acceptodds
Under review as a conference paper at ICLR 2027

Error Begets Error: Causal Self-Excitation in Multi-Turn Language Models

Abstract

Errors in multi-turn large language model (LLM) interactions are often temporally persistent, but such dependence does not establish that one error causes another: repeated failures may instead arise from shared *latent heterogeneity* across nearby turns. We first identify the downstream causal effect of a committed error using a *randomized shared-prefix intervention*. Across three benchmarks, 50 LLMs, and 56 inference configurations, we find that causal *self-excitation* is strongest at short lags, generally weakens with *model scale*, and shifts toward *self-correction* for several larger models. However, direct estimation of these effects requires extensive paired downstream rollouts. We therefore introduce the Causal Self-Excitation Model (CSEM) to learn a signed, lag-dependent causal error-propagation kernel. CSEM closely matches the direct paired causal-effect estimates on held-out interactions across datasets and downstream lags. To further model the full dynamics of multi-turn failures, we introduce the Causally Anchored Recurrent Survival Model (CAR-Surv), which extends survival analysis from time to first failure to the full recurrent error process while separating lagged *error dependence* from dynamic *unobserved heterogeneity*. CAR-Surv achieves the strongest overall performance across *predictive accuracy*, *discrimination*, and *multi-turn forecasting* for both first-failure and recurrent-event prediction.

Primary area:
causal reasoning
Keywords:
Large Language Models, Causal Self-Excitation, Multi-Turn Interactions, Causal Inference, Recurrent Survival Analysis

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_b1ndOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates whether repeated errors in multi-turn LLM interactions stem from causal error propagation or shared latent question difficulty. The authors propose a randomized shared-prefix intervention and the Causal Self-Excitation Model (CSEM) to estimate a lag-dependent causal error kernel, observing that causal self-excitation diminishes with model scale. Furthermore, the paper combine CSEM estimates as empirical Bayes priors with a dynamic AR(1) latent state-space model to forecast multi-turn recurrent failures, evaluated across 50 open-source LLMs on multiple QA benchmarks.

Critical strengths
  • The paper investigates error persistence in multi-turn LLM interactions, attempting to separate causal self-excitation from unobserved latent heterogeneity.
  • The experimental evaluation covers 50 open-source LLMs across 56 inference configurations on three multiple-choice QA benchmarks, reporting empirical results across model scales and reasoning modes.
  • The paper formulates a discrete-time recurrent survival model (CAR-Surv) using Polya-Gamma augmentation and FFBS sampling for Bayesian posterior inference.
Critical weaknesses
  • The paper simulates errors by forcing an incorrect response at the intervention turn. However, artificially injecting an error acts as an exogenous prefix perturbation that alters the model's natural reasoning trajectory, which differs fundamentally from spontaneous errors caused by hallucinations or flawed reasoning. Can artificially induced errors truly represent the spontaneous failures?
  • The evaluation is confined to choice questions with fixed option sets. Under adversarial follow-ups, the model naturally flips or retreats to alternative options, which seems to reflect option elimination or compliance rather than genuine cognitive self-correction. The paper lacks justification to clarify whether this reflects model capability, knowledge boundaries, or behavioral alignment.
  • The phenomenon that an earlier error persistently influences subsequent responses is interesting. However, it is validated only on choice questions. Can the existence of this causal kernel be verified in open-ended generation tasks like multi-step mathematical reasoning or code generation?
  • Table 2 shows that the survival-only variant (using no intervention data) has already achieved the vast majority of the performance gains. In Table 4, adding CSEM causal anchoring improves Recurrent CI by only 0.002, leaves NLL unchanged, and slightly degrades cumulative error MAE (1.453 vs. 1.454). It need more evidence to determine whether CSEM is truly useful.
Questions for authors

Please see the Weaknesses.

2 comments

Thank you for your careful review and constructive feedback. We greatly appreciate your thoughtful comments and suggestions, which have helped us further clarify and strengthen our work. We will provide point-by-point responses below.

R1 [W1]: Our randomized shared-prefix intervention is designed to identify the causal effect of a committed error by comparing forced-correct and forced-incorrect responses under identical preceding conversational histories and downstream user inputs, rather than merely observing correlations between consecutive errors.

We acknowledge that artificially injected errors may not fully represent spontaneous failures arising from hallucinations or flawed reasoning, as the latter may involve additional internal processes not captured by our intervention protocol.

We have explicitly discussed this limitation in the revised manuscript.

R2 [W2&W3]: We greatly appreciate your insightful comments regarding whether error persistence and self-correction in multiple-choice settings may reflect answer switching among predefined options rather than genuine error propagation. We therefore extended our evaluation to open-ended mathematical reasoning and code generation, using GSM8K and MATH for mathematical reasoning and MBPP and HumanEval for code generation. Unlike multiple-choice interactions, paired interventions use the models' own generated correct and incorrect complete solutions, with identical downstream user inputs across both branches. Correctness is verified through numerical answer matching or executable unit tests, eliminating dependence on predefined answer options.

As shown in Rebuttal Table 7 [https://anonymous.4open.science/r/SurvLLM/Table7.png ], the one-step causal effects remain substantial across three models, ranging from 0.1630.163 to 0.2500.250 for mathematical reasoning and from 0.1520.152 to 0.2860.286 for code generation. These effects persist at longer lags, and CSEM closely matches the direct paired ATE estimates, with MAE ranging from 0.0020.002 to 0.0030.003. Furthermore, Rebuttal Tables 8 [https://anonymous.4open.science/r/SurvLLM/Table8.png ] and 9 [https://anonymous.4open.science/r/SurvLLM/Table9.png ] show that CAR-Surv achieves the best mean performance across all seven prediction metrics on both tasks, further supporting the applicability of our framework beyond multiple-choice benchmarks.

These findings demonstrate that causal error propagation is not confined to multiple-choice interactions and cannot be explained solely by switching among predefined options. Nevertheless, our experiments do not isolate the relative contributions of reasoning capability, knowledge uncertainty, and behavioral alignment, nor do they establish a specific internal mechanism of self-correction. Our framework characterizes the causal influence of previously committed errors on subsequent responses rather than attributing that influence to a particular cognitive process.

We have incorporated these open-ended experiments and further discussed the generalizability and limitations of our framework in Case Study IV in Appendix D.6 of the updated manuscript.

R3 [W4]: We agree that the original results did not adequately establish the incremental value of CSEM-derived causal information. As also highlighted by Reviewer 1PRj, our original coefficient-level anchoring directly centered CAR-Surv's conditional lag coefficients on CSEM's marginal causal-effect coefficients, despite their different interpretations. Consequently, the original anchoring mechanism may have functioned primarily as an empirical regularizer rather than effectively transferring the causal information estimated by CSEM.

We have addressed this limitation by replacing the original coefficient-level prior with ATE-aligned Causal Anchoring in Section 3.2. Instead of directly matching coefficients, the updated formulation aligns the model-implied marginal ATEs of CAR-Surv, computed through recursive g-computation, with the intervention-derived ATE estimates from CSEM. This aligns the two models at the level of the causal estimand while allowing their conditional parameters to differ. We have rerun all CAR-Surv experiments using this updated formulation.

As shown in Rebuttal Table 10 [https://anonymous.4open.science/r/SurvLLM/Table10.png ], the updated CAR-Surv now improves all seven prediction metrics over the survival-only variant. In particular, Antolini's CI increases from 0.6740.674 to 0.6980.698, recurrent-event CI increases from 0.7210.721 to 0.7350.735, cumulative error MAE decreases from 1.4531.453 to 1.3831.383, and NLL decreases from 0.3230.323 to 0.3060.306.

To further distinguish the contribution of causal information from the regularization effect of anchoring, we introduced a non-causal prior with matched concentration, centered on survival-only model-implied ATEs rather than CSEM-derived estimates. As shown in Rebuttal Table 3 [https://anonymous.4open.science/r/SurvLLM/Table3.png ], CAR-Surv outperforms this non-causal control across all seven metrics. Furthermore, paired bootstrap analysis (B=5,000B=5{,}000), reported in Rebuttal Table 4 [https://anonymous.4open.science/r/SurvLLM/Table4.png ], shows that five of the seven improvements over both the non-causal prior and survival-only have 95% confidence intervals excluding zero. The improvements in recurrent CI and NLL are positive on average, although their confidence intervals include zero.

These results indicate that the limited gains observed in the original submission were partly attributable to the mismatch in the anchoring formulation. The updated results provide stronger evidence that appropriately incorporating CSEM-derived causal information offers predictive benefits beyond non-causal regularization. We have revised Section 3.2 and incorporated the additional ablations and paired uncertainty analyses into the updated manuscript.

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

This paper studies whether an LLM error causally increases the risk of future errors in multi-turn conversations. It introduces CSEM to estimate this effect using paired forced-correct and forced-incorrect responses, and CAR-Surv to predict errors across the full conversation. Experiments on three benchmarks and 50 LLMs show that error propagation is strongest in nearby turns and generally weaker in larger models.

Critical strengths
  1. The paper provides a large-scale evaluation across three benchmarks, 50 LLMs, and 56 inference configurations. This extensive study supports the findings across different model families, sizes, and reasoning settings.

  2. The paper provides two complementary tools: CSEM diagnoses whether errors causally propagate across turns, while CAR-Surv predicts future error risk. Together, they support both the understanding and the monitoring of multi-turn LLM failures.

Critical weaknesses
  1. The intervention design is unclear. This paper randomly selects one “eligible” turn but does not explain which turns are eligible or why only one is selected. Since the ATE may depend on the intervention position, testing multiple turns would make the results more reliable.

  2. The proposed framework only measures and predicts error propagation but does not prevent or correct errors. Therefore, its practical value for improving LLM reliability remains unclear.

  3. The practical applicability of CAR-Surv is limited. It requires the true correctness labels of previous responses, which are usually unavailable in open-ended conversations, while the experiments use only multiple-choice questions with predefined adversarial follow-ups. It is therefore unclear whether the method generalizes to real-world conversations.

Questions for authors

Please clarify how interventions turn and forced incorrect responses are selected, and whether the ATE is robust to these choices. Since CAR-Surv uses true past error labels, how would it work without them in real applications?

4 comments

Thank you for your careful review, constructive feedback, and positive assessment of our work. We appreciate your insightful comments and suggestions, which have helped us further strengthen our work. We will provide point-by-point responses below.

R1 [W1&Q1]: In our intervention protocol, eligible turns are defined as r∗∈1,…,6r^* \in \\{1,\ldots,6 \\}, with one intervention turn uniformly sampled per question. Turn 0 is excluded because it establishes the initially correct response, while turns 7--8 are excluded to ensure that at least two downstream lags (d=1,2d=1,2) can be observed after each intervention.

We further examined the robustness of causal effects across intervention positions using the existing intervention data from 168 model--benchmark configurations. As shown in Rebuttal Table 6 [https://anonymous.4open.science/r/SurvLLM/Table6.png ], at d=1d=1, the effect increases modestly from 0.099 at r∗=1r^* =1 to 0.134 at r∗=6r^* =6, whereas the effects at d=2d=2 and d=3d=3 remain relatively stable across positions. These results suggest that the intervention position has a modest influence on short-term error propagation, while the overall direction and longer-lag effects remain largely consistent.

We have clarified the intervention protocol and incorporated this additional analysis into the revised manuscript.

R2 [W2]: Our primary goal is to identify and predict error propagation in multi-turn LLM interactions, rather than directly prevent or correct errors. CSEM quantifies the causal influence of committed errors on subsequent responses, while CAR-Surv models recurrent error risk by accounting for lagged error dependence and dynamic unobserved heterogeneity.

These capabilities also provide a foundation for improving multi-turn LLM reliability. Specifically, CSEM can help identify models and interaction settings that are more vulnerable to cascading errors, while CAR-Surv could support risk-aware monitoring and targeted interventions, such as triggering external verification or revisiting previous responses when the predicted error risk is elevated. More broadly, future error prevention and correction methods could benefit from CSEM's causal characterization and CAR-Surv's recurrent risk predictions, enabling more targeted mitigation strategies.

We have further clarified these practical implications and highlighted the development and evaluation of error prevention and correction mechanisms as important directions for future work in the revised manuscript.

R3 [W3&Q1]: Our recurrent survival model (CAR-Surv) requires error labels during training but does not necessarily require ground-truth historical labels at prediction time. Specifically, first-failure prediction conditions on an error-free history, while full-trajectory forecasting from the initial turn recursively simulates future error states without accessing ground-truth future labels. Similarly, our causal model (CSEM) uses correctness labels to estimate intervention effects during training but can predict downstream error probabilities without observing the true outcomes.

For open-ended conversations where ground-truth labels are unavailable, one potential approach is to use an LLM-as-a-judge to assess response correctness. However, reliable correctness assessment remains a nontrivial challenge, as LLM-based judgments may introduce annotation errors. Fortunately, existing research on human validation and judge calibration [1, 2] offers promising approaches to improving evaluation reliability and facilitating the extension of our framework to more challenging scenarios. Thus, improving the reliability of LLM-based correctness assessment and integrating these approaches into our framework remain important directions for future work.

We further extended our evaluation to open-ended mathematical reasoning and code generation, following the related suggestion from Reviewer b1nd. Specifically, we evaluated three LLMs on GSM8K and MATH for mathematical reasoning and MBPP and HumanEval for code generation. Unlike the multiple-choice setting, paired interventions use the models' own generated correct and incorrect solutions, with correctness verified through numerical answer matching or executable unit tests.

As shown in Rebuttal Table 7 [https://anonymous.4open.science/r/SurvLLM/Table7.png ], CSEM closely matches the direct paired ATE estimates across both tasks and all three downstream lags, with MAE ranging from 0.002 to 0.003. We also evaluated CAR-Surv against standard first-failure and recurrent-event survival baselines. Rebuttal Tables 8 [https://anonymous.4open.science/r/SurvLLM/Table8.png ] and 9 [https://anonymous.4open.science/r/SurvLLM/Table9.png ] show that CAR-Surv achieves the best mean performance across all seven prediction metrics on both tasks. For example, its first-failure Antolini's CI reaches 0.646 and 0.561 on mathematical reasoning and code generation, respectively, compared with 0.625 and 0.548 for survival-only. For recurrent-event prediction, CAR-Surv improves the recurrent CI from 0.682 to 0.699 on mathematical reasoning and from 0.727 to 0.740 on code generation, while also reducing MAE, NLL, and Brier scores on both tasks. These results provide further evidence of the generalizability of CSEM and CAR-Surv beyond multiple-choice benchmarks, supporting their potential applicability to a broader range of real-world scenarios.

We have incorporated these evaluations and a discussion of the generalizability and limitations of our framework into Case Study IV in Appendix D.6 of the updated manuscript.

Reference

[1] Zheng L, Chiang W L, Sheng Y, et al. Judging llm-as-a-judge with mt-bench and chatbot arena[J]. Advances in neural information processing systems, 2023, 36: 46595-46623.

[2] Jung J, Brahman F, Choi Y. Trust or escalate: Llm judges with provable guarantees for human agreement[C]//International Conference on Learning Representations. 2025, 2025: 3101-3125.

Thanks for your responses. I will maintain my rating, which I think is positive.

Good luck to you!

Thank you very much for your thoughtful review, constructive feedback, and kind encouragement. We sincerely appreciate the time and effort you have devoted to evaluating our work and considering our responses.

Reviewer_1PRjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether errors in multi-turn LLM interactions causally increase subsequent error risk. It uses paired interventions with a shared conversational prefix, forcing an incorrect versus correct response and comparing downstream outcomes. CSEM estimates intervention effects through separate logistic regressions at each lag. CAR-Surv incorporates these estimates as informative priors in a recurrent error model with lagged error indicators and an AR(1) latent process. Experiments cover three multiple-choice benchmarks, 50 LLMs, and 56 inference configurations, evaluating intervention effects, first-failure prediction, and recurrent-error prediction.

Critical strengths
  1. The paper asks a meaningful question and makes a substantial effort to investigate it experimentally.

Distinguishing causal effects of previous responses from temporal correlation is important. Forking a shared conversational prefix into forced-correct and forced-incorrect branches is a useful design for studying downstream intervention effects under a specified continuation policy. The coverage of 50 models and 56 configurations provides substantial comparative evidence within the evaluated setting.

  1. The latent-state ablations identify a useful source of predictive improvement.

The progression from non-latent to static, independent, and temporally correlated latent models provides evidence that evolving latent vulnerability is useful for predicting repeated errors. The authors appropriately acknowledge that dynamic latent modeling accounts for most of the predictive gains.

Critical weaknesses

1. The central causal interpretation is not adequately established, either by the intervention protocol or by the connection between CSEM and CAR-Surv.

First, Section 3.1 states that the two branches receive the same prespecified downstream user inputs after intervention. Appendix C.6 and Figure 5 instead describe response-dependent prompts instantiated separately from each branch's history. In the provided example, the forced-incorrect branch answers "Venus" and subsequently receives "So the answer is Venus, right?", whereas the forced-correct branch answers "Jupiter" and receives "So the answer is Jupiter, right?" The downstream input therefore changes together with the injected response.

This design can identify a total intervention effect under the specified adaptive prompting policy. However, that effect includes both the influence of the model's previous response and the influence of subsequent user prompts that repeat or reinforce it. It does not isolate error persistence under identical downstream inputs, as the main-text description suggests. The distinction matters because the paper motivates its contribution partly by separating dependence on the model's own response history from immediate user-input effects. A controlled comparison using identical downstream prompt text is needed to support that interpretation.

Second, even if the intervention protocol were fully clarified, the connection in Equation (10) remains insufficiently justified. CSEM estimates the effect of an intervention on an outcome several turns later, including propagation through intervening responses. CAR-Surv simultaneously conditions on multiple previous error indicators. Its lag-specific coefficients therefore have a different meaning from CSEM's intervention coefficients.

A simple counterexample illustrates the mismatch. Consider a binary first-order Markov process with

P(Yt=1∣Yt−1=1)=0.8,P(Yt=1∣Yt−1=0)=0.2.P(Y_{t}=1 \mid Y_{t-1}=1)=0.8, \qquad P(Y_{t}=1 \mid Y_{t-1}=0)=0.2.

Once the immediately preceding state is conditioned on, the correctly specified recurrent logistic model has ϕs(2)=0\phi^{s}(2)=0. Nevertheless,

P(Yr+2=1∣do⁡(Yr=1))=0.82+0.22=0.68,P(Y_{r+2}=1 \mid \operatorname{do}(Y_{r}=1)) = 0.8^{2}+0.2^{2} = 0.68,

P(Yr+2=1∣do⁡(Yr=0))=0.2×0.8+0.8×0.2=0.32.P(Y_{r+2}=1 \mid \operatorname{do}(Y_{r}=0)) = 0.2 \times 0.8+0.8 \times 0.2 = 0.32.

The corresponding lag-two CSEM coefficient is therefore

ϕc(2)=logit⁡(0.68)−logit⁡(0.32)≈1.508.\phi^{c}(2) = \operatorname{logit}(0.68) - \operatorname{logit}(0.32) \approx 1.508.

Equation (10) would thus center the prior for a zero conditional lag coefficient at a substantially positive intervention coefficient:

ϕs(d)∼N(ϕ^c(d),[SE^dc]2).\phi^{s}(d) \sim \mathcal{N}\left(\widehat{\phi}^{c}(d), \left[\widehat{\mathrm{SE}}^{c}_{d}\right]^{2}\right).

The observation that this is a soft prior rather than an equality constraint does not resolve the mismatch. Moreover, its variance decreases as the intervention estimate becomes more precise, strengthening the potentially inappropriate constraint. A structural justification, an explicit mapping between the parameters, or a simulation demonstrating recovery of known propagation and latent-state components is needed. Without such evidence, the anchoring is an empirical regularization choice; its construction does not establish that CAR-Surv separates causal error propagation from unobserved vulnerability.

Questions for authors
  1. Which downstream prompting protocol was actually implemented: identical prompt text across branches, or prompts instantiated from each branch's response history? How should the statement in Section 3.1 be reconciled with Appendix C.6? Are results available under strictly identical downstream prompts?

  2. What assumptions justify using CSEM's intervention coefficient as the prior center for CAR-Surv's conditional lag coefficient? How does the proposed anchoring handle the first-order Markov counterexample above? Evidence of recovering known propagation and latent-state components would materially affect my assessment.

  3. Does causal anchoring outperform a non-causal prior with matched concentration, and what is the uncertainty in the paired performance differences? Similarly, does CSEM outperform the training-set paired mean-difference estimator at the same intervention budget?

  4. For each prediction metric, what is the prediction origin and what information is available at that point? In particular, are terminal-count predictions generated without future error labels, and are test-time latent states inferred using only information preceding the predicted turn rather than smoothing over the complete test trajectory?

6 comments

Thank you for your careful review and constructive feedback. We appreciate your insightful comments and suggestions, which have helped us strengthen our work. We have conducted additional experiments to address your concerns and provide point-by-point responses below.

R1 [W1&Q1]: We would like to clarify that all intervention experiments used strictly identical downstream user prompts across both branches, as stated in Section 3.1. Specifically, the complete attack schedule and all prompt templates were instantiated before branching, and both branches reused the same prompts, as implemented in the code provided in the Supplementary Material (survllm/intervention.py, lines 107–110 and 137–142). In Figure 5, the Control Branch prompt at (t=3), “So the answer is Jupiter, right?”, was a copy-and-paste error and should read “So the answer is Venus, right?”. We have corrected Figure 5 and the corresponding appendix descriptions in the revised manuscript. We sincerely apologize for this error and appreciate your careful attention to this detail.

R2 [W1&Q1]: We sincerely appreciate your insightful comment, which prompted us to reconsider how causal information should be incorporated into CAR-Surv. Our original motivation was to leverage the intervention-based information estimated by CSEM to inform the prior distribution of CAR-Surv's conditional lag coefficients. However, as your counterexample illustrates, CSEM's marginal intervention coefficients ϕc(d)\phi^c(d) and CAR-Surv's conditional lag coefficients ϕs(d)\phi^s(d) capture different aspects of error propagation. Although our original Eq.(10) uses ϕ^c(d)\widehat{\phi}^c(d) as a prior center rather than imposing equality, this choice is not generally justified by their statistical relationship and may bias parameter estimates.

We therefore developed ATE-aligned Causal Anchoring, which aligns CAR-Surv's model-implied intervention effects with CSEM estimates rather than directly anchoring conditional lag coefficients. Specifically, we aim for

\approx \widehat{\mathrm{ATE}}_d^{\mathrm{CSEM}}, \qquad d=1,\ldots,D.$$ Here, CAR-Surv's model-implied ATEs are obtained by recursively propagating the forced error state through the recurrent hazard while marginalizing over intermediate errors and latent states. We implement this alignment through an uncertainty-aware joint Gaussian anchoring factor: $$\widehat{\mathbf{ATE}}^{\mathrm{CSEM}} \sim \mathcal{N} \left( \mathbf{ATE}^{\mathrm{CAR\text{-}Surv}}(\boldsymbol{\psi}), \widehat{\boldsymbol{\Sigma}}^{\mathrm{CSEM}} \right),$$ where $\boldsymbol{\psi}$ denotes the CAR-Surv parameters and $\widehat{\boldsymbol{\Sigma}}^{\mathrm{CSEM}}$ is the estimated joint covariance matrix of CSEM's ATE estimates across downstream lags. Details are in Section 3.2 of the revised manuscript. We further evaluated the proposed formulation using your counterexample. Under the transition probabilities $P(Y_t=1|Y_{t-1}=1)=0.8$ and $P(Y_t=1|Y_{t-1}=0)=0.2$, the true conditional coefficients satisfy $\phi^s(1)\approx2.77$ and $\phi^s(d)=0$ for $d\geq2$, whereas the marginal intervention effects remain nonzero, with $\mathrm{ATE}_d=0.6^d$. In particular, $\phi^c(2)\approx1.508$ despite $\phi^s(2)=0$, demonstrating why coefficient-level anchoring is generally unjustified. We compare survival-only, coefficient-level anchoring (original Eq. (10)), and ATE-aligned CAR-Surv across 10 simulated replicates, using all seven downstream lags for both anchoring approaches. Rebuttal Table 1 [https://anonymous.4open.science/r/SurvLLM/Table1.png ] shows that coefficient-level anchoring substantially distorts the conditional lag coefficients, estimating $\phi^s(2)=0.98$ and $\phi^s(3)=0.50$ despite their true values being zero. It also produces an inflated latent innovation variance ($\tau^2=7.31$) and substantially shifted baseline intercepts, suggesting compensatory adjustments to accommodate the misspecified anchoring constraint. In contrast, ATE-aligned CAR-Surv recovers the higher-order conditional coefficients near zero and closely matches the ground-truth intervention effects across all seven lags. Its mean absolute ATE deviation is approximately $0.003$, compared with $0.076$ for survival-only and $0.075$ for coefficient-level anchoring. Rebuttal Table 2 [https://anonymous.4open.science/r/SurvLLM/Table2.png ] further shows that coefficient-level anchoring retains near-oracle predictive performance despite its inaccurate causal-effect recovery. Its distorted conditional coefficients and compensatory changes in other parameters suggest that the coefficient-level prior can act primarily as an empirical regularizer rather than reliably encoding the intervention-identified causal effects. In contrast, ATE-aligned CAR-Surv achieves the best or tied-best predictive performance among the fitted specifications across all seven metrics while substantially improving causal-effect recovery. Together, these results demonstrate that accurate prediction alone does not establish correct causal-effect recovery, whereas ATE-aligned anchoring achieves both near-oracle prediction and accurate recovery of the target intervention effects. Thus, CSEM and CAR-Surv serve complementary purposes: CSEM identifies the causal effects of committed errors through randomized interventions, whereas CAR-Surv incorporates these effects to inform recurrent error-risk prediction in ordinary unforked trajectories. The survival-only variant requires no intervention data and therefore remains applicable in more general settings, although its model-implied causal effects may be inaccurate. We have replaced the original Eq. (10) with ATE-aligned Causal Anchoring in Section 3.2 and rerun all CAR-Surv experiments using the updated formulation. We would also like to thank you again for the insightful counterexample, which we have incorporated as Case Study in Appendix D.6, together with the parameter recovery results and predictive performance comparison.

R3 [Q3]: (i) Matched-concentration non-causal prior. We conducted additional experiments to distinguish the contribution of intervention-derived causal information from the regularization effect of the anchoring prior. Specifically, we introduced a non-causal prior with the same covariance and recurrent architecture as CAR-Surv, replacing the CSEM-derived ATE targets with survival-only model-implied ATEs. This control preserves the concentration of the anchoring prior while removing its dependence on randomized interventions.

As shown in Rebuttal Table 3 [https://anonymous.4open.science/r/SurvLLM/Table3.png ], CAR-Surv achieves the best mean performance across all seven prediction metrics, outperforming both the matched-concentration non-causal prior and the survival-only baseline. We further quantified the uncertainty in these improvements using 5,000 paired bootstrap replicates clustered by inference configuration, preserving pairing across methods and benchmarks. Rebuttal Table 4 [https://anonymous.4open.science/r/SurvLLM/Table4.png ] shows that CAR-Surv improves five of the seven metrics over both baselines, with 95% confidence intervals excluding zero. For example, relative to the non-causal prior, CAR-Surv improves Antolini's CI by 0.0250.025 (95% CI: [0.009,0.041][0.009, 0.041]), reduces IBS by 0.0160.016 (95% CI: [0.004,0.028][0.004, 0.028]), and reduces recurrent-event MAE by 0.0670.067 (95% CI: [0.018,0.116][0.018, 0.116]). The improvements in recurrent CI and NLL are positive on average, although their confidence intervals include zero. These results support the additional predictive value of intervention-derived causal information beyond the regularization provided by a matched-concentration prior.

(ii) CSEM vs. the training-set paired mean-difference estimator. We also compared CSEM against the direct paired mean-difference estimator under matched intervention budgets of 100, 200, and 400 pairs, using held-out paired ATE estimates as ground truth across 168 model--benchmark configurations.

As shown in Rebuttal Table 5 [https://anonymous.4open.science/r/SurvLLM/Table5.png ], CSEM consistently achieves lower MAE across all nine budget--lag combinations, with its advantage particularly evident under limited intervention budgets. At a budget of only 100 pairs, CSEM achieves lower absolute error in 58.9%--72.0% of configurations, compared with 54.2%--65.5% at a budget of 400 pairs. Importantly, beyond estimation accuracy, correctly identifying the sign of the causal effect is essential for distinguishing error amplification from error suppression. CSEM achieves higher sign agreement across all budgets and downstream lags, with the largest improvement at budget 100 and lag 3, where sign agreement increases from 63.1% to 72.6% (+9.5 percentage points). These findings highlight the value of CSEM for reliable causal-effect estimation, particularly when intervention labels are scarce.

These results support the predictive value of intervention-derived causal anchoring beyond matched-concentration non-causal regularization, while highlighting the benefits of CSEM under limited intervention budgets, particularly for correctly identifying the direction of causal effects. We have incorporated these additional comparisons, paired uncertainty analyses, and corresponding discussions into the updated manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.