- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
There are several sources of instability that arise from sample-and-aggregate classification ICL. The ordering, partition, and privacy noise all affect the classification output. The paper investigates how increasing privacy noise masks the instability arising from the other two (ordering and partition).
The paper uses a sampled demonstration set that is partitioned and fed as a context to a frozen LLM. This produces vote counts for each class in the classification task. Afterwards, the final classification is released by a Gumbel report-noisy-max (RNM) mechanism, which is an exponential-mechanism variant with a closed-form distribution. The tight Gumbel privacy masks upstream order and partition effects. Under strong privacy, the output distribution of this mechanism approaches the uniform distribution.
Critical strengths
- The explanatory framework for the privacy masking of the upstream order and partition effects is interesting.
- The explanatory framework is well supported by empirical evidence, and the diversity of models and datasets are reasonable.
- The authors performed careful evaluation audits including tie-floor and exact order reconstruction checks
- The paper also presents interesting experiments showing the utility advantage of using more smaller voters at intermediate privacy levels.
Critical weaknesses
- The explanatory framework is novel; however, the mechanisms used are not novel, and the paper only considers one type of ICL which is LLMs. For example, tabular foundation models might be sensitive to partitioning but a permutation-invariant architecture might not be sensitive to the ordering of samples within each partition. For these reasons, I believe that the paper's contributions are limited.
- Proposition 2 decomposes the total prediction instability by first conditionon on , then ordering conditioned on the partition . However, section 3 describes the pipeline as proceeding from ordering to partitioning , and then to private release. Both yield valid decompositions through the law of total variance but they can assign different contributions to ordering and partitioning.
- The fixed-record experiment demonstrating the transmission window uses only one model-dataset combination, providing limited evidence that this phenomenon generalizes.
Questions for authors
- Would the same explanatory framework apply to other ICL methods, such as permutation-invariant tabular foundation models?
- Can the authors clarify why this particular decomposition is used and why it is not consistently described in the paper? Also, whether using one over the other would affect the paper's conclusions?
- Would the transmission window conclusion generalize beyond the single model-dataset combination?
4 comments
Applicability beyond LLM-based ICL. The attribution identity is not specific to a language model. It applies to categorical releases from internal vote counts under the stated execution model. For a predictor whose votes are exactly invariant to demonstration order, R_ord = 0 and G = R_priv + R_part. Partitioning can still matter. Retaining the same full-domain Gumbel channel also retains the tight-privacy limit and count-gap bounds; a finite crossover depends on partition variation and the tie floor. This theoretical applicability is distinct from empirical validation on tabular foundation models, which we have not performed (§3.4, p. 6).
We agree that the release mechanism itself is established. The contribution is the attribution analysis and the measured finite-budget consequences for LLM-based sample-and-aggregate classification, not a new private-ICL mechanism.
Evidence beyond the original transmission-window condition. To test the fixed-72-record result on a second model, we evaluated Mistral-7B-v0.3 on Banking77 during discussion. It uses the same 100 queries, labels and ordered 72-record pools as the Qwen comparison, five partitions, and the same endpoints and paired query bootstrap. No pass/fail threshold was set for this run.
At ε = 1, replacing eight nine-example voters with 24 three-example voters gives:
| Endpoint | Eight voters | 24 voters | Difference [95% interval] |
|---|---|---|---|
| Released accuracy | .222 | .542 | +.320 [.256, .384] |
| Plurality fidelity | .301 | .960 | +.660 [.639, .681] |
| Pre-release plurality accuracy | .602 | .545 | −.056 [−.098, −.021] |
Pre-release accuracy is a secondary endpoint. Its decrease shows that the released-accuracy gain does not require higher average plurality accuracy before noise (§5.3, p. 9; Table 18).
The curve shape does not replicate. In the post-hoc Mistral replay, the accuracy difference is −.014 [−.049, .017] at ε = 2 and −.058 [−.100, −.023] at ε = 4 (pointwise intervals; Table 19). Qwen's advantage instead becomes small at the deterministic reference. We therefore use “transmission window” for budgets where more voters improve mean released accuracy, without asserting a common shape or an inferentially established window boundary. The 72-record result now covers two models, still on one task.
Complementary cross-task evidence. The separately frozen n = 24 panel covers Mistral-7B and Qwen2.5-1.5B on DBpedia and TREC-fine. Each subset casts its most frequent prediction across all within-subset orders, so this is a different voting protocol. Comparing (S, s) = (24, 1) with (8, 3), released accuracy increases at ε = 1 in all four conditions (+.036 to +.139), then decreases by ε = 8 in post-hoc replay (Table 3). TREC-fine endpoints were primary; DBpedia was descriptive. This supports budget-dependent allocation trade-offs across tasks, but is not a replication of the Banking77 curve. Section 5.3 and §7 now state these distinctions directly.
Thank you for the detailed responses and revisions. I appreciate the effort invested in addressing my concerns.
I believe that the authors have addressed the conditioning-order issue and strengthened the empirical evidence.
However, my main concern remains the significance of the central contribution. The observation that stronger privacy masks upstream variation follows directly from the behavior of the well-known exponential mechanism.
Therefore, while I acknowledge the improvements in the revisions, I will maintain my original rating of 2.
Thank you for reading the revision and for confirming that the conditioning-order issue is resolved. We agree that masking as ε → 0 follows directly from the exponential mechanism. Our intended contribution is the empirical characterization of configuration choices at finite privacy budgets. With the same 72 private records and ε = 1, switching from 8 voters with 9 demonstrations each to 24 voters with 3 demonstrations each increased released accuracy by 37.9 percentage points for Qwen2.5-7B and 32.0 points for Mistral-7B. In the Mistral experiment, this improvement occurred despite lower pre-noise plurality accuracy. The limiting argument alone does not determine the magnitude or practical consequences of this trade-off. We respect your assessment of its significance and appreciate your constructive review.
Thank you for identifying the inconsistency between the pipeline description and the conditioning order. We have corrected the exposition and made the attribution convention explicit.
Why partition first, and what changes under reversal? The private pipeline partitions the pool, orders examples within each subset, collects one vote per subset, and releases a private label. The revised setup and Figure 1 caption state this order. Section 3.1 is now “Full-context diagnostic: signed margins”; that diagnostic is separate from the voting pipeline.
For the same joint law of partition Π, order O and release probabilities q, reversing the conditioning leaves two quantities unchanged:
Consequently their crossover is unchanged. The attribution of upstream variation to order versus partition can change. With independent population randomizers, write the upstream variance as A_Π + A_O + I, where A_Π and A_O are main effects and I is their interaction. Partition-first attribution gives R_part = A_Π and R_ord = A_O + I; reversing the conditioning gives R_ord′ = A_O and R_part′ = A_Π + I (Appendix D.1, Lemma 4).
The sampler coupling matters. Our pool is uniformly shuffled and sliced into ordered blocks, represented by Π. An independent O specifies a position permutation within each block. Fixing O maps the uniform whole-pool shuffle bijectively onto another uniform shuffle. Thus E[q | O] = E q and the population order main effect A_O is zero under this coupling. Intuitively, a position permutation acts on different examples in different partitions. Sorting each block's initial arrangement can break this symmetry; the appendix includes a toy counterexample.
This is not a statement that order-sensitive models have no order effect. Partition-first R_ord measures variation when membership is held fixed and the examples are reordered. It includes the interaction and answers the fixed-subset question we study. An order-first attribution answers a different question and, under this coupling, assigns the interaction to partition instead. Our order-dominance claims therefore concern the partition-first estimand, not a convention-independent ranking of the two factors.
The zero-main-effect statement is about the population sampler, not a finite empirical estimate. The logged schedules are generated independently for different partitions and subsets; shared schedule indices do not form a crossed partition-by-order table. We do not treat them as such.
Where this is now stated. Section 3 opens with the sampling order and attribution convention (p. 3). “Why partition first?” (§3.4, p. 6) explains the choice. Appendix D.1 gives the reverse-conditioning formulas and exact toy enumeration. The privacy term, total-upstream conclusions and crossover survive reversal; the separate order/partition shares are explicitly convention-dependent.