acceptodds
Under review as a conference paper at ICLR 2027

SENECA: Small-Sample Discrete Entropy Estimation via Self-Consistent Missing Mass

Abstract

Discrete entropy estimation is a classic information theory problem, wherein the average information content of a discrete random variable is estimated from samples alone. Naive approaches, such as the plugin method, fail to account for the probability mass associated with members of the random variable's support that are unobserved in a given sample, known as the "missing mass." The resulting systemic underestimation is particularly problematic when data is time-consuming or costly to gather. We propose SENECA, an entropy estimation scheme based on a novel "self-consistent" missing mass calculation. Extensive numerical experiments indicate that our approach outperforms many state-of-the-art alternatives overall in the small-sample setting. We then apply SENECA to two practical use cases, namely biodiversity estimation and the detection of incorrect large language model responses, where our method is competitive with domain-specific approaches. Our work advances SENECA as an effective drop-in replacement for small-sample entropy estimation, with broad utility across several domains.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
entropy, functional estimation, uncertainty quantification, missing mass, hallucination, biodiversity, under-sampled

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_SL65OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The manuscript proposes SENECA, a new entropy estimator for discrete distributions in the small-sample and under-sampled regime. The central idea is to estimate missing mass through a fixed-point equation that combines observed empirical label frequencies with an estimate of the number of unobserved labels, obtained from a support-size estimator. The resulting missing-mass estimate is then plugged into a Chao–Shen/Horvitz–Thompson-style entropy estimator. The authors provide existence and lasymptotical consistency results, extensive simulation experiments over 72 synthetic distributions, an application to 58 biodiversity collections, and an application to LLM uncertainty quantification.

Critical strengths

The methods presented in the manuscript demonstrated competitive performance in the simulation experiments.

Critical weaknesses

1 The manuscript provides a proof of the asymptotical consistency of its method. However it is well known that several commonly used entropy estimators possess asymptotical consistency. The manuscript did not discuss whether the method possesses more interesting statistical characteristics such as unbiasedness or sample complexity.

2 Uniqueness and convergence of the fixed-point estimator are not established. Lemma 1 proves existence of a fixed point of Eq. 5, but not uniqueness. Algorithm 1 uses Steffensen iteration from an initial value of 0.5. If multiple fixed points exist, it is unclear which one is selected and whether the iteration is guaranteed to converge to the intended solution.

In summary, there is room for further improvement in the theoretical solidity of the manuscript.

Questions for authors

The methods presented in the manuscript seem to perform less well on real data compared to that on simulated data. Could the author provide substantive analysis and explanation on this matter?

3 comments

We thank Reviewer SL65 for their time spent reviewing our submission, as well as their constructive comments. We are encouraged that the Reviewer points out ”competitive performance” in ”extensive simulation experiments” and finds no substantive concerns about the paper’s soundness.

We also appreciate the Reviewer’s patience, as we prepared a thorough response to accommodate as many of the Reviewers’ helpful comments as possible (our efforts proved a bit more time-consuming than initially anticipated). Our revised manuscript highlights substantive additions in blue. We have also added a small table of contents to the appendix section with auxiliary proofs, for the Reviewer’s organizational convenience.

We respond to the Reviewer’s comments below:

<<W1>> The manuscript provides a proof of the asymptotical consistency of its method. However it is well known that several commonly used entropy estimators possess asymptotical consistency. The manuscript did not discuss whether the method possesses more interesting statistical characteristics such as unbiasedness or sample complexity.

We agree that some entropy estimators (e.g., Plugin and Chao-Shen) are known to be asymptotically consistent (which we note in Section 6.2), but the results the Reviewer describes can be quite challenging to produce and are often not established in the works introducing the methods. For example, we are not aware of general proofs of consistency, unbiasedness, or sample complexity in the papers introducing the Chao-Shen, Chao-Wang-Jost, James-Stein, or Bonachela estimators. Consequently, empirical analyses are more typical, like the ones we perform. In Figure 8, we characterize the included entropy estimators' biases empirically, and we hope this is satisfactory to the Reviewer. We conclude by noting that, in general, unbiased entropy estimators do not exist.

<<W2>> Lemma 1 proves existence of a fixed point of Eq. 5, but not uniqueness. Algorithm 1 uses Steffensen iteration from an initial value of 0.5. If multiple fixed points exist, it is unclear which one is selected and whether the iteration is guaranteed to converge to the intended solution.

This is a good point, and we agree that establishing the uniqueness of the fixed point is important. In the revision, we have identified a minimally restrictive upper bound on υ\upsilon which allows us to prove uniqueness of the fixed-point solution on [0,1][0, 1] (Thms. 1 and 2). After repeating our experiments with this upper bound, we discern no difference in the headline results arising from this update (e.g., Table 1 is identical).

<<Q1>> The methods presented in the manuscript seem to perform less well on real data compared to that on simulated data. Could the author provide substantive analysis and explanation on this matter?

We have added a paragraph discussing this topic to the Discussion. For convenience, we reproduce it below:

Finally, we observe that any advantage SENECA demonstrates in Section 5 appears more modest than that of Section 4. We attribute this to two factors. First, SENECA targets a particular setting: the small-sample and under-sampled regime, but many of the species distributions in the biodiversity study are well-sampled (≈ 29% for 𝑛 = 10). With support sizes as low as 2, the degenerate case of ∣X≠n∣=1|X_\neq^n|=1 is possible, where SENECA is equivalent to the plugin estimator (Corollary 1). Second, the LLM UQ study is a bipartite ranking task that does not strictly assess entropy estimation ability (note the highest-ranking method is a support size estimator). In general, there exists no discrete entropy estimator that is optimal in every setting (La Torre et al., 2025). Nonetheless, our results indicate that SENECA has broadly favorable performance and may be utile to many diverse technical disciplines.

We reiterate our gratitude for the Reviewer's engagement, which has helped refine our work. Please let us know if any questions remain.

Dear Reviewer SL65,

As the discussion period comes to a close, we are reaching out to gently follow up on this exchange. Your initial feedback was greatly appreciated, and we believe our efforts to address these concerns have led to an improved submission.

In short, we responded with regard to proofs of unbiasedness and sample complexity (W1) and added a proof of uniqueness of the fixed point solution (W2).

Please let us know if you feel that outstanding items persist, and we will do our best to address them in the time remaining.

Thank you,

Authors

Dear Reviewer SL65,

As we enter the final hours of the discussion period, we wanted to reach out once more to continue this productive dialogue, wherein the Reviewer pointed out several interesting topics. We are happy to answer any lingering questions, to the extent time allows.

We found the Reviewer’s feedback thoughtful and constructive, and we are grateful they lent their time to the consideration of our work.

All the best,

Authors

Reviewer_T5FjOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a discrete Shannon entropy estimator designed for the small-sample regime (n=10–50n = 10\text{–}50), particularly when the support size ∣X∣|\mathcal{X}| exceeds the sample size nn (the under-sampled regime). SENECA retains the exact Horvitz–Thompson coverage-adjusted entropy formulation of Chao & Shen (2003), which scales empirical frequencies by estimated sample coverage (1−M^)(1 - \hat{M}) and divides each summand by its inclusion probability 1−(1−p^u)n1 - (1 - \hat{p}_u)^n.

Instead of using the classical Good–Turing missing mass estimator M^GT=Φ1/n\hat{M}^{GT} = \Phi_1 / n, the authors introduce a "self-consistent" missing mass estimator M^SC\hat{M}^{SC}. Specifically, they decompose the expected missing mass E[M]=∑u∈Xpu(1−pu)n\mathbb{E}[M] = \sum_{u \in \mathcal{X}} p_u(1 - p_u)^n into observed and unobserved components over a realized sample XnX^n, plug in isotropically shrunk empirical frequencies (1−m)p^uML(1 - m)\hat{p}_u^{ML} for the observed classes and a uniform split m/υm / \upsilon across υ=max⁡(∣X^∣−∣X≠n∣,1)\upsilon = \max(|\widehat{\mathcal{X}}| - |X^n_{\neq}|, 1) unobserved classes (where ∣X^∣|\widehat{\mathcal{X}}| is obtained from an off-the-shelf support-size estimator such as Regularized Weighted Chebyshev, Chien & Milenkovic, 2019), and solve the 1D fixed-point equation μ(m∗,υ,Xn)=m∗\mu(m^*, \upsilon, X^n) = m^*. The authors prove weak consistency of SENECA over finite support (extending Vu et al., 2007) and validate the estimator on: (1) synthetic item-capture simulations across 8 distribution families and 9 support sizes at n=10n=10 and n=20n=20; (2) biodiversity estimation on 58 insect sweep-sample collections from Janzen (1973b); and (3) black-box LLM hallucination detection across 5 models and 6 QA datasets.

Critical strengths
  1. Practical, low-variance stabilization of Chao–Shen coverage adjustment in the ultra-small-sample (n=10n = 10) under-sampled regime. Classical coverage-adjusted entropy estimators (Chao & Shen, 2003; Chao et al., 2013) rely on only one or two frequency-of-frequency (FoF) counts (Φ1,Φ2\Phi_1, \Phi_2). At n=10n = 10, Φ1\Phi_1 exhibits high variance and suffers from severe boundary degeneracies—collapsing to the negatively biased plugin estimator when Φ1=0\Phi_1 = 0 and becoming undefined (0log⁡0/00 \log 0 / 0) or requiring ad-hoc clipping when all samples are singletons (Φ1=n\Phi_1 = n). By defining M^SC\hat{M}^{SC} as the fixed point of μ(m,υ,Xn)=m\mu(m, \upsilon, X^n) = m, SENECA smoothly couples all observed empirical frequencies with an external support-size estimate ∣X^∣|\widehat{\mathcal{X}}| while guaranteeing M^SC∈[0,1)\hat{M}^{SC} \in [0, 1). This yields a favorable bias–variance tradeoff and consistently lower RMSE than 7 baseline estimators in under-sampled synthetic benchmarks (Tables 1–2) and across the 58 ecological distributions (Figure 3), which is consequential because practitioners rarely know in advance whether their sample is well-sampled or under-sampled.
Critical weaknesses
  1. Inaccurate analysis of support-size dependence in Eq. 4 and unstated assumptions on ∣X^∣|\widehat{\mathcal{X}}| in Theorem 1. On p. 4, the authors claim SENECA is insensitive to ∣X^∣|\widehat{\mathcal{X}}| "since the 'unobserved' component of Eq. 4 is small." However, rearranging the fixed-point equation μ(m,υ,Xn)=m\mu(m, \upsilon, X^n) = m yields Observed(m)=m[1−(1−m/max⁡(υ,1))n]≈nm2/υ\text{Observed}(m) = m[1 - (1 - m/\max(\upsilon, 1))^n] \approx n m^2 / \upsilon. Because m(1−m/υ)n→mm(1 - m/\upsilon)^n \to m as υ\upsilon increases, the unobserved term is not small; rather, m∗m^* depends directly on υ=max⁡(∣X^∣−∣X≠n∣,1)\upsilon = \max(|\widehat{\mathcal{X}}| - |X^n_{\neq}|, 1). The low sensitivity in Figure 5 appears to stem from RWC and Wu–Yang sharing the same pmin⁡p_{\min} regularization heuristic at n=10n=10 (Appendix B.2). Relatedly, the consistency proof in Lemma 3 (line 845) asserts max⁡(υ,1)≤∣X∣\max(\upsilon, 1) \le |\mathcal{X}|, which is not guaranteed when υ\upsilon is computed from an estimator ∣X^∣|\widehat{\mathcal{X}}| that can overestimate ∣X∣|\mathcal{X}| unless conditions on ∣X^∣|\widehat{\mathcal{X}}| are explicitly assumed in Theorem 1.

  2. Omission of bias-corrected entropy baselines and marginal gains on LLM UQ (Section 5.2). While Sections 4 and 5.1 compare SENECA against six bias-corrected entropy estimators (such as Chao–Shen, Chao–Wang–Jost, and James–Stein), Section 5.2 compares SENECA's semantic entropy only against the naive Plugin estimator (DSE). Moreover, SENECA's aggregated AUROC improvement over DSE is very small (~0.71 vs. ~0.70 in Figure 4, with heavily overlapping confidence intervals across Figure 10), making it unclear whether SENECA provides a meaningful advantage over standard Chao–Shen or DSE for LLM hallucination detection.

Questions for authors
  1. Role of ∣X^∣|\widehat{\mathcal{X}}| in Eq. 4 and Theorem 1: Since μ(m)=m\mu(m) = m is equivalent to Observed(m)=m[1−(1−m/max⁡(υ,1))n]\text{Observed}(m) = m[1 - (1 - m/\max(\upsilon, 1))^n], m∗m^* is directly governed by υ\upsilon. Can you clarify the claim on p. 4 that the unobserved component is small, and state the exact condition required on the support-size estimator ∣X^∣|\widehat{\mathcal{X}}| (e.g., ∣X^∣=oP(n)|\widehat{\mathcal{X}}| = o_P(n)) for Lemma 3 and Theorem 1 to hold?
  2. Other entropy estimators on LLM UQ: How do Chao–Shen, Chao–Wang–Jost, and the coverage-adjusted semantic entropy estimators from McCabe et al. (2026) perform on the LLM UQ benchmark in Figure 4 compared to SENECA?
  3. Minor corrections: In Lemma 4 (line 861), H^M^HT→p0\hat{H}_{\hat{M}}^{HT} \xrightarrow{p} 0 should be H^M^HT→pH(P)\hat{H}_{\hat{M}}^{HT} \xrightarrow{p} \mathbb{H}(P). In Table 2's caption (line 1203), please clarify whether the entries are normalized MSE or unnormalized RMSE.
4 comments

We appreciate Reviewer T5Fj’s careful reading and insightful remarks. We are thrilled that the Reviewer has highlighted our “practical, low-variance stabilization’’ that ”yields a favorable bias–variance tradeoff” and is ”consequential” for practitioners.

We also appreciate the Reviewer’s patience, as we prepared a thorough response to accommodate as many of the Reviewers’ helpful comments as possible (our efforts proved a bit more time-consuming than initially anticipated). Our revised manuscript highlights substantive additions in blue. We have also added a small table of contents for the appendix of auxiliary proofs, for the Reviewer’s organizational convenience.

We respond to the Reviewer’s comments below:

<<W1a>> the unobserved term is not small

We agree and have removed this sentence.

<<W1b>> The low sensitivity in Figure 5 appears to stem from RWC and Wu–Yang sharing the same pminp_{min} regularization heuristic at n=10n=10 (Appendix B.2).

This is an interesting point, and it may be partially related. However, the RMSE is also comparable with the Chao1 estimator, which does not employ the heuristic. Averaged over all possible fingerprints at n=10n=10, the four support size estimators (Chao1, RWC, Wu-Yang, Valiant-Valiant) differ mutually by just ≈0.38k\approx 0.38k, where kk is the number of observed categories. Taken together, these observations substantiate our interpretation that our results are not particularly sensitive to the choice of support size estimator. We have added a brief remark to this effect in Appendix B.2.

<<W1c>> and <<Q1>> Regarding conditions on |\hat{\mathcal{X}| for Lemma 3 and Theorem 1 (now Theorem 3)

We thank the Reviewer for pointing out this oversight, which we have corrected. Our revised proof requires an upper bound on υ\upsilon that is sub-quadratic in nn.

Further, we identify and implement one such upper bound that is not restrictive (Appendix A.6.1) and allows us to prove uniqueness of the fixed-point solution on [0,1] (Thms. 1 and 2). After repeating our experiments with this upper bound, we discern no difference in the headline results arising from this update (e.g., Table 1 is identical). Appendix A.6 visualizes and discusses the impacts of this bound.

<<W2>> and <<Q2>> Regarding additional baselines for Section 5.2

We have updated the LLM experiment to include all alternative estimators from the prior sections. Bonachela, Chao-Shen, Chao-Wang-Jost, James-Stein, and SENECA all achieve aggregated AUROC point estimates that exceed those of SE and DSE. Interestingly, these ``advanced’’ entropy estimators perform very similarly on this task, with AUROC point estimates within ≈0.00055\approx 0.00055 of one another.

We also added two more estimators: the “hybrid” entropy estimator from McCabe et al. that you suggested and a simple variant of SENECA (namely, SENECA-U) which uses the semantic support size estimator UEigVU_{EigV} in place of the ordinary RWC. Both out-rank all other methods (besides UEigVU_{EigV}, which remains top-performing): the order is UEigVU_{EigV}, hybrid, SENECA-U.

<<Q3>> Minor corrections

Thank you for pointing these out; we have corrected them. The mention of normalized MSE was an error; these are averaged unnormalized RMSE values, like in Table 1.

We reiterate our gratitude for the Reviewer's engagement, which has helped refine our work. Please let us know if any questions remain.

Dear Reviewer T5Fj,

As the discussion period comes to a close, we are reaching out to gently follow up on this exchange. Your initial feedback was greatly appreciated, and we believe our efforts to resolve these concerns have led to an improved submission.

In short, our revision strengthens the consistency proof (W1) and adds the recommended estimators to the LLM experiment (W2).

Please let us know if you feel that outstanding items persist, and we will do our best to address them in the time remaining.

Thank you,

Authors

Thank you for these clarifications and the new experiments, these are helpful to see. I have raised my score.

Dear Reviewer T5Fj,

We thank the Reviewer for their time, thorough review, and constructive feedback.

We are glad to have meaningfully addressed the Reviewer’s concerns and grateful for the constructive exchange.

All the best,

Authors

Reviewer_XgNoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SENECA addresses the problem of estimating discrete entropy from very few observations. The main contribution is an estimator of missing mass: the probability that the next observation belongs to a previously unseen category. The method first estimates how many categories remain unobserved. Then, it adjusts the observed frequencies to reserve probability for those categories and solves a fixed-point equation so that the assumed missing mass agrees with the expected missing mass implied by the estimated distribution. Finally, it incorporates this estimate into a coverage-adjusted entropy calculation based on the Chao–Shen approach. The empirical results demonstrate that SENECA is a useful approach when samples are scarce and unseen categories are common. However, it does not outperform every estimator in every setting, and its LLM improvements are modest.

Critical strengths
  1. The paper addresses a well-motivated, practical problem. When only a few samples are available, standard entropy estimates miss unseen possibilities.
  2. In the main simulation study, SENECA achieves the lowest average error across all evaluated eight distribution families. The paper also evaluates across diverse applications.
  3. For LLM uncertainty estimation, the method doesn’t require internal LLM probabilities.
Critical weaknesses
  1. A step in the consistency proof assumes a bound on the estimated number of unseen categories that the paper does not establish. The theorem needs clarification or a revised proof.
  2. The method estimates how many categories remain unseen and assigns them equal probabilities, but doesn’t address what happens when these assumptions fail.
  3. The paper compares SENECA with standard empirical entropy and several other LLM uncertainty methods. But it does not include Chao–Shen and the other established entropy corrections in that LLM experiment, even though it evaluates them in its synthetic and biodiversity experiments. This makes it harder to identify the specific benefit of SENECA. Does SENECA’s particular correction improve error detection, or would an existing correction work equally well?
Questions for authors

See Weaknesses above.

3 comments

We thank Reviewer XgNo for their time and detailed feedback. We are pleased that the Reviewer finds that our work “addresses a well-motivated, practical problem,” and that our method is “a useful approach when samples are scarce and unseen categories are common.”

We also appreciate the Reviewer’s patience, as we prepared a thorough response to accommodate as many of the Reviewers’ helpful comments as possible (our efforts proved a bit more time-consuming than initially anticipated). Our revised manuscript highlights substantive additions in blue. We have also added a small table of contents for the appendix of auxiliary proofs, for the Reviewer’s organizational convenience.

We respond to the Reviewer’s comments below:

<<W1>> A step in the consistency proof assumes a bound on the estimated number of unseen categories that the paper does not establish.

We thank the Reviewer for pointing out this oversight, which we have corrected. Our revised proof requires an upper bound on υ\upsilon that is sub-quadratic in nn.

We identify and implement one such upper bound that is not restrictive (Appendix A.6.1) and allows us to prove uniqueness of the fixed-point solution on [0,1][0, 1] (Thms. 1 and 2). After repeating our experiments with this upper bound, we discern no difference in the headline results arising from this update (e.g., Table 1 is identical). Appendix A.6 visualizes and discusses the impacts of this bound.

<<W2>> The method estimates how many categories remain unseen and assigns them equal probabilities, but doesn’t address what happens when these assumptions fail.

We have added a section (Appendix D) discussing this matter. For convenience, we reproduce it below:

In our problem statement, one does not know a priori the shape of the unobserved portion of the data-generating distribution. For this reason, we employ a “clipped” estimate of the unobserved support size and assign all elements of the estimated support equal prevalences. In practice, these support size and prevalence estimates may diverge from ground truth, but the results of Section 4 do not indicate that this is particularly problematic. First, our analysis of the “support-risky” regime indicates that SENECA performs favorably even where no consistent estimator for support size exists. Additionally, the “equal probabilities” assumption fails by construction for several distributions therein - e.g., Zipf-1.5, where SENECA nonetheless has the lowest regime-averaged error in the under-sampled setting (Table 1).

<<W3>> But it does not include Chao–Shen and the other established entropy corrections in that LLM experiment, even though it evaluates them in its synthetic and biodiversity experiments. This makes it harder to identify the specific benefit of SENECA. Does SENECA’s particular correction improve error detection, or would an existing correction work equally well?

We have updated the LLM experiment to include all alternative estimators from the prior sections. Bonachela, Chao-Shen, Chao-Wang-Jost, James-Stein, and SENECA all achieve aggregated AUROC point estimates that exceed those of SE and DSE. As the Reviewer suspected, these ``advanced’’ entropy estimators perform comparably, with AUROC point estimates within ≈0.00055\approx 0.00055 of one another. Grassberger Valiant-Valiant rank lower on this task.

We also added two more estimators: the “hybrid” entropy estimator requested by Reviewer T5Fj and a simple variant of SENECA (namely, SENECA-U) which uses the semantic support size estimator UEigVU_{EigV} in place of the ordinary RWC. Both out-rank all other methods (besides UEigVU_{EigV}, which remains top-performing): the order is UEigVU_{EigV}, hybrid, SENECA-U.

We have added the following reflection to the Discussion section, which may be of interest to the Reviewer, as well:

Finally, we observe that any advantage SENECA demonstrates in Section 5 appears more modest than that of Section 4. We attribute this to two factors. First, SENECA targets a particular setting: the small-sample and under-sampled regime, but many of the species distributions in the biodiversity study are well-sampled (≈ 29% for 𝑛 = 10). With support sizes as low as 2, the degenerate case of ∣X≠n∣=1|X_\neq^n|=1 is possible, where SENECA is equivalent to the plugin estimator (Corollary 1). Second, the LLM UQ study is a bipartite ranking task that does not strictly assess entropy estimation ability (note the highest-ranking method is a support size estimator). In general, there exists no discrete entropy estimator that is optimal in every setting (La Torre et al., 2025). Nonetheless, our results indicate that SENECA has broadly favorable performance and may be utile to many diverse technical disciplines.

We reiterate our gratitude for the Reviewer's engagement, which has helped refine our work. Please let us know if any questions remain.

Dear Reviewer XgNo,

As the discussion period comes to a close, we are reaching out to gently follow up on this exchange. Your initial feedback was greatly appreciated, and we believe our efforts to resolve these concerns have led to an improved submission.

In short, our revision strengthens the consistency proof (W1), discusses what happens when the support size estimate and “equal probabilities” assumption diverge from ground truth (W2), and adds the recommended estimators to the LLM experiment (W3).

Please let us know if you feel that outstanding items persist, and we will do our best to address them in the time remaining.

Thank you,

Authors

Dear Reviewer XgNo,

As we enter the final hours of the discussion period, we wanted to reach out once more to continue this productive dialogue, wherein the Reviewer pointed out several interesting topics. We are happy to answer any lingering questions, to the extent time allows.

We found the Reviewer’s feedback thoughtful and constructive, and we are grateful they lent their time to the consideration of our work.

All the best,

Authors

Comments on the paper

1 comment

As the discussion period is soon coming to a close, we take this opportunity to reiterate our appreciation for the Reviewers’ careful assessments. The Reviewers highlighted several strengths, including those pertaining to the performance of our method (XgNo, T5Fj, SL65) and the practical value of the problem we tackle (XgNo, T5Fj).

The Reviewers also offered constructive feedback, and we appreciate the opportunity to refine our work in response. Through the discussion period, we aimed to respond thoroughly to all reviewer comments, with updates including:

  • improved consistency proof (Section 3 and Appendix A.2; Reviewers XgNo and T5Fj),

  • additional baselines in the LLM experiment (Section 5; Reviewers XgNo, T5Fj),

  • further discussion of what happens when certain assumptions fail (Appendix B.2; Reviewer XgNo), and

  • proof of uniqueness of the fixed point solution (Section 3 and Appendices A.3-A.5; Reviewer SL65).

Finally, we also updated the code included as supplementary material, in case that is of interest in the following stage.

To conclude, we thank the Reviewers and Area Chair again for their invaluable contributions throughout this process.

All the best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.