- Rating
- 3
- Confidence
- 3(+1)
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies multilingual DP fine-tuning under cross-lingual tokenization fragmentation. Sequence-level averaging disproportionately attenuates gradient signals for languages with longer token sequences, while DP noise remains length-independent, leading to what they call Entropic SNR Collapse. Under an idealized semantic-signal conservation assumption, they derive an SNR upper bound. To address this issue, the paper proposes ToPA, which combines token-wise ghost clipping with fixed-anchor aggregation to control token-level gradient contributions without active-length normalization. An adaptive variant privately calibrates the token clipping threshold using the Exponential Mechanism.
Critical strengths
strength1. The paper identifies a practically relevant issue in multilingual DP fine-tuning: tokenization fragmentation can lead to substantial sequence-length disparities across languages, which may interact unfavorably with standard DP-SGD. This is a meaningful problem in realistic multilingual LLM settings and has received relatively limited attention.
strength2. ToPA remains computationally efficient through its token-wise ghost clipping implementation. The reported experiments show only moderate runtime and memory overhead compared with standard DP-SGD, making the method practically applicable to parameter-efficient fine-tuning.
strength3. The proposed approach is conceptually straightforward and does not require modifying the pretrained tokenizer or model architecture. The main components can be integrated into the existing DP fine-tuning pipeline without substantial architectural changes.
Critical weaknesses
weakness1. Authors claim that tokenization fragmentation leads to longer sequences, which weakens the gradient signal under sequence averaging while the DP noise remains unchanged. However, the experiments don't directly verify this claim. Figure 1 only shows that cross-token gradient correlation decreases with token distance; it doen't show that increasing tokenization length causes worse DP fine-tuning performance. A more direct experiment in introduction would keep the semantic content and language fixed, vary only the tokenization length, and compare downstream performance under DP and non-DP fine-tuning. This would help separate the effect of tokenization length from other differences across languages.
weakness2. Proposition 1 relies on the assumption that the cumulative semantic signal remains roughly constant for equivalent content under different tokenizations. However, it is unclear how this “semantic signal” is defined or measured from actual gradients. In particular, can be directly computed or separated from the other components of the gradient in practice? This assumption is central to the claimed SNR degradation. A clearer mathematical definition or an empirical validation showing that the cumulative signal is approximately preserved as tokenization length changes would substantially strengthen the theoretical motivation for the proposed method.
weakness3. ToPA replaces the standard active-length normalization with a fixed anchor . This modification does not only address gradient attenuation; it also changes the effective training objective from sequence-uniform to token-uniform, giving longer sequences proportionally larger update mass. Since the most fragmented languages are also those with the longest sequences, it is difficult to determine how much of the observed improvement comes from mitigating the proposed SNR collapse versus simply reweighting long-sequence examples. Although the paper provides a component ablation, a more controlled comparison that matches the effective objective/length weighting across methods would help isolate the contribution of the proposed privacy mechanism.
Questions for authors
See Weaknesses.
6 comments
Table 8: Controlled Single-Language English Thai Tokenization Fragmentation Studies (; cells report spBLEU_th / chrF++_th; bold denotes best DP score)
Part A: Canonical Monolingual English Thai Natural Fertility Tertiles (Low: 0.434, Med: 0.476, High: 0.513 tok/char) & Overall Monolingual Control
| Model | Method | Low Fertility | Med Fertility | High Fertility | Overall Monolingual Thai | Step SNR |
|---|---|---|---|---|---|---|
Llama-3.2-1B | Non-Private (, ) | 11.09 / 30.18 | 8.71 / 27.02 | 9.19 / 26.40 | 9.64 / 27.95 | — |
Llama-3.2-1B | Non-Private (, ) | 11.32 / 30.26 | 9.39 / 27.57 | 9.52 / 26.65 | 10.06 / 28.24 | — |
Llama-3.2-1B | Vanilla DP-SGD () | 5.42 / 19.73 | 5.00 / 18.35 | 4.78 / 17.39 | 5.07 / 18.55 | |
Llama-3.2-1B | Fixed-Anchor Only (, No Tok. Clip) | 4.41 / 15.74 | 3.71 / 14.80 | 3.70 / 12.75 | 3.93 / 14.50 | |
Llama-3.2-1B | Adaptive ToPA (, Ours) | 5.77 / 21.58 | 5.47 / 20.43 | 5.45 / 19.14 | 5.56 / 20.45 | |
Llama-3-8B | Non-Private (, ) | 20.87 / 39.65 | 20.59 / 37.59 | 20.07 / 36.88 | 20.51 / 38.11 | — |
Llama-3-8B | Non-Private (, ) | 20.97 / 39.86 | 19.94 / 37.28 | 20.13 / 36.83 | 20.34 / 38.07 | — |
Llama-3-8B | Vanilla DP-SGD () | 18.58 / 36.34 | 16.59 / 33.23 | 15.73 / 30.95 | 16.94 / 33.63 | |
Llama-3-8B | Fixed-Anchor Only (, No Tok. Clip) | 17.49 / 35.07 | 15.60 / 31.61 | 15.14 / 30.25 | 16.05 / 32.42 | |
Llama-3-8B | Adaptive ToPA (, Ours) | 19.36 / 37.23 | 17.61 / 34.36 | 16.80 / 32.73 | 17.90 / 34.88 |
Part B: Controlled Intra-Sentence BPE Subword Splitting on Identical English–Thai Sentences (8B Realized & Expansion Ratio; 1B within ±0.02 tok)
| Split Setting | 8B Realized | 8B Non-Private () | 8B Vanilla DP-SGD | 8B Adaptive ToPA | 1B Non-Private () | 1B Vanilla DP-SGD | 1B Adaptive ToPA |
|---|---|---|---|---|---|---|---|
| Canonical () | 58.31 () | 20.51 / 38.11 (PPL 2.61) | 16.94 / 33.63 | 17.90 / 34.88 | 9.64 / 27.95 (PPL 4.19) | 5.07 / 18.55 | 5.56 / 20.45 |
| Split-1.35 | 78.36 () | 7.08 / 23.02 (PPL 6.39) | 15.73 / 31.87 | 15.89 / 32.04 | 2.80 / 15.19 (PPL 7.91) | 4.81 / 17.38 | 4.62 / 17.54 |
| Split-1.70 | 92.55 () | 5.41 / 20.35 (PPL 15.35) | 15.68 / 31.78 | 15.71 / 31.98 | 1.89 / 13.06 (PPL 18.85) | 4.25 / 16.20 | 4.28 / 16.23 |
| Split-2.10 (Ceiling) | 93.41 () | 5.19 / 20.30 (PPL 16.31) | 15.71 / 31.78 | 15.80 / 32.06 | 1.68 / 12.71 (PPL 19.73) | 4.40 / 16.26 | 4.28 / 16.10 |
Q2. Proposition 1 relies on the assumption that the cumulative semantic signal () remains roughly constant for equivalent content under different tokenizations. However, it is unclear how this “semantic signal” is defined or measured from actual gradients. In particular, can () be directly computed or separated from the other components of the gradient in practice? This assumption is central to the claimed () SNR degradation. A clearer mathematical definition or an empirical validation showing that the cumulative signal is approximately preserved as tokenization length changes would substantially strengthen the theoretical motivation for the proposed method.
Because token positions vary with sentence realization and tokenizer, we formalize the semantic signal at the sequence level (Section 3.2 (Finding 1), Table 1; Appendix Section E.6, Table 16; formalizing Proposition 1 in Section 2.2) over a distribution of latent semantic propositions realized as text in language/tokenizer at a fixed LoRA state (warmed up for 15 non-private steps so and both and are active):
where and captures zero-mean lexical/syntactic and example-sampling fluctuations around the mean token signal ( identically, distinct from DP Gaussian noise). Crucially, requires no token-to-token alignment: its empirical estimator is directly computable by averaging unnormalized sequence gradients across a parallel multi-way corpus at fixed . Whenever with , sequence-level averaging attenuates signal power as against length-independent DP noise.
Evaluating across all parallel FLORES-200 dev sentences in 8 languages at using each model's native tokenizer target lengths (Review 1, Table 5), we find on Llama-3.2-1B, on Gemma-2-2B (reflecting high English norm 206.20 leveraged across 2B's narrow non-English length band while non-English scripts stay bounded in ), and on Llama-3-8B—strictly below linear growth ( ) on every model, with 6 languages (tur, ind, rus, arb, swh, tha) forming an explicit conservation plateau on Llama-3-8B across . Without imposing conservation, unconditional empirical DP SNR under averaging decays with (1B), (2B), and (8B), dropping from English to Thai by 11.1x on 1B and 71.2x on 8B.
Q1. Authors claim that tokenization fragmentation leads to longer sequences, which weakens the gradient signal under sequence averaging while the DP noise remains unchanged. However, the experiments don't directly verify this claim. Figure 1 only shows that cross-token gradient correlation decreases with token distance; it doesn't show that increasing tokenization length causes worse DP fine-tuning performance. A more direct experiment in introduction would keep the semantic content and language fixed, vary only the tokenization length, and compare downstream performance under DP and non-DP fine-tuning. This would help separate the effect of tokenization length from other differences across languages.
We thank Reviewer rV2L for engaging closely with our core thesis: how tokenization fragmentation impacts Differential Privacy in multilingual settings, and for suggesting a direct controlled experiment holding semantic content and language fixed while varying only tokenization length. To isolate tokenization fragmentation from language differences while holding semantic content fixed, we present three complementary controlled experiments (ε=4.0,δ=10−5, Table 8 and Table 5 below; incorporated directly into Section 1 (Figure 1b–c) & Section 3.2 (Findings 1–2), Tables 1–4 & Figure 2d and detailed in Appendix Section E.6, Tables 16–17):
devtestsentences,Gemma-2-2B's 256k SentencePiece vocabulary tokenizes Thai into Ntha=49.51 tokens and suffers only a-3.7%relative DPspBLEU_thdrop from Non-Private (15.38) to Vanilla DP-SGD (14.81), whereasLlama-3.2-1B's 128k BPE tokenizer fragments the exact same Thai sentences into Ntha=58.82 tokens and suffers a-35.4%relative DP drop (8.23 → 5.32). Similarly, onLlama-3-8Bover the same parallel sentences, higher-fragmentation Thai (Ntha=58.82) suffers a-14.2%relative DP drop (19.50 → 16.73), 8.8× steeper than lower-fragmentation Arabic (Narb=43.67,-1.6%,20.50 → 20.17).dev, 897 pairs) and stratifyingdevtest(1,012 pairs) by within-Thai tokenizer fertility (tok/chardensity: Low0.434, Medium0.476, High0.513, where target token counts remain matched in57.43–60.69), onLlama-3-8BNon-Private (1/Ni)spBLEU_thdrops by only-0.80from Low to High fertility (20.87 → 20.07;chrF++_th:-2.77), whereas Vanilla DP-SGD (1/Ni) suffers a3.56xsteeperspBLEU_thdrop of-2.85(18.58 → 15.73;chrF++_th:-5.39). Adaptive ToPA (1/Nmax) improves High-fertility Thai quality over Vanilla DP-SGD on both8B(16.80 / 32.73vs.15.73 / 30.95,+1.07spBLEU_th/+1.78chrF++_th) and1B(5.45 / 19.14vs.4.78 / 17.39,+0.67spBLEU_th/+1.75chrF++_th).58.31(1.00x) to78.36(1.34x),92.55(1.59x), and93.41(1.60x, saturating at the vocabulary's character/byte sub-token floor for target parameters1.35, 1.70, 2.10). Because the base LLM embeddings are frozen, artificial sub-token sequences lie off the pretrained BPE manifold: unregularized Non-Private fine-tuning on 897 pairs rapidly overfits (PPL:2.61 → 16.31on8B), whereas DP clipping and noise regularize off-manifold drift (PPL:2.54 → 2.64). Comparing DP methods head-to-head on identical fragmented inputs, Adaptive ToPA outperforms Vanilla DP-SGD across all four fragmentation levels on8B(on bothspBLEU_thandchrF++_th). On lower-capacity1B, ToPA leads at canonical1.00x(5.56 / 20.45vs.5.07 / 18.55,3.06xhigher Step SNR) andSplit-1.70(4.28 / 16.23vs.4.25 / 16.20), while off-manifold sub-token gradients push C∈[25.43,26.50] near the Mmax=30.0 grid ceiling—tradingSplit-1.35(4.62 / 17.54vs.4.81 / 17.38, ToPA higher onchrF++_th) andSplit-2.10(4.28 / 16.10vs.4.40 / 16.26) while maintaining 1.48×–1.52× higher Step SNR across the artificial sub-token split levels.