261.4
Steering473.57756.5063.5
Target-state465.00748.0066.4
Dual-stream483.74748.0159.3
Next-state599.11759.7584.4
Entropy1395.11935.7877.9
Route+Entropy1390.90935.7881.7

Dear Reviewer,

Thank you again for your constructive feedback. We have updated our manuscript to incorporate additional analyses including L24 carry-sensitive heads, causal connection between generation and routing, cross-operation generalization and so on, as discussed in our rebuttal. We would greatly appreciate it if you could take another look.

If you feel that these revisions have adequately addressed your concerns, we would be very grateful if you would consider updating your assessment and score. Should you have any further questions or concerns, we are happy to address them and remain available for further discussion.

Thank you once again for your time and thoughtful consideration!

Reviewer_HNvDOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper investigates how errors arise and propagate during multi-digit addition in language models, specifically focusing on carry-related mistakes. The authors conduct causal intervention experiments to identify internal components involved in generating incorrect digits and show how previously generated digits influence subsequent predictions through carry-dependent mechanisms. Building on these findings, they propose a training-free correction method that combines mechanistic signals with predictive entropy to improve arithmetic accuracy.

Critical strengths
  1. The paper studies a well-scoped question with clear motivation - how arithmetic errors arise and propagate through autoregressive generation. The distinction between error generation, propagation, and downstream compatibility provides a useful perspective for studying arithmetic failures in language models.
  2. The experiments are well designed and sufficiently support the claims made in the paper.
  3. The paper also demonstrates a practical application of its findings through a training-free correction method that substantially improves arithmetic accuracy. I appreciate the effort to connect mechanistic interpretability findings with improvements in model behavior.
Critical weaknesses
  • The paper provides a useful high-level explanation of how arithmetic errors arise and propagate. However, I had difficulty understanding how the model actually performs addition and how the identified components contribute to this computation. For example, Section 3 identifies L24 attention as important for target-digit selection, while Section 4 characterizes L34H24/H27 as routing subsequent computation toward lower- or higher-carry continuations. Could the authors clarify how these findings fit together in the broader arithmetic computation? In particular, would inducing or correcting an error through L24 also change the routing behavior of L34H24/H27?
  • Section 3 identifies L24 attention, along with some late-layer MLPs, as causally influential in target-digit selection. However, the investigation stops at the attention-block and MLP-block level. Could the authors further investigate which attention heads or pathways contribute to the carry-sensitive effect at L24, and provide some insights into what specific computations they perform?
  • I'm unclear whether the characterization of L34H24 and L34H27 as lower- and higher-carry routing heads fully captures their computational roles. In particular, is this a general mechanism used during addition, or a pattern specific to the off-by-one errors studied in the paper? For example, would manipulating these heads during correct generation redirect the model toward a different carry-consistent continuation?
  • Minor issue: Related work could provide a broader discussion of prior mechanistic studies of arithmetic in language models. Although following papers primarily investigate how models perform arithmetic rather than how errors arise and propagate, they provide useful context for positioning the paper's mechanistic findings.
    • Ameisen, et al., "Circuit Tracing: Revealing Computational Graphs in Language Models", Transformer Circuits, 2025.
    • Stolfo, Alessandro, Yonatan Belinkov, and Mrinmaya Sachan. "A mechanistic interpretation of arithmetic reasoning in language models using causal mediation analysis." Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.
    • Mamidanna, Siddarth, et al. "All for one: Llms solve mental math at the last token with information transferred from other tokens." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.
Questions for authors
  • My understanding is that Section 3 identifies L24 attention as important for generating the target digit, while Section 4 identifies L34H24/H27 as important for routing the generated digit at the next position. However, I am unclear how these two mechanisms are connected. I'm wondering if inducing or correcting an error through L24 also changes the routing behavior of L34H24/H27? This may also help establish a causal connection between error generation and propagation?
  • Figure 5 supports the interpretation of L34H24 and L34H27 as lower- and higher-carry routing heads, but it is unclear how consistently this pattern holds across examples. Could the authors report the proportion of examples that follow the expected routing pattern and it would great if also characterize the cases where it does not hold?
4 comments

Thank you for your time and constructive feedback! We have carefully addressed your comments below, and are incorporating some of these new results into our manuscript, which we look forward to sharing soon.

All newly added figures are prefixed with "R" (e.g., Fig. R1) for distinction, which can be viewed in the Supplementary Material.

Q1. L24 Head Localization and Information Flow

We localize carry-sensitive generation within L24 attention (Fig. R4). Using the steering protocol of Eq. (4) and Fig. 3(b) (λ=0.5\lambda=0.5), Fig. R4(a) shows that H5 alone has a substantial effect, H5+H29 is stronger, and H5+H29+H30 nearly reproduces the full L24 block for both FE Rescue (DpFED_p^{FE}) and CC Induction (DpCCD_p^{CC}).

Fig. R4(b) shows that their summed residual contributions organize by incoming carry (0/1/2), with clearer separation for correct donors X′X' and incorrect FE receivers concentrated near adjacent-carry boundaries, consistent with Fig. 5.

We also find that these heads aggregate lower-order operand V information: replacing only these V contents nearly reproduces the full three-head steering effect. H5/H29/H30 also remain the top three influential heads across 2-, 3-, and 4-operand addition (in different orders), arguing against a simple one-head-per-operand correspondence.

Q2. Causal Connection Between Generation and Routing

Of course, these two stages are naturally connected through the emitted digit: L24 controls target-position generation (Sec. 3), and its autoregressive read-back determines subsequent L34 carry routing (Sec. 4). We further investigate whether this is the only pathway.

Using the original Cases A/B/C\mathcal{A}/\mathcal{B}/\mathcal{C}, we apply the L24 Attention rescue (Eq. 4) to A\mathcal{A} using C\mathcal{C} as the donor, but force the original wrong digit dpWd_p^W during read-back, denoted AR,W\mathcal{A}^{R,W}. Thus, A→B\mathcal{A}\rightarrow\mathcal{B} isolates the digit effect, while A→AR,W\mathcal{A}\rightarrow\mathcal{A}^{R,W} isolates the L24-induced cache effect.

Table R3. Next-position effects.

CaseRead-back digitNext Acc.Marginp+1\mathrm{Margin}_{p+1}
A\mathcal{A}Wrong42.25%−0.70
B\mathcal{B}GT98.50%+14.71
AR,W\mathcal{A}^{R,W}Wrong59.38%+0.87

Changing only the digit improves next-position accuracy by 56.25% (+15.41 logits), while changing only the cache improves it by 17.13% (+1.57 logits), revealing a weaker cached-context pathway.

To test whether this additional effect involves the same routing mechanism, we patch the two H24+H27 target QK scores (Eq. 6), preserving all other recipient values. Throughout the following experiments, X←Y\mathcal{X}\leftarrow\mathcal{Y} denotes inserting the two target scores from donor Y\mathcal{Y} into recipient X\mathcal{X}; ΔMarginp+1\Delta\mathrm{Margin}_{p+1} is next-position margin gap defined in Sec. 4.

Table R4. Causal contribution of L34 H24+H27 QK edges.

Edge interventionΔMarginp+1\Delta\mathrm{Margin}_{p+1}
A←AR,W\mathcal{A}\leftarrow\mathcal{A}^{R,W}+1.16
AR,W←A\mathcal{A}^{R,W}\leftarrow\mathcal{A}−0.53

The two edges recover 73.9% of the cache-induced margin gain, while reverse patching removes 33.8%, demonstrating their substantial but partial causal contribution.

Together, generation and routing are connected through both the emitted digit and the updated cached context, with the identified L34 QK edges contributing to both pathways.

Q3. Consistency of the Routing Pattern

We quantify the routing consistency in Fig. 5 over DpFED_p^{FE} at p=1p=1. Following Eq. (7), a case is resolved when the supplied digit belongs to the local argmax set of exactly one of H24/H27; a switch requires both cases to be resolved and assigned to different heads.

Table R5. Head-switch consistency across all first-error triplets.

ComparisonSwitch / AllBoth Resolved / AllSwitch / Resolved
A→B\mathcal{A}\rightarrow\mathcal{B}767/800 (95.88%)767/800 (95.88%)767/767 (100%)
A→C\mathcal{A}\rightarrow\mathcal{C}563/800 (70.38%)664/800 (83.00%)563/664 (84.79%)

Digit replacement (A→B\mathcal{A}\rightarrow\mathcal{B}) switches heads in 95.88% of all cases and 100% of resolved pairs. Context replacement (A→C\mathcal{A}\rightarrow\mathcal{C}) switches in 70.38% of all cases, indicating weaker but substantial context-dependent routing.

This also further clarifies Q2: the L24-only cache intervention A→AR,W\mathcal{A}\rightarrow\mathcal{A}^{R,W} changes the assigned head in just ~1.5% cases, indicating that its downstream effect primarily modulates QK interactions rather than discrete head identity.

Q4. Generalization Beyond Off-by-One Errors

We test whether the neighboring-carry routing mechanism in Sec. 4 extends to other read-back digits and non-off-by-one natural errors, using the same two-edge patching.

First, we extend the off-route analysis in Table 1. Patching from A\mathcal{A} into alternative-digit branches yields ΔMarginp+1=−4.15\Delta \mathrm{Margin}_{p+1}=-4.15 logits, while the reverse yields +7.90+7.90 logits, demonstrating causal influence beyond the original A/B\mathcal{A}/\mathcal{B} comparison.

Second, we examine 800 off-by-one (DpFED_p^{FE}) and 200 non-off-by-one natural first errors. Using the same A←B\mathcal{A}\leftarrow\mathcal{B} two-edge intervention and the head-switch criterion from Q3, we obtain:

Table R6. Routing behavior across natural error types.

First-error typeNext Acc.Distinct Unique HeadsPatching ΔMarginp+1\Delta \mathrm{Margin}_{p+1}
Off-by-one42.25%95.88%+16.05
Non-off-by-one79.50%13.50%+0.66

Non-off-by-one errors rarely show opposite unique head assignments or substantial two-edge effects, while their wrong-digit read-backs often preserve correct next-digit predictions. Thus, the identified edges have broader causal influence, but their paired lower-/higher-carry organization primarily characterizes neighboring carry alternatives.

Q5. Routing Manipulation During Correct Generation

Sec. 4 already shows carry-dependent head preferences on naturally correct controls (Fig. 5). We now directly test their causal role in correct read-back trajectories, using the same two-edge patching protocol from Q2.

Table R7. Two-edge interventions on correct continuations.

CaseNext Acc.Marginp+1\mathrm{Margin}_{p+1}
B\mathcal{B}98.50%+14.71
B←A\mathcal{B}\leftarrow\mathcal{A}79.38%+4.59
C\mathcal{C}100.00%+18.30
C←A\mathcal{C}\leftarrow\mathcal{A}88.50%+12.03

Patching reduces B\mathcal{B} accuracy by 19.12% (ΔMarginp+1=−10.12\Delta\mathrm{Margin}_{p+1}=-10.12 logits). Crucially, patching the naturally correct trajectory C\mathcal{C}, which generates the same read-back digit as A\mathcal{A}, also reduces accuracy by 11.50% (ΔMarginp+1=−6.27\Delta\mathrm{Margin}_{p+1}=-6.27 logits).

Under the head-switch criterion from Q3, C←A\mathcal{C}\leftarrow\mathcal{A} also changes the assigned routing head in 67.38% both resolved cases.

These results demonstrate that L34H24/H27 causally influence subsequent arithmetic computation even during correct generation, rather than functioning only after an error occurs.

Q6. Related Work

Thank you for highlighting these relevant studies! We are glad to expand Sec. 7 to these. These works primarily investigate arithmetic computation and information transfer, our work focuses on how naturally occurring errors arise and propagate across autoregressive steps. We will clarify this distinction in the revision.

Dear Reviewer,

Thank you again for your constructive feedback. We have updated our manuscript to incorporate additional analyses including L24 carry-sensitive heads, causal connection between generation and routing, cross-operation generalization and so on, as discussed in our rebuttal. We would greatly appreciate it if you could take another look.

If you feel that these revisions have adequately addressed your concerns, we would be very grateful if you would consider updating your assessment and score. Should you have any further questions or concerns, we are happy to address them and remain available for further discussion.

Thank you once again for your time and thoughtful consideration!

Thank you for the detailed responses and additional experiments. Thank you for the detailed responses and revisions. I appreciate the additional experiments and clarifications incorporated into the manuscript. In particular, the analyses connecting L24 carry-sensitive computation to downstream L34 routing, along with the head-level characterization and routing-consistency results, help clarify the proposed mechanism.

I will maintain my original score of Weak Acceptance, as my initial assessment was already positive about the paper's contributions, and the revisions further strengthen that assessment.

Reviewer_McPAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how carry-related errors arise and propagate in LLMs, focusing on the addition of three ten-digit integers. Using causal interventions such as activation patching, the authors investigate (1) why an incorrect digit is generated, (2) how feeding back this digit affects subsequent predictions, and (3) whether downstream representations retain evidence of arithmetic inconsistency. The authors identify specific attention components involved in carry-related digit selection and subsequent computation, demonstrating that interventions on their activations can causally influence the model's predictions and even correct or induce errors. Furthermore, the paper shows that models retain detectable signals of arithmetic inconsistency despite following an incorrect continuation. Building on these findings, the authors propose a training-free inference-time correction method that improves answer accuracy.

Critical strengths

The paper’s key strengths are listed below:

  • Sound and well-structured mechanistic analysis. The paper provides a detailed analysis of carry-related arithmetic failures in the addition of three ten-digit integers. Through carefully constructed counterfactuals, activation patching (Sections 3–4), fine-grained attention-head and query–key interventions (Section 4), and downstream compatibility analysis (Section 5), the paper provides convincing causal evidence that the identified mechanisms are indeed responsible for the model’s failure behavior within the task.
  • Effective translation of mechanistic insights into practical improvements. The paper uses the identified routing and uncertainty signals to develop a training-free inference-time correction method (Section 6). This underlines the causal influence of the identified mechanisms and shows that mechanistic interpretability can lead to actionable insights. Additional experiments across different model families and architectures further strengthen the paper's contributions.
Critical weaknesses

The paper’s main weaknesses are listed below:

  • Limited impact and generalizability. While the paper’s analysis is great and convincing, its task scope (carry-related failures in three-operand addition) is very narrow, which limits the broader impact of its findings. The authors evaluate different models, operand lengths, and answer positions in the Appendix. However, it remains unclear whether the identified mechanisms generalize to other settings, such as different numbers of operands, operands with varying lengths, or operand orderings. Furthermore, the analysis is restricted to integer addition, leaving open whether the higher-level findings generalize to floating-point addition or other arithmetic operations (e.g., subtraction or multiplication). Evaluating such variations would help establish whether the findings reveal a more general mechanism or are specific to the narrow task studied.
  • Interpretation of the lower-/higher-carry interpretation of routing heads. Section 4 (Figure 5) interprets L34H24 and L34H27 as representing the lower and higher of two locally relevant carry hypotheses. However, it is unclear how this pair is determined. In particular, when the correct carry is (cp=1c_p=1), both 0 and 2 are feasible alternatives, making carry 1 either the higher or lower hypothesis depending on the comparison. The ambiguity is also reflected in Figure 5, where correct-control samples with similar carry potentials are assigned to different heads. While the counterfactual experiments establish context-dependent routing, they do not fully justify the proposed carry interpretation. The authors should clarify how the relevant carry pair is determined and quantitatively evaluate each head's preferences across all three carry candidates, particularly for (cp=1c_p=1).
Questions for authors

Please find a list of questions below:

  • Figure 5: Why are the markers plotted at different vertical positions within the same attention-head category? Since the y-axis represents a categorical assignment to either L34H24 or L34H27, should the markers not be aligned vertically?
  • Inference-time computational overhead: The proposed correction method in Section 6 evaluates multiple candidate digits at each position, introducing additional inference costs. How much computational overhead does this introduce compared to standard greedy decoding?
5 comments

Thank you for your time and constructive feedback! We have carefully addressed your comments below, and are incorporating some of these new results into our manuscript, which we look forward to sharing soon.

All newly added figures are prefixed with "R" (e.g., Fig. R1) for distinction, which can be viewed in the Supplementary Material.

Q1. Generalization Across Addition Settings

We examine six addition settings varying operand length, operand count, length alignment, decimal arithmetic, operand order, and reasoning context (Fig. R1). Across all six settings, the lower-/higher-carry routing pattern in Fig. 5 largely persists, with frequent head switches under both digit replacement (A→B\mathcal{A}\rightarrow\mathcal{B}) and carry-context changes (A→C\mathcal{A}\rightarrow\mathcal{C}). In mixed-length addition, errors are less concentrated near carry-potential boundaries, potentially reflecting digit-alignment effects, while the routing pattern remains visible.

Q2. Generalization Across Arithmetic Operations

We extend the bidirectional head-output scan in Fig. 12 to subtraction and multiplication (Fig. R2). Excluding the low-level digit-transfer head L1H22, L34H24/H27 remain the top two heads in subtraction, as in addition, and rank second and third in multiplication.

We then apply the original A/B\mathcal{A}/\mathcal{B} read-back protocol and patch only the two L34H24/H27 target QK scores (Eq. 6).

Table R1. Cross-operation routing effects.

TaskNext Acc. A\mathcal{A} / B\mathcal{B}A←B\mathcal{A}\leftarrow\mathcal{B} Patched Acc.ΔMarginp+1A←B\Delta\mathrm{Margin}_{p+1}^{\mathcal{A}\leftarrow\mathcal{B}}ΔMarginp+1B←A\Delta\mathrm{Margin}_{p+1}^{\mathcal{B}\leftarrow\mathcal{A}}A→B\mathcal{A}\rightarrow\mathcal{B} Head Switch
Addition42.25% / 98.50%98.50%+16.05−10.1296.50%
Subtraction17.0% / 100.0%97.0%+19.93−13.5295.5%
Multiplication40.5% / 64.0%44.5%+1.65−0.6429.0%

Head switch denotes distinct resolved H24/H27 assignments under Eq. (7).

Subtraction shows strong transfer: two-edge patching nearly reproduces full L34 attention-pattern patching. Multiplication shows partial transfer, with weaker causal effects and fewer head switches.

In multiplication, the additional head L36H26 appears to relay contextualized next-digit prediction information through its self-position V content, rather than selecting carry routes through target QK edges.

Q3. Relative Carry Routing at cp=1c_p=1

L34H24/H27 encode relative rather than absolute carry preferences.. When cp=1c_p=1, the correct carry is the higher alternative in the 0/10/1 pair near Φp=1\Phi_p=1 (H27), but the lower alternative in the 1/21/2 pair near Φp=2\Phi_p=2 (H24).

To make this explicit, Fig. R3 expands the categorical head assignment of Fig. 5 into a continuous preference score

where ℓh(d)\ell_h(d) is the target-edge scaled QK score from Eq. (7), h∈{L34H24,L34H27}h\in\{\mathrm{L34H24},\mathrm{L34H27}\}. Positive Δh\Delta_h means that head hh prefers the supplied digit over the two neighboring candidates.

Fig. R3 confirms this transition in naturally correct samples (left), while under-carry errors (1→01\rightarrow0, middle) favor H24 and over-carry errors (1→21\rightarrow2, right) favor H27. These continuous preferences explain the overall routing pattern in Fig. 5, including ambiguous assignments near Φp≈1.5\Phi_p\approx1.5.

Thus, H24/H27 exhibit context-dependent lower-/higher-carry preferences, reflecting the ambiguity between neighboring carry hypotheses.

Q4. Vertical Offsets in Figure 5

The y-axis denotes categorical head assignments. Fixed offsets separate correct controls and Cases A/B/C\mathcal{A}/\mathcal{B}/\mathcal{C} within each head, with slight jitter for visibility. These offsets have no quantitative meaning. We will clarify this in the caption.

Q5. Computational Overhead

As stated in Appendix E, our correction experiments serve as functional validation of the identified mechanism, rather than a practical arithmetic-solving method. For real-world usage, using an external calculator is clearly preferable. We nevertheless report latency and FLOPs in the following table for completeness, without claiming computational efficiency or deployment advantages.

Table R2. Inference-time accuracy and computational cost.

MethodLatency (ms)GFLOPsQues Acc (%)
Original450.08747.9460.9
Re-prompting490.321436.8261.4
Steering473.57756.5063.5
Target-state465.00748.0066.4
Dual-stream483.74748.0159.3
Next-state599.11759.7584.4
Entropy1395.11935.7877.9
Route+Entropy1390.90935.7881.7

Dear Reviewer,

Thank you again for your constructive feedback. We have updated our manuscript to incorporate additional analyses including L24 carry-sensitive heads, causal connection between generation and routing, cross-operation generalization and so on, as discussed in our rebuttal. We would greatly appreciate it if you could take another look.

If you feel that these revisions have adequately addressed your concerns, we would be very grateful if you would consider updating your assessment and score. Should you have any further questions or concerns, we are happy to address them and remain available for further discussion.

Thank you once again for your time and thoughtful consideration!

Reviewer_McPAOpenReview ↗

Response to Authors

I thank the authors for their responses to my questions and concerns. Regarding my concerns:

  • W1 (generalization across addition): Thank you for adding the experiments for different addition setups. These results are interesting, and it is great to see that they are largely consistent. However, I think it is important to also discuss the inconsistent results (e.g. mixed operand lengths), at least briefly, in the main text of the revised paper. If you have already done so, I might have missed it. Could you point me to the relevant passage?
  • W1 (generalization across arithmetic operations): Great, thanks for these experiments. Again, very interesting. It seems that the effects shown in Figure R2 are less pronounced for subtraction, and particularly for multiplication (see also your patching experiment). Do you plan to discuss this further in the paper or provide additional evidence supporting your hypothesis that “in multiplication, the additional head L36H26 appears to relay contextualized next-digit prediction information through its self-position V content, rather than selecting carry routes through target QK edges.”?
  • W2: Thanks for the analysis. I think this addresses my initial concern.

Overall, I think the authors have done a good job of addressing my concerns and I will increase my initial score to weak acceptance. However, while I consider the paper technically sound, I still think it studies a rather narrow and specific question. I leave it to the AC to assess its relevance to the broader ICLR community.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.