acceptodds
Under review as a conference paper at ICLR 2027

Does Mean Pooling Erase Text?

Abstract

Mean pooling and normalization aggregate a token sequence into one vector, but do they erase its lexical identity? We study reconstruction from native Nemotron-3-Embed-8B passage embeddings, with controlled raw-token pooling across six models as complementary evidence. For a specified fixed-length causal transformer parameter family, we establish generic spherical injectivity in exact arithmetic, without guaranteeing injectivity of trained checkpoints or efficient reconstruction. Under open-weight and victim-query access, our method combines a learned base inverter, target-conditioned prefix guidance, residual-conditioned iterative correction, and victim re-embedding with a persistent candidate archive. Reconstructing eight-token AG-News fragments from native passage embeddings reaches 87.54% token accuracy and 72.60% exact match. On separate eight-token WikiText-2 raw-pooling benchmarks, Nemotron-3 and Jina-v4 reach 94.40% and 97.20% exact match, respectively. Recovery depends strongly on model and domain and deteriorates markedly in the tested 16- and 32-token raw-pooling settings. Candidate-oracle analyses show that most failed reconstructions lack the complete target in the generated archive, while numerical audits show that small residuals need not identify the original text. These findings distinguish architectural identifiability, numerical separation, and computational recoverability, exposing short-text reconstruction risks without implying universal invertibility.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
embedding inversion, text reconstruction, mean pooling, representation learning, spherical injectivity, privacy

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_REzZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies input reconstruction from native Nemotron-3-Embed-8B passage embeddings obtained through mean pooling. It shows that most parameter configurations for several Transformer families map unique inputs to unique mean-pooled embeddings (generic spherical injectivity), without covering trained checkpoints or an efficient reconstruction algorithm. The method presented to reconstruct input from mean-pooled embeddings combines a learned base inverter, target-conditioned prefix guidance, residual-conditioned iterative correction, and victim re-embedding. Empirical evidence shows that 8-token inputs can be reconstructed; analysis of deteriorated reconstruction capabilities at longer input lengths shows that the correct text may never be generated. Analysis of component contributions and numerical ambiguity is also performed.

Critical strengths
  • The proofs presented are insightful
  • The questions are structured well: is the text unique in theory, is it separable numerically, and whether search can find it
  • Modern 1-8B parameter embedders and base LLMs are tested
Critical weaknesses
  • Vec2Text achieves 92% at 32 tokens, but the method presented in this paper gets 0% at 32 tokens for raw pooling on out-of-domain data. This difference makes sense given that the domain, training scale, inverter design, and victim models differ, but it is not clear whether this is because these victims are harder to invert or whether the method itself is weaker.
  • The Vec2Text-style re-implementation is compared only on Llama-2, which is the easiest victim (99.7% EM in Table 1), so it is not clear how the author's method compares to Vec2Text on other, more difficult victims.
  • No hyperparameter search is reported. The paper attributes the length cliff to budget but does not vary the budget at the longer lengths where recovery fails; so this may be a limitation of the configuration and not the method, and the claim may be unsupported.
  • Sampler seeds are varied, but the sampling method itself and its parameters are not. It is unclear what effect the sampling method has on the accuracy of the attack. In addition, the failure analysis says that the right answer is never generated, which may be affected by sampler parameters.
  • The attacker's training seeds are fixed, so it's unclear how much effect training noise has on the accuracy of the attack approach and other results.
Questions for authors
  • Can the authors re-train the original Vec2Text and evaluate on native Nemotron-3 at 8 and 32 tokens?
  • On native Nemotron-3, how does performance change with search budget (rounds, beam) and sampler settings (temp, top-k, etc)?
  • Can Table 3 be repeated with Nemotron-3 or a setting where there is more room for a difference?
  • How would the results vary when the attacker is trained with multiple training seeds?
1 comment

Critical Weaknesses 1:

We agree that the published 92% and our 0% were measured on different victims and data, so the comparison cannot distinguish victim difficulty from differences in training or reconstruction. Our new raw-pooling Nemotron-3 experiments on AG-News identify the immediate bottleneck in our pipeline: at every tested budget, search failed to generate a single complete 32-token target for the final selector to choose (§5.3; Appendix E.6, Table 29). See our reply to Question For Authors 1 below.

Critical Weaknesses 2:

Revised Table 3 extends the comparison to native Nemotron-3 at T=8. On the same AG-News windows and at matched inference GPU time, our method achieved higher exact match than the retrained Vec2Text baseline (§5.4). See our reply to Question For Authors 1 below.

Critical Weaknesses 3:

In new experiments on AG-News, we tested raw-pooling Nemotron-3 at T=16 and T=32 using beam widths of 64 and 128, each with 50 and 100 correction rounds. Doubling both budgets increased T=16 exact recoveries from 1 to 2 of 128 windows and T=32 candidate-oracle token accuracy from 27.12% to 31.84%. T=32 exact match remained zero in all four settings (§5.3; Appendix E.6, Table 29). These results show that budget affects recovery, but not that budget alone explains the decline with length.

Critical Weaknesses 4:

In our new native Nemotron-3 T=8 tests, deterministic token-beam decoding generated the correct sequence on fewer windows than the default sampler. The tested temperature and top‑k changes showed no clear difference from the default. Whenever the correct sequence was generated, it was selected, so the exact-match differences arose during generation rather than final selection. These tests do not rule out better samplers at longer lengths (Appendix E.6, Table 30). See our reply to Question For Authors 2 below.

Critical Weaknesses 5:

We independently retrained the attacker for native Nemotron-3 at T=8 with a second training seed, keeping the inference seed fixed. The two runs had similar exact-match rates on AG-News, although they recovered different individual windows (Appendix E.7, Table 31). See our reply to Question For Authors 4 below.

Questions For Authors 1:

At T=8, we retrained a T5-based Vec2Text adaptation and evaluated both methods on the same native Nemotron-3 observations of the same AG-News windows. At matched inference GPU time (0.40% difference), exact match was 73.30% for our method versus 16.70% for Vec2Text (§5.4, Table 3; Appendix E.3). We do not report a corresponding native T=32 Vec2Text comparison.

Questions For Authors 2:

We conducted new rebuttal experiments on native Nemotron-3 at T=8 using AG-News. With beam width 64, exact match increases from 54.20% to 73.30% to 79.30% at 10, 50 and 100 correction rounds, respectively. In the correction-only control, increasing beam width from 64 to 128 at 50 rounds improves exact match from 70.80% to 73.80% (Appendix E.6, Table 28).

On a paired sampler subset, the default temperature 0.8 and top-k 32 achieves 73.44% exact match. Temperatures 0.6 and 1.0 yield 70.31% and 75.00%, while increasing top-k to 64 yields 74.22%. These temperature/top-k differences are not statistically conclusive. The tested deterministic token-beam variant performs substantially worse, at 42.19% (Appendix E.6, Table 30). Thus, larger search budgets improve recovery, with diminishing gains from additional rounds, whereas the tested temperature and top-k changes offer no clear improvement over the default.

Questions For Authors 3:

Yes. Revised Table 3 now includes the inference-time-matched comparison on native Nemotron-3 at T=8 using AG-News (§5.4; Appendix E.3; results in Question For Authors 1 above). On a separate set of 1,000 WikiText-2 windows using the same native interface and length, exact match was 99.10% for our method versus 57.00% for Vec2Text (Appendix E.5, Table 26). Unlike the AG-News comparison, the WikiText-2 comparison used different beam widths and was not matched for inference time.

Questions For Authors 4:

In a new rebuttal experiment, we independently retrained the base inverter, corrector and Twist with an additional training seed, regenerating their training caches while keeping the target training data and inference settings fixed. On native Nemotron-3 T=8 with AG-News, training seeds 1234 and 2025 achieve 88.65% and 89.50% token accuracy, and 73.30% and 73.70% exact match, respectively. The second fit recovers 75 previously unsuccessful windows but loses 71 previously successful ones, giving a net 0.40-percentage-point exact-match difference (Appendix E.7, “Independent Training Replication”). Thus, aggregate performance is similar across these two independently trained attackers, although individual reconstructions vary. This is a limited replication, not a reliable estimate of training-seed variability across many runs.

Reviewer_bTjAOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
3
Read review
Paper summary

The paper studies whether normalized mean-pooled text embeddings hide the original text. It proves a generic injectivity result for a specified family of causal transformers in exact arithmetic. It also proposes a Vec2Text-style reconstruction framework (base inverter, iterative correction with re-embedding, and a new prefix-guidance module). The main result is 72.60% exact match on eight-token AG-News fragments from native Nemotron-3-Embed-8B embeddings, with raw-pooling controls on six models.

Critical strengths
  • The manuscript explicitly distinguishes architectural identifiability, numerical separation, and computational recoverability, and it scopes its claims carefully.

  • The failure analysis is informative: failed reconstructions lack the target among generated candidates, which locates the bottleneck in proposal generation.

  • Baselines are matched on GPU time, and the defence study reports retrieval utility alongside attack success.

Critical weaknesses
  • Presentation: The paper is very difficult to follow, starting with the abstract (e.g. "generic spherical injectivity in exact arithmetic, without guaranteeing injectivity of trained checkpoints or efficient reconstruction"). Key terms are left undefined, nearly every claim is hedged, and frequent shifts between interfaces, cohorts, and settings obscure the central contribution.
  • Novelty: The core pipeline (target conditioning, residual correction, re-embedding, sequence beams) is inherited from Vec2Text, as the authors acknowledge. The main new component adds 2.4 points of exact match at ~45% extra compute (Table 2), and gains over the Vec2Text-style baseline at matched GPU time are modest and mixed (Table 3).
  • Practical significance: Exact match falls to at most 8.9% at 16 tokens and 0% at 32 tokens, with strong domain sensitivity. Simple Gaussian noise cuts exact match from 69% to 13% with negligible retrieval loss, at least against the limited adaptive attacker tested. The paper does not motivate how eight-token fragments correspond to realistic retrieval deployments.
  • Theoretical contribution: The exact-real result has limited demonstrated relevance to trained, finite-precision systems. The paper admits this gap, which leaves the theory disconnected from the empirical risk claim.
Questions for authors
  1. What does the method contribute beyond Vec2Text, apart from Twist?
  2. Do the 16- and 32-token results improve with larger search budgets, or do they plateau?
  3. What does the injectivity result imply for trained checkpoints?
3 comments

Critical Weaknesses: Presentation

We apologize for the unclear presentation. Given the page limit, we assumed familiarity with the terminology. We revised the abstract for clarity; §3.1 distinguishes embedding settings, and §§3.2–3.3 explain the theorem and its limits.

Critical Weaknesses: Novelty

We acknowledge the Vec2Text lineage. Our contribution extends beyond Twist to the architecture-specific injectivity construction (§3.2; Appendix A) and analyses distinguishing identifiability, numerical separation and reconstruction difficulty (§3.3 and 5.3; Appendix B). Tables 2–3 report cost–accuracy trade-offs, not uniform efficiency superiority.

Critical Weaknesses: Practical significance

In a new rebuttal experiment, native Nemotron-3 at T=32 achieves 38.28% exact match on 128 WikiText windows versus 0% on 128 AG-News windows (§5.3; Appendix E.5, Table 27). This extends recovery beyond eight tokens but confirms domain sensitivity. Training and cohorts differ from the raw-pooling experiments, so these results do not isolate an interface effect. We embedded the public passage prefix and the target window together; we did not recover an eight-token fragment from an embedding of a longer document (§3.1). Recovery of retrieval-length passages remains unestablished (§6).

Our new raw-pooling budget study shows that doubling beam width and correction rounds improves partial T=32 recovery without restoring exact reconstruction (§5.3; Appendix E.6, Table 29). A separate new native Nemotron-3 T=8 comparison on AG-News achieves 73.30% exact match versus 16.70% for the T5-based Vec2Text adaptation at matched inference time (§5.4, Table 3; Appendix E.3). Attacker capacity and offline training costs are not matched. The submitted Gaussian-noise study demonstrates mitigation under the tested conditions; it does not establish information erasure or protection against stronger adaptive attacks (§5.5; Appendix E.4).

Critical Weaknesses: Theoretical contribution

Our theorem retains the specified architectures’ attention and normalization operations and shows, in exact arithmetic, that pooling need not erase text identity for six model families under specified raw-token settings (Appendices A.4–A.5). Appendix A.6 checks constructed witnesses, not pretrained checkpoints.

Section 3.3 and Appendix B.1 give sufficient selection conditions when the target is present, but their required uniform error bounds are not established by the trained-model diagnostics in Appendix B.2. The privacy-risk claim therefore rests on measured reconstruction under the stated access conditions (§5).

Questions For Authors 1:

We build on Vec2Text’s correction mechanism; our contributions beyond Twist are theoretical and empirical. We evaluate six victim models spanning causal language models and bidirectional LLM-based embedders, including Nemotron-3, GritLM, Jina-v4 and Llama-3.1, with native Nemotron-3 passage embeddings as the primary application and complementary raw-pooling controls (§5.2). We provide an architecture-specific explanation of why normalized mean pooling need not erase text identity, accounting for RoPE, normalization, grouped-query attention and bidirectional attention under the specified raw-token assumptions (§3.2; Appendix A). Finally, our length, domain, numerical and candidate-oracle analyses distinguish missing complete proposals from selection failures, explaining why unsuccessful reconstruction does not itself establish that the representation has lost the text (§3.3 and 5.3; Appendices B.2–B.3).

Questions For Authors 2:

New experiments conducted for this rebuttal show that larger budgets improve partial recovery, but do not resolve long-sequence exact reconstruction. Using AG-News evaluation data and fixed Nemotron-3 raw-pooling models, increasing correction rounds from 50 to 100 at beam 64 raises T=16 token accuracy from 17.43% to 21.09% and exact match from 0.78% to 1.56%. At T=32, token accuracy rises from 3.69% to 4.76% and candidate-oracle token accuracy from 27.12% to 29.54%, while exact match remains zero. Increasing the beam to 128 at 100 rounds raises T=16 token accuracy to 22.95% and T=32 oracle token accuracy to 31.84%, but exact match remains 1.56% and 0%, respectively (§5.3; Appendix E.6, Table 29). Thus, partial recovery has not completely plateaued, but increasing these search budgets does not overcome the missing-complete-proposal bottleneck.

Questions For Authors 3:

The result shows that pooling does not force collisions; it does not show that a particular trained checkpoint is collision-free. For the unprompted, fixed-length inputs covered by §3.2, almost every parameter setting gives distinct texts distinct pooled vectors in exact arithmetic. As Appendix A.3 explains, random initialization alone does not establish this for the weights reached after training. For trained checkpoints, we therefore examine numerical separation (§3.3) and measure actual text recovery (§5).

I thank the authors for their detailed response and additional experiments. While some of my concerns remain, I recognize that the extensive experiments and analyses presented in the paper, including the appendix, constitute a valuable scientific contribution. For those reasons, I am increasing my overall score from 2 to 3.

The abstract is now much easier to follow, and the overall readability of the paper has improved. I have therefore increased my Presentation score from 1 to 2.

My remaining concerns are:

  • Novelty: The additional Twist component does not appear to provide a clear advantage at comparable computational cost. In particular, Table 28 shows that allocating approximately the same GPU time to additional correction rounds without Twist achieves similar results than the full pipeline.
  • Practical applicability: Reconstruction remains highly sensitive to the input domain and sequence length. Increasing the computational budget provides only limited improvements for longer sequences, suggesting that further methodological advances are needed for reliable reconstruction of longer texts. Nevertheless, I find this an interesting insight that future work could build upon.

Finally, regarding the theoretical contribution, I agree with the authors on its scope and acknowledge that its limitations are clearly stated throughout the paper.

Thank you for your time and feedback. We appreciate your decision to raise both ratings. Your points about Twist at comparable inference time (Table 28) and the sensitivity of reconstruction to domain and sequence length are well taken.

Reviewer_rcm5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies textual encoders and whether mean pooling removes information about the original text from language model embeddings. The authors provide a theoretical analysis showing that, under certain assumptions and in exact arithmetic, different token sequences of a fixed length produce different normalized mean-pooled embeddings for almost all parameter settings of the specified architectures. Although mean pooling itself ignores order, the contextual token representations can already encode it. The paper also investigates whether the original text can be recovered in practice, using a learned inverter, prefix guidance and Vec2Text-style iterative correction. When the model weights are available, the authors recover short eight-token sequences (e.g., 72.6% exact match from native Nemotron-3-Embed-8B embeddings), but recovery largely fails for sequences of 16-32 tokens.

Critical strengths
  1. The paper presents a theoretical analysis of when normalized mean pooling preserves sequence identity (generic injectivity).
  2. The authors clearly distinguish theoretical identifiability, numerical separation and practical recovery. This helps explain why an embedding may uniquely represent a sequence even when recovering that sequence is difficult.
  3. The experiments cover six victim models with different architectures (causal and bidirectional attention, grouped-query attention), and the explicit theoretical constructions cover the same architecture families.
Critical weaknesses
  1. Practical recovery works mainly for 8-token windows. At 16 tokens, exact match is only 0.2-8.9%, and at 32 tokens no window is recovered exactly, with oracle token accuracy of only 27-33%. Real passages in retrieval systems are usually hundreds of tokens long, so the demonstrated privacy risk is limited to very short fragments. Vec2Text reported 92% exact match for 32-token inputs on a different embedder, so it is unclear whether the drop comes from the embedders, from the method, or from the training and search budget.
  2. The attack assumes the model weights, local query access (including hidden-state sums of candidate prefixes), which makes it applicable only for the open-source models, and not APIs.
  3. The attack is very expensive, and the new component adds little. One 8-token window needs about 50,000 victim forward passes and 60-85 GPU-seconds. The proposed prefix guidance (Twist) adds only 2.4 points of exact match on Nemotron-3 and 0.5 on Llama-2, at the cost of about 45,000 extra victim queries per window. Most of the gain comes from the Vec2Text-style corrector, which the authors themselves describe as inherited.
Questions for authors
  1. Your method seems to be domain-specific. Did you try launching and evaluating your attacks on completely different domains?
  2. Why does recovery fail for longer sequences? Is the main limitation the search algorithm or numerical precision?
  3. How does the attack compare with Vec2Text under the same model, access conditions, and computational budget?
3 comments

Critical Weaknesses 1:

In a new rebuttal experiment on native Nemotron-3, we exactly recovered 49 of 128 WikiText-2 windows (38.28%) at 32 tokens, but none of 128 AG-News windows. In every failed case on both datasets, the complete original was absent from the generated candidates. Even perfect final selection among those candidates could not have recovered another window (§5.3; Appendix E.5, Table 27). Recovery therefore extends beyond eight tokens in the tested WikiText-2 setting, but we do not establish recovery of typical retrieval-length passages.

In a separate raw-pooling study on 128 AG-News windows, doubling both beam width and correction rounds raised 32-token candidate-oracle token accuracy from 27.12% to 31.84%, but no tested configuration generated a complete target (§5.3; Appendix E.6, Table 29). The native and raw studies differ in training and evaluation cohorts, so their results do not isolate the effect of the embedding interface.

Vec2Text’s published 92% result uses a different embedder and experimental setting. Our new same-victim, inference-time-matched native Nemotron-3 T=8 comparison achieves 73.30% exact match versus 16.70% for a separately trained T5-based Vec2Text adaptation (§5.4, Table 3; Appendix E.3). Attacker capacity and offline training are not matched, and relative performance at T=32 remains unresolved.

Critical Weaknesses 2:

We agree. Our threat model explicitly assumes open-weight, query-enabled access, including candidate-prefix hidden states (§3.1); we do not claim that the full pipeline works against embedding-only APIs. This setting remains relevant when embeddings are produced by publicly available models.

Critical Weaknesses 3:

We acknowledge the computational cost and the corrector’s Vec2Text lineage. For the native reference run, 48,604.6 forwarded victim sequences correspond to 759.45 batched model calls per window (Appendix E.3). Twist improves accuracy at fixed correction rounds, but these gains do not establish a matched-time advantage: native Nemotron-3 achieves 75.90% exact match without Twist at 75 rounds versus 73.30% with Twist at 50 rounds, taking 76.85 and 74.34 seconds per window, respectively (Appendix E.3, “Comparable-time control without Twist”; Appendix E.6, Table 28).

Questions For Authors 1:

Yes. Our native Nemotron-3 attacker is trained on WikiText-2 and evaluated without domain-specific retraining on AG-News, achieving 72.60% exact reconstruction (Sections 5.1–5.2, Table 1). This demonstrates transfer from encyclopedic text to news. We acknowledge that this is cross-corpus natural-language evidence, not a guarantee of generalization to arbitrary domains.

Questions For Authors 2:

The evidence points to a candidate-generation bottleneck under the tested training/search budgets (Section 5.3). On 128 raw Nemotron-3 T=32 windows, increasing correction rounds from 50 to 100 at beam 64 raises oracle token accuracy from 27.12% to 29.54%; increasing the beam to 128 at 100 rounds raises it to 31.84%. Selected and oracle exact match remain zero: complete targets are still absent from the generated candidates (Appendix E.6, Table 29).

To test numerical verification separately, we added the original sequence to 32 incorrect candidates for eight development windows at each of T=16 and T=32. Against the unchanged stored target, the original ranked strictly first in all tested executions (Appendix B.2, “Longer-length fixed-target probes”). It was never supplied during reconstruction. These probes find no final-ranking failure within the tested pools, but do not rule out numerical effects during generation or pruning.

Questions For Authors 3:

Table 3 (Section 5.4; Appendix E.3) compares our method with a local Vec2Text-style baseline on Llama-2 T=8, using the same victim, observations, evaluation windows and access assumptions, with inference GPU time matched within 5%. At approximately 31 and 84 GPU-seconds per window, our exact match is 98.80% and 99.60%, versus 92.60% and 98.60% for the baseline. At the lowest budget, the baseline performs better: 92.60% versus 90.40%.

In the revised paper (§5.4, Table 3; Appendix E.3), we additionally report a native Nemotron-3 T=8 comparison under the same observation and permitted-access conditions. On the same AG-News evaluation windows, our pipeline achieves 73.30% exact match versus 16.70% for the retrained T5-based Vec2Text adaptation, at matched inference GPU time: 74.34 versus 74.64 seconds per window, a 0.40% difference. Only inference time was matched; victim-query count, attacker model capacity and offline training cost were not.

Thank you for the new experiments. The comparison with Vec2Text and the results on longer texts answered most of my questions. I still have one concern about Twist, on the eight-token AG-News examples, the version without Twist did better in about the same time. Overall, I will raise my score to weak accept.

Thank you for your time and feedback. We agree that on eight-token AG-News windows with native Nemotron-3, omitting Twist allowed 75 correction rounds rather than 50 at comparable inference time, and this gave higher exact match (75.90% vs. 73.30%; Appendix E.3). At beam width 64, adding Twist raised exact match from 70.80% to 73.30% at 50 rounds and from 75.90% to 77.10% at 75 rounds (Appendix E.6, Table 28).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.