acceptodds
Under review as a conference paper at ICLR 2027

Saturation-Insensitive Dueling Bandits with General Function Approximation

Abstract

We study contextual dueling bandits with general function approximation under the Bradley–Terry–Luce (BTL) preference model. A key challenge in this setting is the *saturation* of the preference model: when the current reward model can already distinguish two actions with high confidence, the resulting preference feedback becomes weakly informative, making it difficult to further improve reward estimation. Consequently, existing sample-complexity analyses often depend on the inverse-derivative factor which can be prohibitively large when the link function saturates for large reward gaps . To address this issue, we introduce `SI-CDB`, an algorithm that selects opponent arms using a carefully designed heuristic for arm selection. This design enables saturation-insensitive reward learning and recovers the near-optimal dependence for linear reward classes, eliminating the unfavorable factor. The core of our analysis is a localized Eluder dimension framework tailored to dueling bandits with general function approximation. Our theoretical results also explain why two-arm regret analysis is crucial for improving single-arm performance in dueling bandits.

Primary area:
learning theory
Keywords:
Saturation, dueling bandit, localized eluder dimension

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Tm6wOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies realizable contextual dueling bandits under the Bradley–Terry–Luce model with bounded latent rewards. It proposes SI-CDB, combining a log-loss confidence set, a greedy target arm, and an optimistic comparator. The asymmetric selection rule bounds two-arm regret through estimation errors on the queried pair, enabling a localized excess-loss analysis. The general-class guarantee is expressed through covering numbers and localized eluder dimensions. For linear rewards, the leading term is O~(dBT)\widetilde O(d\sqrt{BT}), while an additive term retains exponential dependence on BB. Validation consists of theoretical proofs; no experimental or application study is provided, and the guarantees assume exact optimization.

Critical strengths
  1. The asymmetric comparator objective bounds regret using estimation errors on the queried pair, providing a useful bridge from preference prediction to two-arm latent-reward regret.

  2. For linear rewards, the leading regret term improves the explicit reward-scale dependence while preserving optimal dimension and horizon dependence, providing a concrete theoretical advance.

Critical weaknesses
  1. The practical significance of the improvement is unvalidated. There are no experiments, application studies, or ablations. This matters because Corollary 5.5 retains an additive term proportional to 1/[σ′(2B)]21/[\sigma'(2B)]^{2}, which can dominate the bound at finite horizons. Consequently, the paper does not establish whether the proposed comparator improves query efficiency over existing methods at realistic budgets. Appendix G describes approximate optimization procedures, but provides neither empirical evaluation nor guarantees for their approximation errors.

  2. The benefit beyond linear reward classes is insufficiently substantiated. The general theorem depends on localized eluder dimensions and metric entropy, but the paper explicitly bounds these quantities only for linear rewards. It therefore remains unclear which representative nonlinear classes enjoy the claimed favorable reward-scale dependence. In particular, removing an explicit inverse-curvature factor does not by itself exclude such dependence through the complexity measures. This limits the demonstrated significance of the paper’s central general-function-approximation claim.

Questions for authors
  1. Can the concentration and counting arguments be corrected explicitly? In Appendix C.1, Lemma E.1 requires a bound on ∣ϕr−ϕˉr∣|\phi_r-\bar\phi_r|, but the proof takes b=2Bb=2B. For Δr∗=2B\Delta r^{*}=2B, Δr=−2B\Delta r=-2B, and o=0o=0, this centered quantity equals 4Bσ(2B)>2B4B\sigma(2B)>2B; b=4Bb=4B appears appropriate. Please also reconcile the historical budget 4βT4\beta_T with (B.5), and the disappearing factor of two before (B.8). Please provide a consistent confidence radius and corrected argument establishing the stated rates.

  2. Under what conditions does the general-class theorem imply the claimed saturation-insensitive dependence? The leading term contains deludτu(R)d_{\mathrm{elud}}^{\tau_u}(\mathcal R), and the nonlocalized contribution depends on log⁡NT\log N_T. Which conditions ensure that these quantities do not introduce exponential dependence on BB, or stronger-than-logarithmic dependence on TT? A representative nonlinear class with explicit complexity bounds would materially strengthen my assessment.

  3. Could the incremental contribution relative to the closest work be stated more precisely? Which ingredients beyond the asymmetric regret decomposition are new relative to Bakhtiari et al.’s localized-eluder analysis and existing local-curvature dueling methods? Please also make Section 2 consistent with Appendix F.1 regarding Chen et al.: SE-POPO provides final-policy latent-reward guarantees and a general-function extension, so the distinction should be stated in terms of the learning objective and guarantee, rather than suggesting that the prior work only controls preference regret.

  4. Could the authors provide a controlled synthetic comparison across reward scales and time horizons against a curvature-aware dueling baseline and an ablation of the comparator rule? Reporting cumulative two-arm regret, the distribution of queried gaps or link curvatures, and computational cost would help determine whether the asymptotic improvement translates into a practical advantage. Evaluating the approximate implementation from Appendix G would also clarify how much the result depends on exact optimization.

5 comments
AuthorsOpenReview ↗

Response to Reviewer Tm6w

Q1: No experiments; please compare across reward scales and horizons with a baseline and a comparator ablation, reporting two-arm regret, queried link curvatures and computational cost.

A1: We have added these experiments in Appendix H of the revision.

Setup. We evaluate a contextual dueling bandit with nonlinear rewards rθ(x,a)=Btanh⁡(x⊤Wa+∑jqjaj2)r_\theta(x,a)=B\tanh(x^\top W a+\sum_j q_j a_j^2) over a continuous action space for B∈{0.5,1,2,4}B\in\{0.5,1,2,4\} (1/σ′(2B)∈[5,3×103]1/\sigma'(2B)\in[5, 3\times 10^3]), T=5000T=5000 rounds, and 5 seeds. Since most existing contextual dueling bandit algorithms are restricted to linear rewards, we adopt the neural dueling bandit algorithms NDB-UCB and NDB-TS as baselines [1].

Cumulative two-arm regret at T=5000T=5000 (mean ±\pm 95% CI).

BBSI-CDBNDB-UCBNDB-TS
0.50.51496 ±\pm 5452198 ±\pm 10871859 ±\pm 824
11946 ±\pm 3011914 ±\pm 4481810 ±\pm 398
221339 ±\pm 2271860 ±\pm 3323891 ±\pm 5305
441117 ±\pm 41223106 ±\pm 999616232 ±\pm 7644

The regret of SI-CDB remains stable across BB and grows sublinearly in TT for every BB (Figure 2), showing that the additive 1/σ′(2B)1/\sigma'(2B)-dependent lower-order term does not dominate at this horizon. SI-CDB achieves the lowest regret across all BB, with a particularly large margin at B=4B=4.

Queried link curvatures. For SI-CDB, fewer than 1% of the queried pairs have σ′(Δr⋆)<0.2\sigma'(\Delta_{r^\star})<0.2 at every BB, so its queries stay away from saturation even at B=4B=4. The baselines behave similarly for B≤1B\le1, but at B=4B=4, 67% (NDB-UCB) and 50% (NDB-TS) of their queries have σ′(Δr∗)<0.1\sigma'(\Delta_{r^*})<0.1.

Computation: See A2.


Q2. Appendix G describes approximate optimization procedures, but provides neither empirical evaluation nor guarantees for their approximation errors. Evaluating the approximate implementation from Appendix G would clarify how much the result depends on exact optimization.

A2:

All our experiments use the approximate implementation (now Appendix H.1). We add an ablation that varies the accuracy of each approximation separately (Appendix H.4, Table 4), in the same setting (B∈{0.5,1,2,4}B\in\{0.5,1,2,4\}, T=5000T=5000, 5 seeds). For the reward-model update, we change the number of Adam steps per round from 8 to 2 and to 32. For the comparator problem, we change the number of (restarts, outer rounds) of the alternating procedure from (3, 8) to (1, 2) and to (6, 16). The table reports the cumulative two-arm regret (mean ± 95% confidence half-width) and the time per round on a single CPU core.

ApproximationB=0.5B=0.5B=1B=1B=2B=2B=4B=4Sec./round
2 Adam steps1692 ± 7371857 ± 27581529 ± 8551595 ± 7860.25
Default (8 Adam steps, (3, 8))1496 ± 545946 ± 3011339 ± 2271117 ± 4120.26
32 Adam steps1235 ± 6541820 ± 17721613 ± 7151937 ± 7830.29
Comparator (1, 2)1464 ± 6451341 ± 2051312 ± 5071879 ± 9840.04
Comparator (6, 16)1021 ± 1301299 ± 6741407 ± 4651268 ± 4890.95

For every BB, the differences between the rows are within the variation across seeds, and a more accurate reward-model update or comparator search does not improve the regret systematically. The regret remains insensitive to BB at every level of accuracy.

Although we do not have a guarantee for the approximation error and our theory assumes exact optimization, the ablation indicates that the empirical behavior does not rely on solving the two problems accurately.

AuthorsOpenReview ↗

Response to Reviewer Tm6w

Q3. The benefit beyond linear reward classes is insufficiently substantiated. The general theorem depends on localized eluder dimensions and metric entropy, but the paper explicitly bounds these quantities only for linear rewards. It therefore remains unclear which representative nonlinear classes enjoy the claimed favorable reward-scale dependence. In particular, removing an explicit inverse-curvature factor does not by itself exclude such dependence through the complexity measures.

A3. Thank you for this question. Theorem 5.3 yields a saturation-insensitive guarantee whenever the following two conditions hold:

  1. log⁡NT\log N_T grows polynomially in BB and logarithmically in TT. This holds, for example, for any class parameterized by a bounded dd-dimensional parameter with a Lipschitz parameterization.
  2. The localized eluder dimension deludτu(R)d^{\tau_u}_{\rm elud}(\mathcal R) at scale ϵ=1/T\epsilon=1/T grows polynomially in BB and polylogarithmically in TT. The quantity deludθu,θu(R)d^{\theta_u,\theta_u}_{\rm elud}(\mathcal R) enters only the TT-independent additive term.

To make the second condition concrete, we have added Appendix D.2 on generalized linear rewards rθ(x,a)=g(⟨θ,μ(x,a)⟩)r_\theta(x,a)=g(\langle\theta,\mu(x,a)\rangle) with polynomial link gg of degree pp, which need not be monotone. Lemma D.7 shows that the second condition holds for any convex class that is linear in some (possibly lifted) feature map, and Lemma D.6 shows that the localized eluder dimension is monotone in the class. Together, they give explicit bounds for representative nonlinear classes:

  • Complexity bounds. log⁡NT≤dlog⁡(1+8LgRT)\log N_T\le d\log(1+8L_gRT) (Proposition D.5) and deludϵ,η(Rg)≤C Dplog⁡(1+pSp2/ϵ)d^{\epsilon,\eta}_{\rm elud}(\mathcal R_g)\le C\,D_p\log(1+pS_p^2/\epsilon) with Dp=(d+pp)−1D_p=\binom{d+p}{p}-1 (Proposition D.8). Neither bound introduces exponential dependence on BB or more than logarithmic dependence on TT.
  • Regret. Corollary D.9 gives a leading term O~(BdDpT)\tilde O(\sqrt{BdD_pT}). For the quadratic link rθ(x,a)=⟨θ,μ(x,a)⟩2r_\theta(x,a)=\langle\theta,\mu(x,a)\rangle^2 (Example D.10), this is O~(d3/2BT)\tilde O(d^{3/2}\sqrt{BT}).
  • Further examples. Two-layer networks with quadratic activation and arbitrary width (Example D.11) and a contextual ideal-point preference model (Example D.12).

Appendix D.2 also notes that the refined analysis of Appendix E improves the leading term to O~(d Dp T)\tilde O(\sqrt{d\,D_p\,T}).


Q4. Can the concentration and counting arguments be corrected explicitly? In Appendix C.1, Lemma E.1 requires a bound on ∣ϕr−ϕˉr∣|\phi_r - \bar\phi_r|, but the proof takes b=2Bb = 2B. For Δr∗=2B\Delta_{r^*} = 2B, Δr=−2B\Delta_r = -2B, and o=0o = 0, this centered quantity equals 4Bσ(2B)>2B4B\sigma(2B) > 2B; b=4Bb = 4B appears appropriate. Please also reconcile the historical budget 4βT4\beta_T with (B.5), and the disappearing factor of two before (B.8). Please provide a consistent confidence radius and corrected argument establishing the stated rates.

A4. Thank you for checking the proof so carefully. All three points are correct. Each affects only numerical constants, and the stated rates are unchanged. We have made the following corrections in the revised manuscript (highlighted in blue).

  1. Range parameter in Lemma E.1 (Now Lemma F.1). Condition 1 of Lemma F.1 concerns the centered quantity ϕr−ϕˉr\phi_r-\bar\phi_r. Lemma B.5 gives ∣ϕr−ϕˉr∣≤4B|\phi_r-\bar\phi_r|\le 4B, so the correct choice is b=4Bb=4B, together with c=2B+4c=2B+4 from Lemma C.1. The confidence radius becomes βt=5Tϵ2+30(3B+2)log⁡(NϵhT/δ),\beta_t=\tfrac{5T\epsilon}{2}+30(3B+2)\log(N_\epsilon h_T/\delta), in place of 60(B+1)log⁡(NϵhT/δ)60(B+1)\log(N_\epsilon h_T/\delta). Its order, βT=O(Blog⁡(NT/δ))\beta_T=O(B\log(N_T/\delta)), is unchanged. We have also added the one-line derivation of ∣ϕr−ϕˉr∣≤4B|\phi_r-\bar\phi_r|\le 4B to the proof of Lemma B.5.
  2. Historical budget in (B.5). Lemma B.4 bounds the historical excess loss by 4βT4\beta_T, and this is the budget that enters Lemma F.2. The corrected (B.5), which also fixes the typo ϕ→ϕˉ\phi\to\bar\phi on the left-hand side, reads ∑l=1kϕˉrjtl(ztl)≤(2B+4βTlog⁡(1+2B/ϵ))deludϵ,τu(R)+2B+Kϵ,\sum_{l=1}^k\bar\phi_{r_j^{t_l}}(z_{t_l})\le\big(2B+4\beta_T\log(1+2B/\epsilon)\big)d_{\mathrm{elud}}^{\epsilon,\tau_u}(\mathcal R)+2B+K\epsilon, so Γu=(2B+4βTlog⁡(1+2BT))deludτu(R)+2B+1\Gamma_u=\big(2B+4\beta_T\log(1+2BT)\big)d_{\mathrm{elud}}^{\tau_u}(\mathcal R)+2B+1.
  3. Factor of two before (B.8). We have ∣ITc∣≤2(8βTu2σ′(2B)+1)deludθu,θu(R)+2|I_T^c|\le 2\big(\tfrac{8\beta_T}{u^2\sigma'(2B)}+1\big)d_{\mathrm{elud}}^{\theta_u,\theta_u}(\mathcal R)+2. With Rt≤4BR_t\le 4B, this gives (I2)≤8B(8βTu2σ′(2B)+1)deludθu,θu(R)+8B(I_2)\le 8B\big(\tfrac{8\beta_T}{u^2\sigma'(2B)}+1\big)d_{\mathrm{elud}}^{\theta_u,\theta_u}(\mathcal R)+8B.

With these corrections, Theorem 5.3 reads Regret(T)≤inf⁡u>0[32TΓuσ′(4u)+8B(8βTu2σ′(2B)+1)deludθu,θu(R)]+8B,\mathrm{Regret}(T)\le\inf_{u>0}\Big[3\sqrt{\tfrac{2T\Gamma_u}{\sigma'(4u)}}+8B\Big(\tfrac{8\beta_T}{u^2\sigma'(2B)}+1\Big)d_{\mathrm{elud}}^{\theta_u,\theta_u}(\mathcal R)\Big]+8B, with βT\beta_T and Γu\Gamma_u as above. The statement of Theorem 5.3 and Appendices B–C have been updated accordingly.

AuthorsOpenReview ↗

Response to Reviewer Tm6w

Q5. Could the incremental contribution relative to the closest work be stated more precisely? Which ingredients beyond the asymmetric regret decomposition are new relative to the localized-eluder analysis of [2] and existing local-curvature dueling methods [3]?

A5. Thank you for the question. We will add a paragraph to Section 6 that separates the imported tools from the new steps.

Relative to [2]. We import their uniform concentration (Lemma F.1), variance bound (Lemma C.1), localized-eluder counting lemma (Lemma F.2), and the witness rescaling behind Prop. D.2. Their regret-to-loss bridge does not transfer to dueling bandits, for two reasons:

  • The learner observes only the queried reward difference, while the regret involves the unqueried optimal action.
  • Regret is measured in latent rewards, so converting excess log loss into a reward-gap error costs 1/σ′(⋅)1/\sigma'(\cdot) at the queried gap, which can be eΘ(B)e^{\Theta(B)}.

Beyond the decomposition Rt≤e1,t+2e2,tR_t\le e_{1,t}+2e_{2,t} (Lemma B.2), the new ingredients are:

  1. Query-gap localization from two-armed regret. Since RtR_t penalizes both queried actions, ∣Δr∗(zt)∣≤Rt|\Delta_{r^*}(z_t)|\le R_t. On rounds where both errors are at most uu, all queried gaps are therefore at most 4u4u. Lemma B.3 then gives the two-sided bound σ′(4u)2ej,t2≤ϕˉrjt(zt)≤u2/8\tfrac{\sigma'(4u)}{2}e_{j,t}^2\le\bar\phi_{r_j^t}(z_t)\le u^2/8:

    • the left inequality converts loss into regret with local, rather than worst-case, curvature;
    • the right inequality places the losses at the scale τu\tau_u where the localized eluder dimension is evaluated.

    This localization takes place in query space and is produced by the comparator choice. Single-action bandits have no analogue, and it is the reason single-arm regret alone would not suffice.

  2. Two coupled estimators. The two errors come from different model sequences, the MLE and the optimistic model. We show that both have historical excess loss at most 4βT4\beta_T (Lemma B.4), so the localized counting applies to each. This is done at scale τu\tau_u on localized rounds and at scale θu\theta_u when counting the remaining rounds.

Relative to local-curvature dueling methods (RCDB-S [3]). These methods are specific to linear rewards. They estimate local link derivatives to build a derivative-weighted covariance matrix and rely on elliptical-potential arguments; for sigmoid feedback they obtain a leading term O~(dB3/2T)\tilde O(dB^{3/2}\sqrt T). Our analysis uses no covariance matrix and applies to general reward classes through the localized eluder dimension of the excess-loss class. Specialized to linear rewards, it gives the leading term O~(dBT)\tilde O(d\sqrt{BT}), which improves the BB-dependence from B3/2B^{3/2} to B\sqrt B. With the refined analysis of the new Appendix E, the leading term further improves to O~(dT)\tilde O(d\sqrt T) (Corollary E.3).


Q6. Please make Section 2 consistent with Appendix F.1 regarding [4]: SE-POPO provides final-policy latent-reward guarantees and a general-function extension, so the distinction should be stated in terms of the learning objective and guarantee, rather than suggesting that the prior work only controls preference regret.

A6. Thank you. As you note, SE-POPO uses preference-based regret as an intermediate tool and ultimately provides latent-reward guarantees for the final policy, with polynomial dependence on the reward scale and an extension to general function classes. The distinction lies in the learning objective. [4] bound the suboptimality of the learned policy. We bound the cumulative two-armed latent-reward regret of an online learner, in which both queried actions are penalized at every round, so exploratory comparisons themselves incur regret. We have replaced the sentence in Section 2 with:

[4] avoid exponential reward-scale dependence in online RLHF through preference-based exploration with an iteratively updated sampler. They obtain final-policy latent-reward guarantees with polynomial dependence on the reward scale, including for general function classes. Their objective is the suboptimality of the learned policy, whereas we study cumulative two-armed reward regret, in which both queried actions incur regret at every round. For this online objective, removing the exponential dependence from the leading regret term for general reward classes has remained open.

We have also aligned the wording of Appendix F.1 (now G.1) with this description. Both changes are highlighted in blue.


References

[1] Verma et al. Neural Dueling Bandits: Preference-Based Optimization with Human Feedback. 2025

[2] Alireza Bakhtiari, Alex Ayoub, Samuel Robertson, David Janz, and Csaba Szepesvári. Eluder dimension: Localise it! 2026.

[3] Qiwei Di, Jiafan He, and Quanquan Gu. Nearly optimal algorithms for contextual dueling bandits from adversarial feedback. 2025.

[4] Mingyu Chen, Yiding Chen, Wen Sun, and Xuezhou Zhang. Avoiding exp(R) scaling in RLHF through preference-based exploration. 2025

Reviewer_EL8POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies contextual dueling bandits with a general realizable reward class under the Bradley–Terry–Luce model, focusing on the saturation of the sigmoid link: when the reward gap of a queried pair is large, comparisons are nearly deterministic, the link derivative is exponentially small, and existing analyses pay a factor κ=1/σ′(2B)=eΘ(B)\kappa = 1/\sigma'(2B) = e^{\Theta(B)} in the leading regret term. The proposed algorithm, SI-CDB, picks a target action greedily under the maximum-likelihood reward estimate and a comparator action, together with an optimistic reward from a log-loss confidence set, by maximizing J(b,r)=2r(x,b)−r1(x,b)−2r(x,a)J(b, r) = 2r(x, b) - r_1(x, b) - 2r(x, a). The coefficients are chosen so that the two-arm regret is bounded by the reward-gap estimation errors on the queried pair. Combining this with the localized eluder dimension of Bakhtiari et al. (2026), the paper proves a regret bound that separates a T\sqrt{T} term depending on local curvature from an additive term that keeps the exponential factor. For linear rewards this gives O~(dBT+d2B2/σ′(2B)2)\tilde O\big(d\sqrt{BT} + d^2B^2/\sigma'(2B)^2\big), matching the known lower bound in d and T and improving the B-dependence over Di et al. (2025).

Critical strengths

(1) The paper addresses a real and well-identified gap: removing the exponential link-curvature factor from the leading regret term of dueling bandits. (2) The asymmetric selection rule and its decomposition (Lemma B.2) are simple and elegant. (3) The linear specialization is meaningful and well connected to prior work.

Critical weaknesses

(1) The technical contribution beyond Bakhtiari et al. (2026) seems to be thin. The confidence set, concentration inequality (Lemma E.1), localized eluder lemmas (Lemma E.2), variance bound (Lemma C.1) and the linear localized eluder bound (Prop. D.2) are imported, and the new part seems to be a few lines of algebra in Lemma B.2 plus the split into localized and non-localized rounds. The paper also lacks a lower bound showing that the B\sqrt{B} leading dependence or an additive exponential term is necessary. So my questions are: Beyond assembling the tools of Bakhtiari et al. (2026), what is the main technical challenge you overcome? Also, can you provide a lower bound on the B-dependence of the leading term, or show that an additive exponential term is unavoidable? (2) In the linear case, Corollary 5.5 appears to drop a factor of B coming from log⁡(1/θu)≈4B\log(1/\theta_u) \approx 4B in deludθu,θud_{\mathrm{elud}}^{\theta_u,\theta_u}, and could you recheck Corollary 5.5. (3) The algorithm is not computationally tractable in general (an ERM and a joint optimization over the confidence set per round), the practical procedure in Appendix G has no guarantee, and there are no experiments at all. The manuscript also has unfinished references ("Definition xx", "Lee et al. (2024); ?.").

Questions for authors

See my questions above. I am on the borderline side for this paper and would like to revise my score if my questions are well solved.

5 comments
AuthorsOpenReview ↗

Response to Reviewer EL8P

Thank you for your detailed review. We will address your concerns one by one.

Q1. The technical contribution beyond [1] seems to be thin. The confidence set, concentration inequality, localized eluder lemmas, variance bound and the linear localized eluder bound are imported, and the new part seems to be a few lines of algebra in Lemma B.2 plus the split into localized and non-localized rounds. Beyond assembling the tools of [1], what is the main technical challenge you overcome?

A1. Thank you for the careful reading. We agree that several probabilistic tools come from [1]. We will state this explicitly in the revision. These tools control how expected excess losses ϕˉr(zt)\bar\phi_r(z_t) accumulate along adaptively chosen queries. They do not, however, turn excess losses into latent-reward regret for dueling bandits without paying 1/σ′(2B)1/\sigma'(2B). This conversion is the main challenge we address.

Why the argument of [1] does not transfer. In their single-action setting, the learner observes a noisy version of the quantity that defines regret. Optimism gives ft(At)≤η(a⋆)f_t(A_t)\le\eta(a^\star), so the regret η(At)−η(a⋆)\eta(A_t)-\eta(a^\star) is at most the prediction error at the played action, and the triangle condition of the log loss bounds this error by ϕˉft(At)\bar\phi_{f_t}(A_t). In dueling bandits this bridge breaks in two places:

  1. The feedback ot∼Ber(σ(Δr⋆(zt)))o_t\sim\mathrm{Ber}(\sigma(\Delta_{r^\star}(z_t))) depends only on the reward difference of the queried pair, whereas Rt=2r⋆(xt,at⋆)−r⋆(xt,at)−r⋆(xt,bt)R_t=2r^\star(x_t,a_t^\star)-r^\star(x_t,a_t)-r^\star(x_t,b_t) involves the unqueried optimal action and both played actions.
  2. Regret is measured in latent rewards, but the log loss controls preference probabilities. Even at the queried pair, ϕˉr(z)=12σ′(ξ)(Δr(z)−Δr⋆(z))2\bar\phi_r(z)=\tfrac12\sigma'(\xi)\big(\Delta_r(z)-\Delta_{r^\star}(z)\big)^2 (Lemma B.3), so converting excess loss into a reward-gap error costs 1/σ′(ξ)=eΘ(B)1/\sigma'(\xi)=e^{\Theta(B)} whenever the queried comparison is saturated. Applying the triangle condition and then converting with the global curvature would put 1/σ′(2B)1/\sigma'(2B) back into the leading term, which is exactly the dependence we remove.

New ingredients.

  • (i) Queried-pair regret decomposition (Lemma B.2). With a greedy target under the MLE and an optimistic comparator selected jointly with a model in the log-loss confidence set, Rt≤e1,t+2e2,tR_t\le e_{1,t}+2e_{2,t}, where ej,t=∣Δrjt(zt)−Δr⋆(zt)∣e_{j,t}=|\Delta_{r_j^t}(z_t)-\Delta_{r^\star}(z_t)|. Both terms are reward-gap errors at the pair that was actually queried, which is what the loss history can control. No error at at∗a_t^* remains, and no error in absolute rewards, which BTL feedback cannot identify. Natural alternatives lack this property: optimism on individual arms leaves absolute-reward errors, and a fixed or reference comparator leaves a queried gap that may be saturated.
  • (ii) Query-gap localization from two-armed regret. Because RtR_t penalizes both queried actions, it bounds the true queried gap: ∣Δr⋆(zt)∣≤Rt|\Delta_{r^\star}(z_t)|\le R_t. Combined with (i), on rounds where both errors are at most uu, we get ∣Δr∗(zt)∣≤3u|\Delta_{r^*}(z_t)|\le 3u and ∣Δrjt(zt)∣≤4u|\Delta_{r_j^t}(z_t)|\le 4u, so the comparison is not saturated. Lemma B.3 then gives the two-sided bound σ′(4u)2ej,t2≤ϕˉrjt(zt)≤u2/8\tfrac{\sigma'(4u)}{2}e_{j,t}^2\le\bar\phi_{r_j^t}(z_t)\le u^2/8. The left inequality converts excess loss into regret through the local curvature σ′(4u)\sigma'(4u). The right inequality places these losses at the scale τu\tau_u on which the localized eluder dimension is defined. This localization takes place in the space of queries and is produced by the selection rule; single-action bandits have no counterpart, since there is no comparator whose choice controls the curvature of the observation. It also explains why single-arm regret is insufficient: an optimal ata_t paired with a poor comparator btb_t still produces a saturated query.
  • (iii) Two coupled estimators. The two errors come from different model sequences, the MLE r1tr_1^t and the optimistic r2tr_2^t. We show that both have historical excess loss at most 4βT4\beta_T (Lemma B.4), so the localized counting applies to each sequence: at scale τu\tau_u on the localized rounds, and at scale θu\theta_u when counting the non-localized rounds.

On "a few lines of algebra". The proof of Lemma B.2 is short because the selection rule was designed for it: the rule makes the per-round regret depend only on queried-pair gap errors of models in the confidence set, which are the only quantities the localized machinery can control. Steps (ii)–(iii) then turn this into a leading term O~(dBT)\tilde O(d\sqrt{BT}) in the linear case. With the refined analysis of the new Appendix E, it further improves to O~(dT)\tilde O(d\sqrt T) (Corollary E.3). Both improve the O~(dB3/2T)\tilde O(dB^{3/2}\sqrt{T}) of the local-curvature method of [2], and our analysis extends to general reward classes, where matrix-based local-curvature arguments are unavailable.

AuthorsOpenReview ↗

Response to Reviewer EL8P

Q2. The paper lacks a lower bound showing that the B\sqrt{B} leading dependence or an additive exponential term is necessary. Can you provide a lower bound on the BB-dependence of the leading term, or show that an additive exponential term is unavoidable?

A2. Thank you for this question. We address the two parts separately.

The B\sqrt B factor in the leading term is not necessary. In the new Appendix E, we tighten the confidence radius to βt=log⁡(2NT/δ)+2\beta_t=\log(2N_T/\delta)+2, which no longer depends on BB, and analyze it through the squared Hellinger distance (Theorem E.1). For linear rewards, Corollary E.3 gives Regret(T)≤O~(dT+d2B2σ′(2B)).\mathrm{Regret}(T)\le\tilde O\Big(d\sqrt T+\frac{d^2B^2}{\sigma'(2B)}\Big). The leading term has no polynomial dependence on BB and matches the Ω(dT)\Omega(d\sqrt T) lower bound of [3] in dd and TT. Consequently, our new regret guarantee matches the lower bound in [3], which indicates that it cannot be further improved.

The additive exponential term. Corollary E.3 also improves the additive term from d2B3/[σ′(2B)]2d^2B^3/[\sigma'(2B)]^2 to d2B2/σ′(2B)d^2B^2/\sigma'(2B), that is, from poly(B) e4B\mathrm{poly}(B)\ e^{4B} to poly(B) e2B\mathrm{poly}(B)\,e^{2B}. Whether this term can be removed is open. A comparable term also appears in the logistic-bandit literature, and a lower bound is difficult because comparisons between the candidate optimal actions of two hypotheses are never saturated. We discuss both points in detail in our response to Reviewer N2kk (Q1).


Q3. In the linear case, Corollary 5.5 appears to drop a factor of BB coming from log⁡(1/θu)≈4B\log(1/\theta_u) \approx 4B in deludθu,θud_{\mathrm{elud}}^{\theta_u,\theta_u}. Could you recheck Corollary 5.5?

A3. Thank you for catching this; you are right. Proposition 5.4 gives deludϵ,η(R)≤Cdlog⁡(1+B2/ϵ)d_{\mathrm{elud}}^{\epsilon,\eta}(\mathcal R)\le Cd\log(1+B^2/\epsilon), and with u=σ′(2B)u=\sqrt{\sigma'(2B)} we evaluate it at ϵ=θu=[σ′(2B)]2/2≈e−4B/2\epsilon=\theta_u=[\sigma'(2B)]^2/2\approx e^{-4B}/2. Then log⁡(1/θu)=4B+O(log⁡B)\log(1/\theta_u)=4B+O(\log B), which is linear in BB and should not be absorbed into O~(⋅)\tilde O(\cdot). Hence deludθu,θu(R)=O(dB)d_{\mathrm{elud}}^{\theta_u,\theta_u}(\mathcal R)=O(dB), and the corrected bound is Regret(T)≤O~ ⁣(dBT+d2B3[σ′(2B)]2).\mathrm{Regret}(T)\le\tilde O\!\left(d\sqrt{BT}+\frac{d^2B^3}{[\sigma'(2B)]^2}\right). The leading term is unaffected: there the localized dimension is evaluated at ϵ=1/T\epsilon=1/T, which contributes only log⁡(BT)\log(BT). We have corrected Corollary 5.5 and added this explanation before it in the revised manuscript (changes are highlighted in blue). In addition, the refined analysis in the new Appendix E (Corollary E.3) improves this bound to O~(dT+d2B2/σ′(2B))\tilde O\big(d\sqrt T+d^2B^2/\sigma'(2B)\big); see also our response to Reviewer N2kk (Q1).


Q4. The algorithm is not computationally tractable in general (an ERM and a joint optimization over the confidence set per round), and the practical procedure in Appendix G has no guarantee.

A4. We agree that SI-CDB is an information-theoretic algorithm stated in terms of optimization oracles. This is also the case for most optimism-based algorithms with general function approximation, e.g., eluder-based UCB [4], GOLF [5], and the ℓ\ell-UCB algorithm of [1], all of which optimize over a version space. We add three remarks.

  1. The linear case is tractable. For rθ(x,a)=⟨μ(x,a),θ⟩r_{\theta}(x,a)=\langle\mu(x,a),\theta\rangle, LtL_t is a convex logistic loss, so the ERM is a convex program and Conft\mathrm{Conf}_t is a convex sublevel set. For each fixed bb, the comparator objective 2rθ(xt,b)−r1t(xt,b)−2rθ(xt,at)2r_\theta(x_t,b)-r_1^t(x_t,b)-2r_\theta(x_t,a_t) is linear in θ\theta. The joint step therefore reduces to ∣A∣|\mathcal A| convex programs per round, one for each candidate bb.

  2. Approximate oracles. The analysis is robust to optimization error. Suppose the ERM is εL\varepsilon_{\mathrm{L}}-approximate in loss, the target is εa\varepsilon_{\mathrm a}-approximate, and the comparator pair is feasible and εc\varepsilon_{\mathrm c}-approximate. Then r⋆∈Conftr^\star\in\mathrm{Conf}_t still holds, Lemma B.4 holds with βt\beta_t replaced by βt+εL/2\beta_t+\varepsilon_{\mathrm L}/2, and Lemma B.2 becomes Rt≤e1,t+2e2,t+εa+εcR_t\le e_{1,t}+2e_{2,t}+\varepsilon_{\mathrm a}+\varepsilon_{\mathrm c}. The rest of the proof goes through with minor changes to constants, adding T(εa+εc)T(\varepsilon_{\mathrm a}+\varepsilon_{\mathrm c}) to the regret. Optimization errors of order 1/T1/\sqrt T therefore preserve the rate. The procedure in Appendix H.1 enforces feasibility explicitly by checking the confidence constraint, so it is covered whenever its alternating updates reach this accuracy.

  3. Scope of Appendix G (Now Appendix H.1). We do not claim that the alternating procedure in Appendix H.1 always attains this accuracy for nonconvex reward models. We will state clearly that Appendix H.1 is a practical heuristic, and add the approximate-oracle statement above as a remark.

AuthorsOpenReview ↗

Response to Reviewer EL8P

Q5. There are no experiments.

A5. Thank you for raising this. We have added numerical experiments in Appendix H using the practical implementation in Appendix H.1.

Setup. We evaluate a contextual dueling bandit with nonlinear rewards rθ(x,a)=Btanh⁡(x⊤Wa+∑jqjaj2)r_\theta(x,a)=B\tanh(x^\top W a+\sum_j q_j a_j^2) over a continuous action space for B∈{0.5,1,2,4}B\in\{0.5,1,2,4\} (1/σ′(2B)∈[5,3×103]1/\sigma'(2B)\in[5, 3\times 10^3]), T=5000T=5000 rounds, and 5 seeds. Since most existing contextual dueling bandit algorithms are restricted to linear rewards, we adopt the neural dueling bandit algorithms NDB-UCB and NDB-TS as baselines [6].

Results (Cumulative two-armed regret, mean ±\pm 95% CI).

BBSI-CDBNDB-UCBNDB-TS
0.50.51496 ±\pm 5452198 ±\pm 10871859 ±\pm 824
11946 ±\pm 3011914 ±\pm 4481810 ±\pm 398
221339 ±\pm 2271860 ±\pm 3323891 ±\pm 5305
441117 ±\pm 41223106 ±\pm 999616232 ±\pm 7644

Consistent with our theory, the regret of SI-CDB remains stable across BB because its leading term is free of 1/σ′(2B)1/\sigma'(2B). SI-CDB achieves the lowest regret across all scales.


Q6. The manuscript has unfinished references.

A6. Thank you for pointing these out. We have fixed both unfinished references in the revised manuscript (changes are highlighted in blue), and we will carefully proofread the rest of the paper.


References

[1] Alireza Bakhtiari, Alex Ayoub, Samuel Robertson, David Janz, and Csaba Szepesvári. Eluder dimension: Localise it! 2026.

[2] Qiwei Di, Jiafan He, and Quanquan Gu. Nearly optimal algorithms for contextual dueling bandits from adversarial feedback. 2025.

[3] Xuheng Li, Heyang Zhao, and Quanquan Gu. Feel-good Thompson sampling for contextual dueling bandits. 2024.

[4] Daniel Russo and Benjamin Van Roy. Eluder dimension and the sample complexity of optimistic exploration. 2013.

[5] Chi Jin, Qinghua Liu, and Sobhan Miryoosefi. Bellman eluder dimension: New rich classes of RL problems, and sample-efficient algorithms. 2021.

[6] Verma et al. Neural Dueling Bandits: Preference-Based Optimization with Human Feedback. 2025

Reviewer_N2kkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies contextual dueling bandits with general function approximation under the BTL preference model. The authors propose SI-CDB, which uses a new opponent-arm selection strategy to reduce the effect of preference saturation on reward learning. Using a localized Eluder dimension analysis, they establish regret guarantees that avoid the unfavorable inverse-derivative dependence in the leading regret term and recover near-optimal rates for linear reward models.

Critical strengths
  1. The asymmetric arm-selection strategy provides a meaningful technical contribution.

  2. The paper extends saturation-insensitive regret analysis to general function approximation.

Critical weaknesses
  1. The regret guarantee still suffers from an exponentially large additive dependence on the reward scale. As shown in Corollary 5.5, when BB is large, this additive term may dominate the regret upper bound. Although the paper removes the worst-case inverse-curvature factor from the leading T\sqrt{T} term, it does not establish a regret bound with polynomial dependence on BB overall.

  2. The complexity measure allows for general function approximation, but the localized Eluder dimension itself may depend on the true reward function and the reward range. Without further control over these complexity measures, their dependence on BB may not be well behaved. In my view, Theorem 5.3 therefore does not automatically imply a regret bound with polynomial dependence on BB for arbitrary reward function classes.

Questions for authors
  1. The regret bound in Corollary 5.5 still contains an additive term with exponential dependence on the reward scale BB. Can the authors further improve this dependence, or establish a lower bound showing that such an exponential additive term is unavoidable?

  2. Could the authors provide explicit bounds on the localized Eluder dimension for nontrivial nonlinear reward function classes beyond the linear setting? I understand that this may be challenging, but it would be helpful to at least provide a concrete example of a nonlinear reward function class for which Theorem 5.3 gives a meaningful regret guarantee.

4 comments
AuthorsOpenReview ↗

Response to Reviewer N2kk

Q1. The regret guarantee still suffers from an exponentially large additive dependence on the reward scale. As shown in Corollary 5.5, when BB is large, this additive term may dominate the regret upper bound. Although the paper removes the worst-case inverse-curvature factor from the leading T\sqrt{T} term, it does not establish a regret bound with polynomial dependence on BB overall. Can the authors further improve this dependence, or establish a lower bound showing that such an exponential additive term is unavoidable?

A1. Thank you for raising this point. Our answer has three parts: (i) such an additive term also appears in prior work, (ii) we have improved both terms of the bound, and (iii) removing the additive term entirely, or proving that it is necessary, appears to be hard.

(i) Prior work has the same type of term. Even for single-parameter logistic bandits, the algorithms that remove the worst-case inverse curvature κ\kappa from the leading term keep an additive lower-order term that scales with κ\kappa. For example, Logistic-UCB-2 of [1] achieves O~(dT+κd2)\tilde O(d\sqrt T+\kappa d^2), and the refinements of [2], [3] and [4] also keep a κ\kappa-dependent lower-order term. In our setting, the counterpart of κ\kappa is 1/σ′(2B)=eΘ(B)1/\sigma'(2B)=e^{\Theta(B)}. For contextual dueling bandits, the guarantees of [5], [6] and [7] are worse: the exponential factor multiplies the poly(T)\mathrm{poly}(T) term. Our bound moves this cost into a low-order additive term, which matches what is currently known in the simpler logistic-bandit setting.

(ii) Improved dependence on BB. We have added a new Appendix E, "Optimizing the Dependence on the Reward Scale BB". It keeps the algorithm unchanged and only tightens the confidence radius to βt=log⁡(2NT/δ)+2\beta_t=\log(2N_T/\delta)+2, which no longer depends on BB. The analysis replaces the excess log loss by the squared Hellinger distance, which is bounded in [0,1][0,1]. Theorem E.1 gives the resulting guarantee for general reward classes, and Corollary E.3 specializes it to linear rewards: Regret(T)≤O~(dT+d2B2σ′(2B)).\mathrm{Regret}(T)\le\tilde O\Big(d\sqrt T+\frac{d^2B^2}{\sigma'(2B)}\Big). Compared with Corollary 5.5, this removes the B\sqrt B factor from the leading term and improves the additive term from d2B3/[σ′(2B)]2≈poly(B)e4Bd^2B^3/[\sigma'(2B)]^2\approx \mathrm{poly}(B)e^{4B} to d2B2/σ′(2B)≈poly(B) e2Bd^2B^2/\sigma'(2B)\approx\mathrm{poly}(B)\,e^{2B}.

(iii) Why further improvement or a lower bound is hard.

  • Improvement. The remaining 1/σ′(2B)1/\sigma'(2B) comes from rounds in which a model in the confidence set is far from r∗r^* at a saturated query. Such a comparison carries only O(σ′(2B))O(\sigma'(2B)) information. Removing the factor would require showing that the algorithm rarely makes such queries, which the eluder-type counting cannot express.
  • Lower bound. Standard two-point constructions do not apply. To force an Ω(e2B)\Omega(e^{2B}) sample complexity between two hypotheses r1r_1 and r2r_2, every informative query (a,b)(a,b) would need to be saturated in the same tail under both r1r_1 and r2r_2 (e.g., Δr1,Δr2≈2B\Delta_{r_1}, \Delta_{r_2} \approx 2B). However, if r1r_1 and r2r_2 have different optimal actions a1∗≠a2∗a_1^{\ast} \neq a_2^{\ast} in some context xx, the learner can directly query the pair (a1∗,a2∗)(a_1^{\ast}, a_2^{\ast}): by optimality, Δr1(x,a1∗,a2∗)>0\Delta _{r_1}(x, a_1^{\ast}, a_2^{\ast}) > 0 whereas Δr2(x,a1∗,a2∗)<0\Delta_{r_2}(x, a_1^{\ast}, a_2 ^{\ast}) < 0. Because the two gaps have opposite signs (σ(Δr1)>1/2>σ(Δr2)\sigma(\Delta_{r_1}) > 1/2 > \sigma(\Delta_{r_2})), this comparison does not lie in a saturated tail, allowing the learner to distinguish r1r_1 and r2r_2 at poly(B)\mathrm{poly}(B) cost.

We have noted this as future work in Appendix E.

AuthorsOpenReview ↗

Response to Reviewer N2kk

Q2. The complexity measure allows for general function approximation, but the localized Eluder dimension itself may depend on the true reward function and the reward range. Without further control over these complexity measures, their dependence on BB may not be well behaved. In my view, Theorem 5.3 therefore does not automatically imply a regret bound with polynomial dependence on BB for arbitrary reward function classes. Could the authors provide explicit bounds on the localized Eluder dimension for nontrivial nonlinear reward function classes beyond the linear setting? I understand that this may be challenging, but it would be helpful to at least provide a concrete example of a nonlinear reward function class for which Theorem 5.3 gives a meaningful regret guarantee.

A2. Thank you for this question. We agree that Theorem 5.3 by itself does not give a polynomial dependence on BB for an arbitrary class. Following your suggestion, we have extended Appendix D, now titled "Dueling Bandits with (Generalized) Linear Reward", with a new Appendix D.2 on generalized linear reward classes rθ(x,a)=g(⟨θ,μ(x,a)⟩),g(s)=∑k=1pcksk,r_\theta(x,a)=g(\langle\theta,\mu(x,a)\rangle),\qquad g(s)=\textstyle\sum_{k=1}^pc_ks^k, where gg is a known polynomial link of degree pp that need not be monotone. For such classes, the preference probability σ(g(⟨θ,μ(x,a)⟩)−g(⟨θ,μ(x,b)⟩))\sigma\big(g(\langle\theta,\mu(x,a)\rangle)-g(\langle\theta,\mu(x,b)\rangle)\big) is in general not a generalized linear model in θ\theta, and the log-likelihood is nonconvex. Appendix D.2 follows the structure of the linear case (Appendix D.1) and contains:

  • Covering number (Proposition D.5). log⁡NT≤dlog⁡(1+8LgRT)\log N_T\le d\log(1+8L_gRT), computed for the original dd-dimensional parameter.
  • Localized eluder dimension (Proposition D.8). For 0<ϵ≤η≤σ′(2B)/40<\epsilon\le\eta\le\sigma'(2B)/4, deludϵ,η(Rg)≤C Dplog⁡(1+pSp2/ϵ)d^{\epsilon,\eta}_{\rm elud}(\mathcal R_g)\le C\,D_p\log(1+pS_p^2/\epsilon), where Dp=(d+pp)−1D_p=\binom{d+p}{p}-1 is the dimension of a lifted feature space and Sp2=∑kck2R2kS_p^2=\sum_kc_k^2R^{2k}. The bound holds uniformly over r∗∈Rgr^*\in\mathcal R_g and depends on BB only logarithmically. Two supporting lemmas are of independent interest: the localized eluder dimension is monotone in the reward class (Lemma D.6), and the linear bound holds for any convex class that is linear in some feature map (Lemma D.7).
  • Regret (Corollary D.9). Regret(T)≤O~(BdDpT+d Dp B2(B+log⁡(1+pSp2))/[σ′(2B)]2)\mathrm{Regret}(T)\le\tilde O\big(\sqrt{BdD_pT}+d\,D_p\,B^2(B+\log(1+pS_p^2))/[\sigma'(2B)]^2\big). With the refined analysis of Appendix E, the leading term improves to O~(d Dp T)\tilde O(\sqrt{d\,D_p\,T}).

Appendix D.2 also gives three concrete nonlinear examples:

  1. Quadratic link (Example D.10). For rθ(x,a)=⟨θ,μ(x,a)⟩2r_\theta(x,a)=\langle\theta,\mu(x,a)\rangle^2, we have D2=d(d+1)/2D_2=d(d+1)/2 and Regret(T)≤O~(d3/2BT+d3B3/[σ′(2B)]2)\mathrm{Regret}(T)\le\tilde O\big(d^{3/2}\sqrt{BT}+d^3B^3/[\sigma'(2B)]^2\big). Here the reward is not monotone in the index ⟨θ,μ(x,a)⟩\langle\theta,\mu(x,a)\rangle.
  2. Two-layer networks with quadratic activation (Example D.11). For r(x,a)=∑k=1mvk⟨wk,μ(x,a)⟩2r(x,a)=\sum_{k=1}^mv_k\langle w_k,\mu(x,a)\rangle^2 with ∑k∣vk∣∥wk∥22≤B\sum_k|v_k|\|w_k\|_2^2\le B, the localized eluder dimension is O(d2log⁡(B2/ϵ))O(d^2\log(B^2/\epsilon)), independently of the width mm.
  3. Contextual ideal-point model (Example D.12). For r(x,a)=−∥A(μ(a)−Cx)∥22r(x,a)=-\|A(\mu(a)-Cx)\|_2^2, preferences are single-peaked around a context-dependent ideal point, and the same bound holds with D=d(d+1)/2+d dxD=d(d+1)/2+d\,d_x.

References

[1] Louis Faury, Marc Abeille, Clément Calauzènes, and Olivier Fercoq. Improved optimistic algorithms for logistic bandits. 2020.

[2] Marc Abeille, Louis Faury, and Clément Calauzènes. Instance-wise minimax-optimal algorithms for logistic bandits. 2021.

[3] Louis Faury, Marc Abeille, Kwang-Sung Jun, and Clément Calauzènes. Jointly efficient and optimal algorithms for logistic bandits. 2022.

[4] Junghyun Lee, Se-Young Yun, and Kwang-Sung Jun. A unified confidence sequence for generalized linear models, with applications to bandits. 2024.

[5] Aadirupa Saha. Optimal algorithms for stochastic contextual preference bandits. 2021.

[6] Viktor Bengs, Aadirupa Saha, and Eyke Hüllermeier. Stochastic contextual dueling bandits under linear stochastic transitivity models. 2022.

[7] Xuheng Li, Heyang Zhao, and Quanquan Gu. Feel-good Thompson sampling for contextual dueling bandits. 2024.

Thanks for the authors’ response. Their response has largely addressed my concerns. Although the issue of completely eliminating the dependence on the BB-index remains unresolved, the authors have clarified the underlying challenges. Therefore, I will maintain my original positive score.

Thank you for your constructive feedback and support!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.