acceptodds
Under review as a conference paper at ICLR 2027

Beyond Sub-Gaussian Detector Scores: Robust Weighted Profile-Loss Change Point Detection for Human-LLM Text Segmentation

Abstract

Mixed human-LLM documents require locating authorship transitions from detector scores whose reliability varies across text units. Existing weighted mean contrasts are vulnerable to extreme scores, while directly replacing means with robust centers obscures how a misplaced boundary changes the population objective.We propose Robust Weighted Profile-Loss Change Point Detection (RWCP), which combines capped reliability weights, Huber profile gains, and narrowest-over-threshold search in reliability coordinates.Our key analysis expresses the population gap between a true and a displaced split as a merge cost, avoiding a closed-form solution for the nonlinear center of a mixed segment.Under explicit curvature, spacing, and dependence conditions, core RWCP recovers the number of changes and localizes their boundaries; its quadratic-loss limit recovers squared weighted CUSUM.We also study RWCP-R, a separately evaluated decoder that shares source centers across nonadjacent passages.Across five retrospective cached-score benchmark families, core RWCP reduces family-macro WindowDiff by 17.6% relative to weighted change-point detection, and RWCP-R lowers it further.Boundary recovery improves most clearly for isolated changes, while both fixed configurations miss changes in collaborative and densely alternating text.

Primary area:
learning on time series and dynamical systems
Keywords:
Human–LLM Text Segmentation, Robust Change Point Detection, Huber Profile Loss, Reliability Weighting, Authorship Localization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nArgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RWCP for detecting authorship transitions in mixed human–LLM text. It replaces WCP’s weighted mean-based contrast with a Huber profile-loss gain, combined with capped reliability weights and reliability-coordinate interval search. A merge-cost analysis establishes theoretical guarantees under certain assumptions. The paper also evaluates RWCP-R, a separate recurrent-state decoder that shares source centers across nonadjacent passages.

Critical strengths

The paper provides a useful way to analyze robust profile gains without explicitly solving for mixed-segment Huber centers. In Theorem 3.11 establishes finite-sample count-recovery and localization guarantees under explicit conditions in Assumption 3.1-3.6. The experimental results cover five benchmark families and show consistent improvements in WindowDiff (WD) metric. Core RWCP reduces family-macro WP by 17.6% compared with WCP, while RWCP-R achieves further improvements.

Critical weaknesses
  1. There are inconsistencies in the assumptions throughout that substantially weaken the contributions in the paper: W1: For example, how are the Assumptions 3.1-3.6, consistent with WCP? Why is the recurrent decoder in Tables 4 and 17 different from RWCP? Due to such differences, the actual effect of the proposed method cannot be evaluated.

  2. There are several weaknesses in the experimental results particularly in some important regimes. One is the boundary recovery problem. Appendix B.2, Table 6 shows that core RWCP detects no boundaries in 82.4% of CoAuthor documents and 93.1% of multiple-change documents, with Loc. cov. of 0% and 1.0% respectively. This issue persists for RWCP-R. Appendix B.7.2 Table 16 reports an 84.1% no-boundary fraction on CoAuthor, where its WD (0.4869) is only slightly better than the baseline (0.4893). Moreover, Table 17 shows that replacing Huber loss with quadratic loss improves count MAE and boundary F1, and only moderately lower WD (0.3110 to 0.3179). These results raise practical concerns about the porposed work.

Questions for authors
  1. Could the authors directly compare RWCP’s assumptions and localization guarantees with those of WCP both theoretically and experimentally?

  2. Could the author provide a comparison about (WD)-(boundary F1) trade-offs between RWCP and WCP under various detection thresholds or transition penalties? This would be helpful to clarify whether the method actually improves boundary detection, or primarily achieves lower WD by predicting fewer boundaries.

3 comments

Thank you very much for your support and valuable comments. We have carefully reviewed them, which are very helpful. We summarize the key concerns you raised, supplemented the corresponding experiments, and provide our responses as follows.

1. Assumptions and guarantees

Assumptions 3.1–3.6 extend WCP's setting rather than restating it. WCP assumes independent heteroscedastic sub-Gaussian scores with source-specific means and inverse squared proxy weights. Our theorem targets segment-wise Huber locations under local curvature and concentration of bounded transforms, which covers heavy-tailed scores and an explicit dependence condition.

AssumptionRole in the Core analysisRelation to WCP
3.1 Robust locationsFinite Huber risks, one Huber minimizer per segment, interior parameterReplaces the mean; no sub-Gaussian tails needed
3.2 Local curvatureQuadratic risk bounds turn mixed assignments into a separation costAutomatic for quadratic loss; a distributional condition for Huber
3.3 ReliabilityBounded positive weights, wmax⁡≤Δw/24w_{\max}\le\Delta_w/24Inverse-proxy weighting with clipping against dominant units
3.4 DependenceBernstein control of bounded transformsAllows dependent scores (WCP: independence)
3.5 Central massEnough mass in the quadratic region; κmax⁡≤p0r0/16\kappa_{\max}\le p_0r_0/16Local regularity for the robust target
3.6 Spacing and signalκmin⁡2Δw≳ΛN(δ)\kappa_{\min}^2\Delta_w\gtrsim\Lambda_N(\delta), feasible side and guard massesSame signal-spacing principle

Corollary 3.12 transfers the guarantee to estimated weights that are uniformly comparable to oracle weights.

On rates: WCP gives exact-count recovery and weighted localization of order log⁡(N/δ)/κ2\log(N/\delta)/\kappa^2 with κ\kappa a mean gap. Theorem 3.11 gives, on one event of probability at least 1−δ1-\delta, K^=K\widehat K=K and dw(τ^j,τj)≤CΛN(δ)/κj2d_w(\widehat\tau_j,\tau_j)\le C\Lambda_N(\delta)/\kappa_j^2, with ΛN(δ)=qN2log⁡{64(1+N2+N3)/δ}\Lambda_N(\delta)=q_N^2\log\{64(1+N^2+N^3)/\delta\}, κj\kappa_j a Huber-location jump, qN=O(1)q_N=O(1) under independence or fixed-order dependence and O(log⁡2(eN))O(\log^2(eN)) under geometric mixing. When the two targets coincide, the independent-case orders match, so robustness and dependence come at no cost in rate. Proposition 2.2 shows that the quadratic profile gain equals squared weighted CUSUM, so WCP's contrast is the quadratic limit of ours. Theory-aligned simulations will be added.

2. Relationship between Core and RWCP-R

RWCP-R is a separate decoder built on the same weighted Huber loss. It shares source-state centers across nonadjacent passages, fits a scale, and solves a transition-penalized objective by alternating exact decoding and parameter updates; this is why the decoder in Tables 4 and 17 differs from Core. Since RWCP-R carries the empirical results, the revision is organized around it.

MethodMacro WD ↓Count MAE ↓F1 ↑
No boundary0.36223.4330.009
WCP, default0.40702.8080.287
Core RWCP, default0.33553.2290.129
RWCP-R, frozen0.31103.0190.230

Averages are taken within settings, families, then over five families; F1 uses a one-sentence tolerance. RWCP-R has the lowest WD of all methods. Against null it gives ΔWD = −0.0511 [−0.0588, −0.0432], ΔMAE = −0.414 [−0.444, −0.383], ΔF1 = +0.221 [0.203, 0.240] (95% paired source-component bootstrap), so it is clearly better than predicting nothing on all three metrics.

3. Boundary recovery

FamilyZero CP CoreZero CP RWCP-RZero CP WCPExact-count WCPCoreRWCP-R
Single change31.3%31.7%12.3%30.9%61.4%65.8%
CoAuthor82.4%84.1%15.9%4.6%0.0%0.0%
Multiple changes93.1%63.9%19.6%19.5%1.0%8.4%
Text attacks94.7%54.5%32.7%34.9%5.4%40.9%
LLM proportion92.8%66.8%27.8%22.0%6.4%26.6%

Exact-count coverage is the share of documents with changes where K^=K\widehat K=K. RWCP-R recovers the exact number of changes more often than WCP on single changes (65.8% vs 30.9%), text attacks (40.9% vs 34.9%) and LLM proportion (26.6% vs 22.0%), and raises coverage over Core in every family except CoAuthor. Core's conservative behavior comes from its analytic threshold: the first-step maximum gain is below it on 82.4% of CoAuthor, 93.1% of multiple-change, 94.6% of attack and 92.8% of proportion documents. RWCP-R replaces this threshold with an explicit penalty, which reduces the no-boundary rate by 26–40 points on multiple changes, attacks and proportions.

3. Boundary recovery (continued): CoAuthor, localization and Table 17

CoAuthor sessions average 24 sentences and 10.4 changes, 62% of runs are one sentence, the sentence-source AUC is 0.68 and the adjacent-difference boundary AUC is 0.53. Under the frozen setting RWCP-R has the lowest CoAuthor WD (0.4869, null 0.4893). To compare boundary recovery, both methods were tuned on the 289 validation sessions by WD alone or by 0.5 WD+0.25∣K−K^∣/(K+1)+0.25(1−F1)0.5\,\mathrm{WD}+0.25|K-\widehat K|/(K+1)+0.25(1-\mathrm{F1}) and tested once on 290 sessions:

CoAuthorWD ↓MAE ↓F1 ↑Zero CP
WCP, default0.4887.920.30215.9%
WCP, composite0.5246.350.4401.7%
WCP, WD-only0.4899.860.09870.7%
RWCP-R, frozen0.48710.190.05884.1%
RWCP-R, composite0.5326.010.4720.3%
RWCP-R, WD-only0.49010.190.05684.1%
No boundary0.48910.380.031100.0%

With composite selection RWCP-R has the best F1 and the lowest count MAE among the configurations in this table: RWCP-R minus WCP is ΔF1 = +0.032 [0.005, 0.058], ΔMAE = −0.34 [−0.74, 0.04], ΔWD = +0.008 [−0.010, 0.025] (session bootstrap, 2,000 draws). With WD-only selection the difference in F1 is −0.043 [−0.064, −0.022].

Mean absolute boundary error (sentences) on documents where both methods get the count right:

FamilyShared docsWCP ↓RWCP-R ↓
Single change700.6860.743
CoAuthor0--
Multiple changes342.8403.081
Text attacks1361.8161.654
LLM proportion583.8974.121

Once the count is right, both methods place boundaries with similar precision (within 0.3 sentences, RWCP-R ahead on attacks); RWCP-R's advantage is that it gets the count right far more often.

Table 17 ablations:

VariantMacro WD ↓MAE ↓F1 ↑Zero CP
Full, frozen0.31103.0190.23060.2%
Quadratic loss0.31792.9450.23954.9%
Uniform weights0.31182.9170.26151.0%
Fixed scale0.31703.0710.21164.0%
No recurrence0.35843.0720.23049.5%
No penalty0.729910.6490.3330.3%

The full model has the lowest WD. Huber loss is a safeguard against extreme scores: with 5% extreme scores, removing it raises WD by 0.056, versus 0.007 on original scores. With 10% contamination and 20-fold proxy inflation, removing Huber changes WD by −0.022 and removing weights by −0.148; the revision discusses proxy quality accordingly.

4. Threshold and penalty trade-offs

Each method varies one detection control on the same scores: WCP the AIC multiplier rr and NOT threshold mlog⁡Nm\log N (CoAuthor), with (1,1)(1,1) reproducing all 2,684 saved predictions; Core a threshold multiplier; RWCP-R the penalty λ.

SettingWD ↓MAE ↓F1 ↑Zero CP
WCP r/m 0.25/0.250.71576.4460.3541.8%
WCP 0.5/0.50.59804.0170.3494.0%
WCP 1/1 (default)0.40702.8080.28721.7%
WCP 2/1.50.33163.0820.17265.2%
WCP 4/20.33853.3180.07789.3%
WCP 8/30.35673.4160.02098.4%
WCP 16/40.35953.4310.01299.7%
Core ×0.10.59933.3470.3283.8%
Core ×0.250.45672.6570.30513.6%
Core ×0.50.36222.8700.23040.3%
Core ×0.750.34043.1100.16864.7%
Core ×1 (default)0.33553.2290.12978.9%
Core ×1.50.34073.3210.08788.8%
Core ×20.34613.3700.06193.8%
RWCP-R λ=0.250.69839.0620.3440.3%
RWCP-R λ=0.50.66547.6690.3560.3%
RWCP-R λ=10.57254.9510.3820.6%
RWCP-R λ=20.37002.5100.38312.5%
RWCP-R λ=4 (frozen)0.31103.0190.23060.2%
RWCP-R λ=80.33013.2970.11486.6%
RWCP-R λ=160.35243.4010.03996.8%

Every WCP point on this path is dominated in both WD and F1 by an RWCP-R point. At λ = 2 RWCP-R beats default WCP on all three metrics. At matched boundary counts (λ=4 vs WCP 2/1.5: 0.49 vs 0.50 boundaries per document) RWCP-R still has ΔWD = −0.021 [−0.026, −0.015] and ΔF1 = +0.058 [0.044, 0.072], so the gain does not come from predicting fewer boundaries. Minimum WD with F1 at least the family's default-WCP F1:

FamilyF1 floorWCP min WDRWCP-R min WD
Single change0.4350.3960.190
CoAuthor0.3020.4880.507
Multiple changes0.2050.4360.428
Text attacks0.2720.3610.268
LLM proportion0.2200.3540.393
Family macro0.2870.4070.370

At equal F1, RWCP-R lowers macro WD by 0.037 (selection-aware bootstrap −0.037 [−0.048, −0.024]), with large gains on single changes and attacks.

Thank you for the rebuttal. My main experimental concern is not resolved. It appears that RWCP-R performs worse than WCP in aggregate boundary F1 and count MAE. This reinforces my original concern that improvements in WD do not necessarily translate into improved authorship boundary detection. I will keep my original score.

Reviewer_5LGJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper solving the problem of locating human/LLM authorship boundaries in a document as multiple change point detection on each sentence detector scores. The proposed RWCP is built from one prior work WCP by capping reliability weights, replacing squared error with a Huber profile gain and selecting boundaries by narrowest over threshold search over random intervals. The authors then proposed a second variant RWCP-R that replaces the search with a penalized Viterbi decoder over shared author states with a fitted scale. The authors have experiment on 2690 documents from the WCP benchmark. The family-averaged WindowDiff falls from 0.407 (WCP) to 0.336 (core RWCP) and 0.311 (RWCP-R).

Critical strengths

The paper’s main strength is its complete section of proof and recovery argument for a robust split statistic. It scores each candidate boundary by how much the best achievable loss drops when the interval is allowed two centers instead of one, and shows that a wrongly placed boundary always pays a measurable penalty for merging two different authors. It is proportional to the size of the jump and the amount of reliable data misassigned.

Critical weaknesses

The paper spends most of its space on the theory and the core RWCP algorithm, but the experiments show that this algorithm does not work properly. The second variant RWCP-R does improve the headline metric, but the ablations show that the improvement comes from components other than the ones the introduction motivates and proposes. Here are the details:

The method with the theoretical guarantee (core RWCP) barely provide any prediction in practice. In the main content of this paper, on four of the five benchmark families it returns no boundaries in 82–95% of documents. Its apparent improvement on the headline metric is an artifact of that returning no boundaries since WindowDiff rewards predicting nothing. The more informative metrics count error and boundary F1 show it doing worse than the baseline.

The method that does perform well (RWCP-R) is a different algorithm with no theory. More than that, this paper claim that the RWCP-R's gain from its two key component: reliability weights and the Huber loss. However, one ablation study in this paper shows these two component contribute nothing measurable. Reliability weights have no effect on WD and worsen count error and F1, and the Huber loss gains 0.007 WD at the cost of worse count error and F1. The fitted scale is the only component that helps on every metric.

Questions for authors
  1. Core RWCP returns no boundaries in 82–95% of documents in four of five families, and the no boundary predictor with WD 0.362 already beats WCP with 0.407. How much of the reported 17.6% WD reduction of RWCP remains after factoring out the null predictor? Does any improvement in the metric of MAE and F1, where core RWCP is worse than WCP?

  2. The ablations attribute only roughly 0.014 of RWCP-R’s 0.096 WD gain over WCP to the Huber loss, reliability weights and fitted scale combined. Can you isolate the contribution of the recurrent decoder and penalty lambda and explain why the paper’s motivation is not reflected in the components that actually drive the gain?

3 comments

Thank you very much for your suggestions, which are of great help to us. We have carefully analyzed your comments and organized our responses below.

1. Core RWCP and RWCP-R

Core RWCP was the analytical starting point of the work. It introduced the weighted robust profile gain and the merge-cost argument that underlies our analysis. RWCP-R builds on the same robust loss and adds the structure that mixed human/LLM documents actually have: sources recur, so RWCP-R shares one robust center per source across nonadjacent passages, fits a residual scale, and decodes states with a transition penalty. It is a standalone decoder and does not run Core first.

Following the reviewer's suggestion, the revision presents RWCP-R as the single method, and the experiments now separate three questions: the gain beyond a no-boundary predictor, the role of the transition penalty, and the contributions of recurrence and the penalty. The new results are reported below.

2. Comparison with the null predictor

For completeness, the evaluated Core configuration used c=1.345c=1.345, weights clipped to [0.2,5][0.2,5], side and guard masses of 1, 500 sampled intervals, δ=0.05\delta=0.05, Cr=1C_r=1, dependence multiplier 1 and seed 42.

MethodWD ↓Count MAE ↓F1 ↑
No boundary0.36223.4330.009
WCP0.40702.8080.287
Core RWCP0.33553.2290.129
RWCP-R, frozen0.31103.0190.230

Metrics are averaged over documents within settings, settings within families, then the five families. F1 uses a one-sentence tolerance; two empty boundary sets score F1 = 1, which gives the null predictor a small nonzero F1.

The WCP-to-Core WD gap splits as 0.0715=0.0448 (WCP to null)+0.0267 (null to Core)0.0715 = 0.0448\,(\text{WCP to null}) + 0.0267\,(\text{null to Core}): the null predictor covers about 63% of the gap, and Core improves on it by a further 0.0267 (7.4% relative), the remaining 37%. RWCP-R goes clearly further: it improves on the null predictor by 0.0511 in WD, almost twice Core's margin, and also by 0.414 in count MAE and 0.221 in F1.

Paired joint source-component bootstrap (2,000 draws, unrounded results):

ComparisonΔWD [95% CI]ΔMAE [95% CI]ΔF1 [95% CI]
Core − WCP−0.0715 [−0.0843, −0.0585]+0.420 [+0.339, +0.502]−0.158 [−0.176, −0.140]
Core − null−0.0267 [−0.0311, −0.0221]−0.204 [−0.226, −0.184]+0.120 [+0.108, +0.133]
RWCP-R − WCP−0.0959 [−0.1073, −0.0845]+0.211 [+0.126, +0.293]−0.057 [−0.073, −0.041]
RWCP-R − null−0.0511 [−0.0588, −0.0432]−0.414 [−0.444, −0.383]+0.221 [+0.203, +0.240]

All four RWCP-R − null intervals favor RWCP-R, so its gain is not a by-product of predicting fewer boundaries.

On MAE and F1, Core improves on WCP on both metrics for single changes (MAE 1.637 to 0.393, F1 0.435 to 0.487), and this family accounts for about 85% of Core's macro WD gain over null. On attacks and LLM proportion Core lowers MAE; on these one-change documents its analytic threshold is conservative, which is why the revision moves to RWCP-R, whose penalty can be set explicitly.

3. What drives RWCP-R

Each variant changes one component of the frozen configuration on the same cached inputs, with no retuning:

VariantMacro WD ↓ΔWD vs. full [95% CI]MAE ↓F1 ↑Zero CP
Full RWCP-R0.3110-3.0190.23060.2%
No recurrence (one center per segment)0.3584+0.047 [+0.040, +0.055]3.0720.23049.5%
Near-zero penalty0.7299+0.419 [+0.407, +0.431]10.6490.3330.3%
Fixed scale, σ=1\sigma=10.3170+0.006 [+0.002, +0.010]3.0710.21164.0%
Quadratic loss, c=106c=10^60.3179+0.007 [+0.003, +0.011]2.9450.23954.9%
Uniform weights, η=0\eta=00.3118+0.001 [−0.004, +0.006]2.9170.26151.0%
WCP0.4070-2.8080.28721.7%

Recurrence is the largest single contributor (+0.047 WD when removed, about half of the gap to WCP); a short passage borrows the center estimated from all passages of its source. The penalty is indispensable: without it the decoder switches on sentence-level noise (13.8 boundaries per document against 3.4 true). The fitted scale improves WD, MAE and F1 together. Huber loss and weights act as safeguards whose value appears under contamination (below). Since components interact in one objective, leave-one-out effects are not additive.

Recurrent versus non-recurrent decoder (same loss, weights, scale and standardization) along the penalty path:

λRecurrent WDMAEF1Zero CPNon-rec. WDMAEF1Zero CP
0.250.6989.060.3440.3%0.88125.010.2430.3%
0.50.6657.670.3560.3%0.87523.250.2510.3%
10.5734.950.3820.6%0.86820.880.2610.3%
20.3702.510.38312.5%0.84717.000.2731.0%
4 (frozen)0.3113.020.23060.2%0.3583.070.23049.5%
80.3303.300.11486.6%0.3333.300.11685.6%
160.3523.400.03996.8%0.3493.390.04496.1%

Shared centers give lower WD at every λ ≤ 8 and far less over-segmentation. At λ = 2, RWCP-R improves on default WCP in all three metrics (WD 0.370 vs 0.407, MAE 2.51 vs 2.81, F1 0.383 vs 0.287) with fewer empty predictions (12.5% vs 21.7%).

To rule out a gain from reporting fewer boundaries, we paired RWCP-R and WCP settings with matched mean predicted boundary counts (RWCP-R − WCP, paired bootstrap):

PairBoundaries per docΔWD [95% CI]ΔF1 [95% CI]ΔMAE [95% CI]
λ=2 vs WCP default2.49 / 2.25−0.037 [−0.048, −0.024]+0.096 [+0.080, +0.112]−0.30 [−0.41, −0.19]
λ=4 vs WCP (r=2)0.49 / 0.50−0.021 [−0.026, −0.015]+0.058 [+0.044, +0.072]−0.062 [−0.100, −0.023]
λ=8 vs WCP (r=4)0.14 / 0.12−0.008 [−0.012, −0.005]+0.036 [+0.026, +0.047]−0.021 [−0.037, −0.004]

At equal boundary counts RWCP-R is better on every metric.

Contamination test: 5% or 10% of scores replaced by median-centered values at eight robust standard deviations, optionally with the affected reliability proxies inflated 20 times:

ScorespFull WDΔWD, remove HuberΔWD, remove weights
Original-0.3110+0.007+0.001
Extreme5%0.3849+0.056−0.015
Extreme10%0.4449+0.012−0.023
Extreme + 20× proxy5%0.4399+0.064−0.066
Extreme + 20× proxy10%0.5717−0.022−0.148

Positive values mean removal hurts. With 5% extreme scores the Huber loss is worth 0.056 WD, eight times its value on clean scores, which is the regime it was designed for. On paragraph-level GovReport (known K=1K=1), removing the weights raises WD from 0.29–0.35 to 0.41–0.45 for k≥4k\ge4, showing their value when unit lengths vary widely.

4. Revisions

RWCP-R becomes the single method. The new theory targets the recurrent objective: a shared-state merge-cost margin, uniform control of assignment losses, penalized decoding recovery, transfer of parameter-estimation error, and convergence of the conditional updates. Section 5 reports recovery metrics with a null reference, the matched-count and penalty comparisons, the ablations and contamination tests, and all benchmark families are retained.

Thank you for acknowledging all my comment and existing limitation I proposed during this review. I believe fixing every points will be a significant modification to the original manuscript that would not likely to be done during the rebuttal period. I would suggest the authors to further refine their work and submit to other conference. Therefore, my score for this work will remains the same. Good luck.

Reviewer_zDu4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes RWCP, a method that combines capped reliability weights, Huber profile gains, and a narrowest-over-threshold search conducted in a reliability-coordinate system, in order to localize the boundaries between human-written and LLM-written text. Its core proposition is to characterize the population-level cost of an incorrect split through a merge cost, thereby circumventing the need to solve for the Huber center of a mixed distribution. An additional decoder, RWCP-R, shares source-state centers across non-adjacent passages and performs Viterbi decoding together with a fitted scale. Across multiple benchmarks, the proposed method reduces the family-macro WindowDiff.

Critical strengths
  1. The problem definition and motivation are clear.
  2. The theoretical contribution is clear and detailed.
Critical weaknesses
  1. My primary concern lies with the empirical results of Core RWCP, whose performance on several experiments diverges substantially from both the theory and the authors' claims. In Section 4.1, the authors report that "its family-macro count MAE rises from 2.808 to 3.229 and matched-boundary F1 falls from 0.287 to 0.129." Table 6 further shows that the Zero CP rate exceeds 0.9 in multiple experiments. Although the author acknowledges this gap, the disconnect between the theory and the empirical subjects diminishes the practical significance of the theorem.
  2. The novelty of this paper is not entirely clear. Each ingredient (capped reliability weights, Huber profile losses, and narrowest-over-threshold search) has appeared in prior work, and the paper does not explicitly isolate what the specific combination delivers that no single component does. The merge-cost identity and the resulting profile-gain difference analysis constitute a non-trivial technical step, but the paper does not establish when this step yields a strict advantage over existing robust change-point procedures.
Questions for authors

Please see the Critical Weaknesses. Besides, the discussion of computational complexity is confined to asymptotic order and implementation-dependent empirical descriptions. The paper reports no measured runtime, throughput, or constant-factor comparison against baselines.

2 comments

Thank you very much for your suggestions and support. We have carefully analyzed your comments, organized them and supplemented the corresponding experiments, hoping to address your concerns. We summarize the issues you raised as follows:

1. Theory and practice

Theorem 3.11 analyzes Core RWCP: detection requires the empirical profile gain to exceed an analytic threshold, while the population separation grows with reliability mass and the squared jump. The fixed analytic threshold is deliberately conservative, which keeps false boundaries rare but leaves short or weakly separated changes below it. The revision therefore moves to RWCP-R, which replaces the threshold with an explicit transition penalty and shares information across passages of the same source:

J(z,μ,σ)=∑i=1Nρc ⁣(w~i(Y~i−μzi)σ)+Nlog⁡σ+λK(z),K(z)=∑i=2N1{zi≠zi−1},J(z,\mu,\sigma)=\sum_{i=1}^{N}\rho_c\!\left(\frac{\sqrt{\tilde w_i}(\tilde Y_i-\mu_{z_i})}{\sigma}\right)+N\log\sigma+\lambda K(z),\qquad K(z)=\sum_{i=2}^{N}\mathbf 1\{z_i\ne z_{i-1}\},

with σ≥σmin⁡>0\sigma\ge\sigma_{\min}>0 and λ≥0\lambda\ge0; the number of states SS is not the number of transitions KK. On the same cached scores RWCP-R lowers macro WD from 0.4070 (WCP) to 0.3110, has lower WD than WCP in 23 of 25 settings, and at matched boundary counts is better than WCP in WD, F1 and count MAE.

The revision develops theory for the recurrent objective under shared state-specific Huber locations, separated centers, minimum run mass, bounded weights, curvature and an explicit dependence condition:

  • R1. Merge-cost bounds for disjoint, possibly nonadjacent sets, giving identifiability of shared centers up to label permutation.
  • R2. Uniform control of empirical versus population assignment-loss differences, with state number, transition complexity and dependence.
  • R3. Recovery for exact Viterbi decoding with known centers and scale.
  • R4. Pooled-center estimation and transfer of center/scale error to the decoding margin.
  • R5. Objective descent and convergence for the conditional updates, and the effect of tolerances and initialization.

For R3, with Δw=min⁡j∑i=τj−1+1τjw~i\Delta_w=\min_j\sum_{i=\tau_{j-1}+1}^{\tau_j}\tilde w_i and κj=∣μzτj+1⋆⋆−μzτj⋆⋆∣/σ⋆\kappa_j=|\mu^\star_{z^\star_{\tau_j+1}}-\mu^\star_{z^\star_{\tau_j}}|/\sigma^\star, the target is

C1ΛN(δ)≤λ≤C2κmin⁡2Δw,K^=K,dw(τ^j,τj)≲ΛN(δ)κj2,C_1\Lambda_N(\delta)\le\lambda\le C_2\kappa_{\min}^2\Delta_w,\qquad \widehat K=K,\qquad d_w(\widehat\tau_j,\tau_j)\lesssim\frac{\Lambda_N(\delta)}{\kappa_j^2},

which keeps the log⁡/κ2\log/\kappa^2 localization order of the Core result and gives an explicit range for the penalty. The scale equation ∑iψc(ri)ri=N\sum_i\psi_c(r_i)r_i=N is analyzed separately, and the solver guarantees are stated at the scope of the conditional updates.

Experiments in the revision: (1) all five families with WD, count MAE, precision/recall/F1 and Zero CP, plus a null reference; (2) λ sweeps for RWCP-R and the matching WCP parameter, including matched-boundary-count comparisons; (3) CoAuthor selection on 289 validation sessions and testing on 290; (4) simulations varying separation, run mass, recurrence, weight heterogeneity and dependence, including no-change sequences.

2. Components and the overall method

The contribution lies in combining these tools in one recurrent robust objective with a supporting analysis. Each component addresses a specific failure of simpler alternatives:

  • Capped reliability weights: use heterogeneous score precision while limiting the leverage of any single unit. Test: uniform vs. weighted, with accurate, noisy and inflated proxies.
  • Huber emissions: limit the effect of extreme residuals on centers and assignments. Test: Huber vs. quadratic under the same decoding and controlled contamination.
  • Shared recurrent centers: pool information across nonadjacent passages of one source, so short passages are recognized from a well-estimated center. Test: shared vs. per-segment centers.
  • Fitted scale: keeps the penalty's effective strength comparable across documents. Test: fixed vs. fitted scale along the penalty path.
  • Transition penalty: prevents switching on sentence-level noise. Test: near-zero penalty and a full sweep.

The leave-one-out results confirm each role: removing recurrence costs 0.047 macro WD and removing the penalty 0.419; the fitted scale improves WD, MAE and F1 together; and with 5% extreme scores the Huber loss is worth 0.056 WD, eight times its effect on clean scores. On paragraph-level GovReport, where unit lengths vary widely, the weights lower WD from 0.41–0.45 to 0.29–0.35. The components interact in one objective, so their effects are reported jointly rather than as an additive split. RWCP-R is solved by alternating exact conditional decoding with center/scale fitting; this recurrent structure, absent from WCP's contrast-and-NOT pipeline, is the core of our contribution.

3. Merge cost and when it helps

The merge-cost argument analyzes optimized robust losses without solving for a mixed-distribution Huber center, which is what makes a robust profile analysis tractable. For fixed σ\sigma, with QA(θ;σ)=∑i∈AE ρc{w~i(Y~i−θ)/σ}Q_A(\theta;\sigma)=\sum_{i\in A}\mathbb E\,\rho_c\{\sqrt{\tilde w_i}(\tilde Y_i-\theta)/\sigma\} and Dσ⋆(A,B)=inf⁡QA∪B−inf⁡QA−inf⁡QBD^\star_\sigma(A,B)=\inf Q_{A\cup B}-\inf Q_A-\inf Q_B, the definition holds for disjoint nonadjacent sets, is zero for a shared center, and under suitable curvature satisfies

Dσ⋆(A,B)≥m02σ2WAWBWA+WB(μA−μB)2,WA=∑i∈Aw~i.D^\star_\sigma(A,B)\ge\frac{m_0}{2\sigma^2}\frac{W_AW_B}{W_A+W_B}(\mu_A-\mu_B)^2,\qquad W_A=\sum_{i\in A}\tilde w_i .

This is the basis of R1. The link to the penalty is direct: with centers and scale fixed, deleting an internal run II in state bb surrounded by state aa changes the objective by ∑i∈I{ℓi(a)−ℓi(b)}−2λ\sum_{i\in I}\{\ell_i(a)-\ell_i(b)\}-2\lambda, so a run is kept exactly when its fit gain pays for two transitions.

Recurrence helps when sources repeat: with known states and fixed scale, center error is of order σ⋆log⁡(S/δ)/Ws\sigma^\star\sqrt{\log(S/\delta)/W_s}, with WsW_s the reliability mass pooled over all occurrences of state ss rather than a single short passage. Robust emissions help when extreme scores would bias center estimates, and weights help when reliability varies across units. WCP's analysis assumes independent sub-Gaussian scores; ours covers heavy tails and dependence and keeps the same order when the targets coincide. We will also compare with robust penalized segmentation (Fearnhead and Rigaill, 2019) under matched calibration.

4. Runtime and throughput

With prefix sums, a weighted-mean contrast costs O(L)O(L) per interval of length LL; the WCP timings below use dynamic programming over change counts up to N/2N/2. RWCP-R costs O(NS2)O(NS^2) per Viterbi update and O(RTN(S2+SB))O(RTN(S^2+SB)) overall for RR starts, TT iterations and BB fitting steps, with O(NS)O(NS) memory, i.e. linear in NN.

Setup: NVIDIA A800 80GB (PyTorch 2.5.1, float32) and an Intel Xeon Gold 6348 limited to one thread, the same 100 documents per dataset, model loading excluded, GPU warm-up and synchronization. Detector pipelines use Gemma-3-1B with the AdaDetectGPT adapter. VCP/WCP rescore predicted segments for labels; RWCP-R labels its states directly.

Dataset (mean sentences)SenPred s/docVotingVCP/WCPRWCP-RVCP/WCP sent/sRWCP-R sent/s
Single change, Story (32.6)1.721.712.001.7216.319.0
Multiple changes, K=8 (93.0)4.894.995.484.8917.019.0
LLM proportion 40% (36.8)1.911.932.171.9117.019.3
CoAuthor (23.3)1.211.161.431.2116.319.3

RWCP-R's detector stage is 11–15% faster because it needs no rescoring pass.

Segmentation from cached scores, single-thread CPU (ms/doc):

DatasetVCPWCPTextTilingRWCP-R NumbaRWCP-R PythonWCP/Numba
Single change, Story2.92.829.60.642.24.7×
Multiple changes, K=857.858.2160.01.1133.052.9×
LLM proportion 40%3.63.635.10.747.85.1×
CoAuthor1.81.815.20.629.73.0×

The Numba implementation is the fastest method on every dataset, up to 52.9 times faster than WCP; a pure-Python fallback is also provided. Timing CoAuthor with the NOT search that WCP actually uses there (5 repetitions per document, single thread) gives 1.67 ms for WCP and 0.24 ms for RWCP-R (about 7 times faster).

Synthetic two-state documents, single-thread CPU (ms):

N2550100200400800
VCP1.67.453.04755,35371,642
WCP1.57.352.14755,25671,772
RWCP-R Numba0.60.81.01.72.86.2
RWCP-R Python35.077.21152744921,216

RWCP-R scales linearly: at N = 800 it takes 6.2 ms (Numba) or 1.216 s (Python), against about 71.8 s for WCP, more than 10,000 times faster. Summing both stages, RWCP-R has the lowest total time on all four real datasets, in both implementations. Timings include all six starts and iterations; JIT compilation is excluded by a warm-up call.

The paper is rewritten around RWCP-R: abstract and §1 present recurrent source states and transition regularization; §2 defines states, transitions and reliability mass and relates the method to WCP and robust penalized segmentation; §3 covers standardization and weights, the recurrent objective, conditional Viterbi with center/scale updates, initialization and penalty choice,; Figure 1 is redrawn; §4 holds the recurrent theory (R1–R5); §5 reports all benchmark families with a null reference, recovery metrics, matched-count and penalty comparisons, the validation-selected CoAuthor comparison, component and contamination studies, simulations and runtime; §6 concludes; appendices give proofs, solver and ablation details, split provenance and timing.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.