Topic Identification in LLM Input-Output Pairs through the Lens of Information Bottleneck

Igor Halperin

REVIEW 4 major objections 3 minor 11 references

Topic Identification in LLM Input-Output Pairs through the Lens of Information Bottleneck

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that topic identification for LLM hallucination detection should be reformulated as a deterministic information bottleneck problem, and delivers a tractable algorithm, UDIB, that sharpens Semantic Divergence Metrics by mak

desk verdict UDIB is a sound, practical twist on DIB clustering, but the claimed SDM sensitivity gains rest on a confounded baseline comparison. read the letter →

arxiv 2509.03533 v1 pith:DAJCCRFG submitted 2025-08-26 cs.CL cs.LGq-fin.GN

Igor Halperin This is my paper

classification cs.CLcs.LGq-fin.GN MSC 68T5062H3094A17

keywords deterministicinformationbottleneckUDIBtopicidentificationsemanticdivergencemetricsconfabulationdetectionLLMhallucinationsembeddingclustering

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the standard way to find topics in LLM prompt–response pairs—geometric clustering of sentence embeddings—is mismatched to the downstream information-theoretic analysis, and fixes the mismatch by deriving a practical Deterministic Information Bottleneck algorithm. The algorithm, UDIB, is an entropy-regularized K-means that trades spatial tightness against cluster parsimony and is directly optimized to keep the topic representation informative about the prompt–response relationship. Applied within the Semantic Divergence Metrics framework, it yields cleaner topics, more interpretable co-occurrence maps, and scores that separate task types more sharply—for closed-domain prompts the divergence measures become monotonic in task complexity, which the older clustering did not show. If correct, this gives confabulation detection a more sensitive measurement tool without overturning the SDM framework's conclusions.

What carries the argument

UDIB (Upper-bounded Deterministic Information Bottleneck): an iterative hard-clustering algorithm whose per-point loss is (1/n_c) sum_{j in c} ||x_i - x_j||² − τ log q(c), with τ = 2s²/β. It replaces DIB's intractable KL term with a Jensen/Hershey–Olsen upper bound, collapsing smoothing scale s and tradeoff β into one temperature-like parameter; the number of clusters is selected by the Kink Angle heuristic on the information profile H[c].

What would settle it

Fix one embedding model and the same SDM pipeline; run agglomerative clustering and UDIB on the same Set A and Set B prompt–response data with comparable model-selection rules. If the monotonic Set B gradient (0.1628 < 0.2315 < 0.2924) and the wider dynamic range disappear when embeddings are held fixed, UDIB's claimed sensitivity gain dissolves.

Watch

Extended reading notes

Core claim

The paper's central claim is that the topics used to compare LLM prompts with their responses should be chosen by how much information they carry about the prompt–response relationship, not by spatial proximity alone. It makes the Deterministic Information Bottleneck practical for high-dimensional embeddings by replacing the intractable KL divergence between a Gaussian and a Gaussian mixture with the Hershey–Olsen convexity upper bound. The resulting UDIB assignment rule—minimize the mean pairwise squared distance within a cluster plus τ times the negative log cluster size—is an entropy-regularized K-means that prunes uninformative clusters on its own. Fed into the Semantic Divergence Metric

Load-bearing premise

The headline comparisons reuse SDM numbers from an earlier study that clustered with a different algorithm and, in all likelihood, different sentence embeddings; if the embeddings differ, UDIB's apparent advantage could be an artifact of the embeddings rather than of the clustering method, so the paper's central claim would not be established.

Editorial extensions

If this is right

If UDIB topics are used, SDM remains coherent: the core ordinal conclusions—high vs moderate vs low stability in Set A, and forced hallucination as a low-scoring outlier in Set B—are preserved.
The widened dynamic range means tasks can be graded continuously: factual recall, comparison, and forecasting now occupy ordered values on the divergence scale instead of overlapping.
A low-and-stable score on a nonsensical prompt becomes an explicit 'confident confabulation' signature, distinguishable from semantically unstable exploration.
Built-in model selection removes the need to fix the number of topics beforehand, making topic identification more parsimonious.
Because UDIB is a modified K-means with one effective hyperparameter, it can be dropped into existing embedding-based pipelines beyond LLM input-output analysis.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Swapping the clustering algorithm is not the only change between the baseline and new tables; the sentence embeddings are also different. A clean test would fix the embedding model and vary only the clustering method to confirm that UDIB itself, not the new embeddings, drives the sharper SDM scores.
The Kink Angle heuristic's recommended cluster count varies widely across seeds (e.g., 10.50 ± 1.50), so the final k is a meta-statistic over runs. A natural extension would quantify how SDM scores fluctuate when k is perturbed within the stability region of τ.
The single effective temperature τ = 2s²/β means the smoothing scale and the information–compression tradeoff are not separately identifiable; probing with several τ values would show whether topic granularity, rather than the entropy penalty, is doing the work.
The information profile diagnostics could be repurposed as a test-time signal: monitoring the kink angle across dialogue turns might detect topic drift or the onset of confabulation without ground-truth labels.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

The takeaway: the UDIB algorithm is real and the math checks out; the empirical case that it makes SDM more sensitive is not established because the baseline comparison is uncontrolled.

What's new: substituting the Hershey-Olsen convexity bound for the intractable KL between a Gaussian and a GMM in the DIB assignment step is a neat trick. It turns DIB into a practical iteration that is essentially K-means with a pairwise-distance term plus an entropy penalty, and it collapses s and beta into one temperature-like parameter. The Jensen inequality argument linking the loss to K-means is correct. The multi-seed procedure and the kink-angle model selection are reasonable engineering. The qualitative topic analysis in the appendix is actually the most convincing part: the UDIB clusters separate semantic content from prompt artifacts (CONSTRAINTS markers, JSON) and split Hamlet's feigned vs. Ophelia's genuine madness. That suggests the algorithm is doing something right.

Where it gets soft: the headline claim of superior confabulation sensitivity rests on comparing UDIB SDM scores (Tables 3, 6) against numbers 'reproduced' from the author's earlier paper [3]. The paper never says what embedding model [3] used. This paper uses Qwen3-Embedding-0.6B. If the baseline used a different embedding model, the comparison is embedding-vs-embedding, not clustering-vs-clustering. On top of that, the number of topics k is not matched: Table 3 uses k=8,9,9 while Table 1's recommended k for the same prompts under the kink heuristic is 10, 10-11, 9-ish; for Set B the k values differ across prompts (7,7,5,8) and the baseline k isn't reported. Since entropy-based SDM metrics depend on k, the 'monotonic trend' and 'dynamic range' claims could be an artifact. Also, the SDM tables have no error bars, and the Wasserstein rows are numerically identical between baseline and UDIB, which makes me wonder whether those were computed on raw embeddings (legit) or copied (sloppy).

One point where I disagree with the reader's report: Eq (18), H[c|x]=0, is fine for a deterministic assignment rule—given x, c is a function of x. That's not a flaw.

Who should read this: people working on information-bottleneck clustering and anyone building faithfulness detectors who wants a principled alternative to vanilla K-means. The algorithmic contribution is worth citing. The experimental claim needs a redo: same embeddings for baseline, matched k, error bars, and code.

Recommendation: send to peer review. The algorithm is sound and the application is relevant; a good referee would ask for the same-embedding baseline and uncertainty quantification. That's a fixable revision, not a desk reject.

Referee Report

4 major / 3 minor

Summary. The paper proposes UDIB, a tractable version of Geometric Deterministic Information Bottleneck clustering, obtained by replacing the KL divergence between a Gaussian and a Gaussian mixture with the Hershey-Olsen convexity bound. This yields an entropy-regularized K-means-like assignment rule with a single effective temperature tau = 2s^2/beta. The method is applied to joint clustering of Qwen3-Embedding sentence embeddings of prompts and responses, with the number of topics selected by a 'kink-angle' heuristic over multiple seeds. SDM scores computed on UDIB topics are compared with scores from an earlier agglomerative-clustering baseline; the paper claims that the UDIB topic space is more interpretable and makes SDM metrics more sensitive and discriminative, especially through monotonic increases across Set B task complexity.

Significance. The mathematical core is mostly sound and presented with useful detail: the upper-bound substitution is correctly derived (Eqs. 10-14), and the Jensen comparison with K-means (Eq. 17) is correct. The method genuinely reduces the DIB optimization to a practical algorithm with one effective hyperparameter and gives a reproducible initialization/selection protocol (Algorithm 1, multi-seed procedure). If the empirical claims were supported by a controlled comparison, this would be a useful advance for LLM faithfulness-hallucination detection. However, the current experimental evidence does not support the headline sensitivity claims because the baseline is confounded and the reported SDM values lack uncertainty quantification.

major comments (4)

[Section 5.2.1, Section 5.4, Tables 2/3 and 5/6] The central empirical claim of enhanced sensitivity rests on comparisons to Tables 2 and 5, reproduced from [3]. Section 5 specifies Qwen3-Embedding-0.6B for all experiments in this paper, but the embedding model (and k) used in [3] is never given. If the baseline used different embeddings, every SDM entry changes even with identical clustering. Moreover, the UDIB tables use different topic counts (Set A: k=8,9,9; Set B: 7,7,5,8) than the baseline, whose k is not reported; H(P), JSD and KL are all functions of k. The Section 5.4 statements about 'wider dynamic range' and the monotonic Set B trend are therefore not attributable to UDIB without rerunning AC on Qwen3 embeddings with matched k and reporting the baseline k.
[Section 4.5, Eq. (18)] The identity I(c;x)=H[c] requires H[c|x]=0. In the paper's model p(x|i)=N(x; x_i, s^2 I) with finite s, x does not determine the index i uniquely, so p(c|x) is a mixture over indices and H[c|x]>0 in general. The equality holds only in the s->0 limit, which is not the regime used by the algorithm (tau=2s^2/beta is finite). Since Eq. (18) is the stated justification for the information profile used by the kink-angle selector, the model-selection diagnostic needs a corrected derivation or an explicit estimate of H[c|x].
[Tables 3 and 6; Section 5.2] The UDIB SDM tables report single numbers without standard errors or confidence intervals, despite the multi-seed design. Table 1 shows large seed-to-seed dispersion (e.g., Rec. # Clusters 10.00 +/- 1.84 and Kink angle 33.70 +/- 22.56 for High Stability). Since the final k and assignments vary across seeds, all SDM metrics inherit this variability; the claimed monotonic increases in Set B (SH 0.1628, 0.2315, 0.2924) could lie within the run-to-run noise. Report per-seed SDM distributions or a sensitivity analysis over the M runs.
[Section 5.5 vs. Table 3 and Appendix A] There is a discrepancy in the number of topics used for SDM. Table 3 uses k=9 for both Moderate and Low Stability, while Appendix A.2 and A.3 report optimal nc=10 and nc=11, and Table 1 gives Kink Angle mean recommendations of 10.5 and 8.6. The mode-based rule in Section 5.5 is not documented with the actual per-run nc histogram, so the reader cannot see how the final k values were obtained. This is not merely cosmetic: every SDM value in Tables 3 and 6 depends on k.

minor comments (3)

[Throughout] Typographical issues include 'promts' (Introduction), 'We For all experiments' (Section 5.2.1), 'CSummary' (Table 1 header), 'F actual' (Table 5), and 'V ariability' (Section 5.5).
[Figures 2 and 4] The heatmaps would be more informative with a shared color scale and axis labels for topic indices. The qualitative descriptions are somewhat subjective and hard to verify from the small figures.
[Equation (15)] The notation E_{i,c,j in S_c} is informal; clarify the dependence on cluster c and which quantities are random. Also, the phrase 'total loss function' should be defined consistently with the per-point Lagrangian in Eq. (14).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the UDIB derivation is self-contained; the baseline comparison raises a reproducibility concern but is not circular.

full rationale

The paper's central derivation chain is: DIB objective (Eq. 2) + Strouse-Schwab geometric DIB + Hershey-Olsen KL upper bound (Eq. 10) -> tractable UDIB assignment rule (Eq. 14). This chain is self-contained and does not use the downstream SDM metrics as an input; the UDIB loss is an entropy-regularized K-means loss with an externally derived upper bound, and no parameter is fitted to the SDM outputs. Model selection uses the information profile / kink-angle heuristic computed from cluster entropy H[c], not from the SDM scores reported later. The empirical comparison against Tables 2 and 5, which are reproduced from the author's prior paper [3], is a potential confound because the paper states 'For all experiments in this paper, sentence embeddings are done using the Qwen3-Embedding-0.6B model' but does not state the embedding model or exact k values used in [3]. That is a validity/reproducibility risk concerning whether the reported sensitivity gains come from UDIB rather than from a different embedding model or topic count, but it is not a circular derivation: the UDIB results are computed independently and the baseline is an external prior result, so no equation or fitted parameter reduces to the claimed conclusion. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The method depends on Gaussian smoothing, hard assignments, and an approximation H[c|x]=0 for model selection. The experimental comparison adds a dependence on the author's prior SDM paper and on the Qwen3 embedding model. No new physical or conceptual entities are introduced.

free parameters (5)

effective temperature tau = 2s^2/beta = swept; chosen with kink-angle heuristic and stability bounds (e.g., tau_min and tau_max in Table 1)
Controls the balance between geometric fit and cluster-entropy penalty. Not derived from data; selected post hoc by the model selection heuristic.
smoothing scale s = set to a fixed number using Eq. (19) and the condition that the normalized information ratio stays below 1
Declared not to enter the loss directly, but it enters the information profile used for model diagnostics; effectively chosen by hand.
initial number of clusters kmax = not specified in text
Algorithm 1 requires an initial kmax; the value is not reported, so runs are not fully specified.
kink-angle window sizes = 2 and 3 neighbors
Heuristic hyperparameter for model selection; an ad hoc choice.
number of seeds M = 10
Multi-seed framework for robustness; the number is arbitrary.

assumptions (7)

domain assumption Gaussian smoothing: p(x|i) = N(x; x_i, s^2 I)
Section 4.1, Eq (3). The data is modeled as a mixture of isotropic Gaussians, which enables closed-form KL bounds but imposes a specific notion of local geometry.
domain assumption Uniform prior over data indices p(i)=1/N
Section 4.1; standard but arbitrary, weights every sentence equally.
domain assumption Hard cluster assignments q(c|i) is a delta
Section 4.1 and Eq (5); follows from DIB, makes I(X;T)=H(T) but also makes H[c|x]=0 in Eq (18) only approximate for finite s.
domain assumption Cluster entropy H[c] equals mutual information I(c;x)
Eq (18), Section 4.5. Requires H[c|x]=0, which holds only in the limit of zero smoothing; used to define the information profile and kink-angle model selection.
standard math Hershey-Olsen convexity bound for GMM KL divergence
Eq (10), Section 4.2. Known 2007 result; used to make the DIB update tractable.
domain assumption Sentence embeddings (Qwen3-Embedding-0.6B) faithfully represent semantic topics
Section 5 intro. All topic discovery is performed on these embeddings; if embeddings do not capture the relevant semantics, the topics inherit the failure.
domain assumption SDM framework from the author's prior work [3] is a valid measure of semantic divergence
Used as the downstream evaluation; the paper does not re-derive or validate SDM.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Topic Identification in LLM Input-Output Pairs through the Lens of Information Bottleneck." pith.science (2026). https://pith.science/paper/DAJCCRFG

@misc{pith2026250903533,
  author       = {Pith},
  title        = {Pith review of: Topic Identification in LLM Input-Output Pairs through the Lens of Information Bottleneck},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DAJCCRFG}},
  note         = {Machine review of arXiv:2509.03533}
}

read the original abstract

Large Language Models (LLMs) are prone to critical failure modes, including \textit{intrinsic faithfulness hallucinations} (also known as confabulations), where a response deviates semantically from the provided context. Frameworks designed to detect this, such as Semantic Divergence Metrics (SDM), rely on identifying latent topics shared between prompts and responses, typically by applying geometric clustering to their sentence embeddings. This creates a disconnect, as the topics are optimized for spatial proximity, not for the downstream information-theoretic analysis. In this paper, we bridge this gap by developing a principled topic identification method grounded in the Deterministic Information Bottleneck (DIB) for geometric clustering. Our key contribution is to transform the DIB method into a practical algorithm for high-dimensional data by substituting its intractable KL divergence term with a computationally efficient upper bound. The resulting method, which we dub UDIB, can be interpreted as an entropy-regularized and robustified version of K-means that inherently favors a parsimonious number of informative clusters. By applying UDIB to the joint clustering of LLM prompt and response embeddings, we generate a shared topic representation that is not merely spatially coherent but is fundamentally structured to be maximally informative about the prompt-response relationship. This provides a superior foundation for the SDM framework and offers a novel, more sensitive tool for detecting confabulations.

Figures

Figures reproduced from arXiv: 2509.03533 by the authors.

**Figure 2.** Comparison of Averaged Topic Co-occurrence Distributions. The top row (a,b,c) [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Information profiles for the four prompts in Set B, showing the variability of the [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Comparison of Topic Co-occurrence for Set B. Top row (a-d) uses Agglomerative [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

11 extracted references · 8 canonical work pages

[3]

Halperin

I. Halperin. Prompt-response semantic divergence metrics for faithfulness hallucination detection in large language models. arXiv preprint, 2025

work page 2025
[1]

M. Cossio. A comprehensive taxonomy of hallucinations in large language models. arXiv preprint arXiv:2508.01781, 2025

arXiv 2025
[2]

Farquhar, S., Kossen, J., Kuhn, L., and Gal, Y. (2024). Detecting Hallucinations in Large Language Models Using Semantic Entropy. Nature, 630, 625-630

work page 2024
[4]

J. R. Hershey and P. A. Olsen. Approximating the Kullback-Leibler divergence between Gaussian mixture models. In 2007 IEEE International Conference on Acoustics, Speech and Signal Processing-ICASSP’07, vol. 4, pp. IV-317, 2007

work page 2007
[5]

Ji, Z., Lee, N., Frieske, R., et al. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, 55(12), 1-38

work page 2023
[6]

S. Lei, Y. Hao, and L. Mei. Revisiting LLM reasoning via information bottleneck. arXiv preprint arXiv:2502.54321, 2025

work page arXiv 2025
[7]

Strouse and D

D. Strouse and D. J. Schwab. The deterministic information bottleneck. Neural Computa- tion, 29(6):1611–1630, 2017

work page 2017
[8]

D. J. Strouse and D. J. Schwab. The information bottleneck and geometric clustering. arXiv preprint arXiv:1712.09657, 2019

work page Pith review arXiv 2019

Show all 11 references

[9]

Tishby, F

N. Tishby, F. C. Pereira, and W. Bialek. The information bottleneck method. InProceedings of the 37th Annual Allerton Conference on Communication, Control and Computing, pages 368–377, 1999

1999
[10]

Z. Yang, Y. Chen, and L. He. Exploring information processing in large language models: Insights from information bottleneck theory. arXiv preprint arXiv:2501.12345, 2025

2025 arXiv
[11]

‘json ”launch year

Zhang, Y., Li, Y., Cui, L., et al. (2023). Siren ’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models. arXiv:2309.01219 [cs.CL]. 18 Appendix A: Detailed Topic Analysis for Set A Prompts This appendix provides the full output of the topic analysis for the...

2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [3]

Halperin

I. Halperin. Prompt-response semantic divergence metrics for faithfulness hallucination detection in large language models. arXiv preprint, 2025

work page 2025

[2] [1]

M. Cossio. A comprehensive taxonomy of hallucinations in large language models. arXiv preprint arXiv:2508.01781, 2025

arXiv 2025

[3] [2]

Farquhar, S., Kossen, J., Kuhn, L., and Gal, Y. (2024). Detecting Hallucinations in Large Language Models Using Semantic Entropy. Nature, 630, 625-630

work page 2024

[4] [4]

J. R. Hershey and P. A. Olsen. Approximating the Kullback-Leibler divergence between Gaussian mixture models. In 2007 IEEE International Conference on Acoustics, Speech and Signal Processing-ICASSP’07, vol. 4, pp. IV-317, 2007

work page 2007

[5] [5]

Ji, Z., Lee, N., Frieske, R., et al. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, 55(12), 1-38

work page 2023

[6] [6]

S. Lei, Y. Hao, and L. Mei. Revisiting LLM reasoning via information bottleneck. arXiv preprint arXiv:2502.54321, 2025

work page arXiv 2025

[7] [7]

Strouse and D

D. Strouse and D. J. Schwab. The deterministic information bottleneck. Neural Computa- tion, 29(6):1611–1630, 2017

work page 2017

[8] [8]

D. J. Strouse and D. J. Schwab. The information bottleneck and geometric clustering. arXiv preprint arXiv:1712.09657, 2019

work page Pith review arXiv 2019