acceptodds
Under review as a conference paper at ICLR 2027

WorldCup Sampling for Multi-bit LLM Watermarking

Abstract

As large language models (LLMs) generate increasingly human-like text, watermarking has emerged as a promising solution for reliable attribution and traceability beyond mere detection. While multi-bit watermarking enables richer provenance encoding, existing approaches typically extend zero-bit schemes through static logit perturbations and hard counting-based decoding, which can degrade text quality and compromise decoding robustness as the payload increases. In this paper, we propose WorldCup, a multi-bit watermarking framework that extends tournament sampling and embeds message bits through hierarchical competition, in which complementary functions guide the selection of candidate tokens. Moreover, WorldCup incorporates entropy-aware modulation to adapt watermark strength to local uncertainty in the model distribution and preserve text quality, while confidence-aware decoding uses agreement across tournament layers to aggregate evidence for robust message recovery. Comprehensive experiments across multiple LLMs and diverse downstream tasks show that WorldCup offers a favorable trade-off among message capacity, detectability, robustness, text quality, and decoding efficiency, consistently outperforming prior baselines.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM watermark, multi-bit, text generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_X5isOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes WorldCup, a multi-bit watermarking framework that embeds message bits directly into tournament sampling during autoregressive generation. It uses complementary g-value functions to guide token selection, entropy-aware modulation to adapt watermark strength to the model’s local uncertainty, and confidence-aware decoding to aggregate evidence across tournament layers. The paper evaluates message recovery, detectability, text quality, downstream task performance, decoding efficiency, and robustness under editing, copy-paste, and paraphrasing attacks.

Critical strengths

The paper builds on SynthID-Text’s zero-bit tournament sampling and existing multi-bit watermarking research, and uses tournament sampling and hash- or pseudorandom-function-based g-value generation. On this basis, it proposes complementary g-values conditioned on message bits and a complementary-distribution subtraction mechanism, q − λq̄. It further extends the method to k-bit encoding, incorporates entropy-aware modulation, and proposes a vectorized computation of the tournament distribution and confidence-aware decoding.

Critical weaknesses
  1. The authors claim in the introduction section that static logit perturbations interfere with natural generation “especially at higher payloads,” degrading text quality and fluency. However, they provide no evidence that increasing the number of watermark bits causes this effect when watermark strength is held constant. In Table 1, PPL does not systematically increase with bit length, so the results do not show the claimed degradation. The authors should support or qualify this claim.

  2. The notation in the core method is ambiguous. Section 3.1 uses m for the message, whose length is b, while Algorithm 1 uses m for the number of tournament layers. The g-function families contain m layer-specific functions, and Figure 1 shows a tournament depth matching the largest subscript of g, suggesting that message length and tournament depth are linked, although the layer count is a hyperparameter independent of message length. The paper also uses boldface and regular-weight g to distinguish a function family from an individual function, a distinction that is visually easy to miss. Together, these ambiguities make the indexing and function selection difficult to follow.

  3. Appendix G.1 establishes score-level separation but notes that the distinguishability of the induced token distributions depends on the base distribution and tournament parameters. The paper does not provide this distribution-level analysis. Section 5.1 reports how removing the subtraction of complementary g-value distributions and replacing complementary g-value functions with randomly sampled ones affect decoding accuracy, but does not quantify the statistical distance between the induced distributions q and q̄. The link between score-level separation and the reported decoding gains therefore remains unclear.

Questions for authors
  1. Could the authors clarify the notation for the message, tournament layers, and g-function families issue mentioned in “weakness” section?
  2. Can the authors quantify the statistical distance between q and q̄ under complementary and randomly sampled g-value functions (e.g., using symmetric KL divergence or MAE), and analyze how it relates to decoding accuracy?
  3. The paper’s decision-margin argument for complementary g-value functions appears to address stochasticity during generation and sampling. It does not explain why WorldCup is robust to the evaluated post-generation attacks, which affect watermark detection and message extraction through different mechanisms. Could the authors provide an attack-specific mechanistic analysis and identify which components account for WorldCup’s advantage over the baselines? In addition, could they report the attack-induced drops in AUC or TPR@1% FPR relative to the matched clean condition for WorldCup and the baselines? This would clarify whether better post-attack performance arises because a method performs better before the attack or because it is more resistant to the attack.
4 comments
AuthorsOpenReview ↗

Response to Reviewer X5is

We thank the reviewer for careful reading and constructive suggestions. We address each concern below.

W1: Payload and text quality

We agree Table 1 shows no consistent increase in perplexity with message length. The phrase “especially at higher payloads” left the role of watermark strength implicit. At a fixed generation length, a longer message has fewer tokens per bit on average. Maintaining or improving bit accuracy may then require stronger logit perturbations, which can reduce fluency and overall text quality. We will make this condition explicit in the introduction.

W2 & Q1: Notation

In Section 3.1, bold m\mathbf{m} denotes the message, while plain mm denotes the tournament depth. The message length bb and the depth mm are chosen independently. As the reviewer notes, we use bold g\mathbf{g} for the ordered family of scoring functions, following the convention for vectors, and plain gℓg_\ell for the individual function at layer ℓ\ell.

For clarity, we will change the message symbol to M\mathcal{M} and keep mm for the tournament depth. We will write the function family for bit value vv as Gv=(g1(v),…,gm(v))\mathcal{G}_v=(g_1^{(v)},\ldots,g_m^{(v)}) and use this notation consistently in Section 3.1, Algorithm 1, and Figure 1.

W3 & Q2: Distribution-level analysis

We measured symmetric KL and MAE between qq and qˉ\bar q, the token distributions for the two bit values. At each of 512 contexts per model, we fixed the base distribution, candidate set, and g0g_0, then compared g1=1−g0g_1=1-g_0 with an independently sampled g1g_1.

Modelgg-value functionsSymmetric KLMAE (×10−6\times 10^{-6})Bit accuracy (%)
Llama3-8BIndependent5.96025.924575.39
Llama3-8BComplementary8.46827.618780.47
Gemma2-9BIndependent6.46893.329671.48
Gemma2-9BComplementary9.20874.017476.46

Complementary functions increased both distance measures relative to independently sampled functions, which coincided with gains in bit accuracy of 5.08 percentage points on Llama3 and 4.98 points on Gemma2. Therefore, greater separation between qq and qˉ\bar q was associated with higher decoding accuracy on both models.

Q3: Robustness analysis

  • Complementary gg-functions strengthen the evidence that survives an attack. Independent functions can give both bit values the same score. WorldCup sets g1=1−g0g_1=1-g_0, so each layer that favors one value disfavors the other. Copy-paste preserves this evidence throughout intact passages. Deletion and substitution leave fewer votes, while paraphrasing or back-translation may preserve a key fact but change its wording. Phrases that retain their wording and local context still carry their original bit evidence. Complementary scoring raised clean bit accuracy from 98.6% to 99.0% (Table 5).
  • Confidence-aware decoding limits the influence of noisy votes. Hard voting gives a narrow majority the same weight as unanimous agreement. WorldCup weights each token by agreement across layers. After copy-paste, inserted tokens with mixed scores contribute little. After rewriting, even a short retained phrase can guide recovery when its tokens give consistent votes and the rewritten tokens give weaker ones. With the same encoder, this raised clean bit accuracy from 97.8% to 99.0% over hard voting (Table 16).

Moreover, we measured each method’s AUC drop in percentage points from its aggregate clean AUC to its AUC under attack. The settings were the same throughout (Llama3-8B, C4, 16 bits, 256 tokens).

AttackWorldCupBiMarkMPACSegMarkStealthInk
Clean AUC1.0001.0000.9960.9950.997
Word deletion0.9946 (↓0.54%)0.9925 (↓0.75%)0.9140 (↓8.20%)0.8678 (↓12.72%)0.7936 (↓20.34%)
DICT substitution0.9901 (↓0.99%)0.9816 (↓1.84%)0.9040 (↓9.20%)0.8500 (↓14.51%)0.7644 (↓23.26%)
BERT substitution0.9942 (↓0.58%)0.9698 (↓3.02%)0.8711 (↓12.49%)0.8503 (↓14.48%)0.7572 (↓23.98%)
Copy-paste0.9989 (↓0.11%)0.9708 (↓2.92%)0.9304 (↓6.56%)0.9849 (↓1.01%)0.9280 (↓6.90%)
DIPPER-10.9743 (↓2.57%)0.9533 (↓4.67%)0.8111 (↓18.49%)0.7518 (↓24.32%)0.7236 (↓27.34%)
DIPPER-20.9834 (↓1.66%)0.9767 (↓2.33%)0.9147 (↓8.13%)0.8829 (↓11.21%)0.8464 (↓15.06%)
Back-translation0.7236 (↓27.64%)0.6311 (↓36.89%)0.5001 (↓49.59%)0.3958 (↓59.92%)0.4820 (↓51.50%)
GPT-4o paraphrasing0.8500 (↓15.01%)0.8257 (↓17.43%)0.6725 (↓32.35%)0.5770 (↓41.80%)0.6523 (↓34.47%)
Mean (8 attacks)0.9386 (↓6.14%)0.9127 (↓8.73%)0.8147 (↓18.13%)0.7700 (↓22.50%)0.7434 (↓25.36%)

Although WorldCup and BiMark had the same reported clean AUC of 1.000, WorldCup lost less AUC under every attack, averaging 6.14 points versus 8.73 points for BiMark. Therefore, WorldCup’s post-attack advantage reflects greater robustness to all eight attacks, not simply better pre-attack performance.

As the rebuttal deadline approaches, we welcome further discussion, which would help us promptly address any remaining questions. If our responses have addressed your concerns, we hope you will take them into account in your final assessment, and we sincerely appreciate your time and feedback.

Thank you for addressing the questions raised in the review. I would like to keep the rating unchanged.

Thank you for taking the time to consider our responses. We sincerely appreciate your thoughtful feedback and support for our work. We wish you all the best!

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

This paper presents WorldCup, a multi-bit watermarking framework for LLM-generated text. Building on SynthID’s tournament sampling, it uses message-dependent complementary scoring functions. Multiple independent function groups encode several bits through an additive–subtractive combination of their induced token distributions, with entropy-dependent modulation controlling the subtraction strength. The decoder reconstructs message positions from context and aggregates soft evidence across tokens and tournament layers. The evaluation covers bit recovery, detection, text quality, robustness, and computational cost across several LLMs and generation tasks. The reported results show improved bit accuracy in many settings, supported by component ablations, downstream-task evaluations, and Hamming-decoded message recovery results.

Critical strengths

(1) Encouraging empirical results in bit recovery.

In Table 1, for LLaMA3 with 48-bit messages and maximum lengths of 128 and 256 tokens, WorldCup reports bit accuracies of 0.862 and 0.916, compared with the strongest listed baselines at 0.828 and 0.900, respectively. In Table 2, WorldCup has the highest reported bit accuracy in all twelve model–task combinations. These results suggest that the complete method offers useful benefits under the evaluated configurations.

(2) A helpful controlled comparison of decoding strategies.

Table 16 compares decoders using the same WorldCup encoder. For k=2 and 48-bit messages, confidence-aware decoding improves bit accuracy from 0.886 to 0.925, with PPL 11.44 for both. This supports the benefit of retaining soft evidence rather than making token-level hard decisions. The evaluation of whole-message recovery and attacks also helps readers understand the method’s operating range.

Critical weaknesses

(1) The relationship to the closest prior work would benefit from further clarification.

MirrorMark, already cited in Appendix A, includes multi-bit tournament sampling (§3.3.2), complementary one-bit mappings u and 1-u (§3.1, Eq. 10), and soft score-based message decoding. WorldCup’s multi-group distribution subtraction and entropy modulation are potentially useful differences. A more explicit discussion of these differences, together with a matched comparison or a suitable controlled alternative, would help establish their added value. The current results and random-function ablation provide useful evidence, but leave the added value over this closely related approach unresolved. This is the main reason for my current recommendation.

(2)The source and scope of the recovery–quality gains could be characterized more clearly.

The complete method improves bit accuracy, and the controlled decoder comparison is informative. For entropy modulation, however, the current ablation also changes effective watermark strength, which makes the specific role of adaptivity less clear. A comparison that better separates these effects would help explain the contribution of this mechanism.

Tables 2 and 4 also indicate that improved bit accuracy is not always accompanied by preserved downstream-task quality or reliable whole-message recovery at high payloads. Strong rewriting remains a challenging setting as well. The results support a favorable trade-off in several evaluated settings, while the broader claim of robust, high-capacity watermarking without quality loss remains only partially supported. A clearer account of these operating conditions, together with comparisons under comparable quality and payload constraints, would help establish where the claimed advantages hold.

Questions for authors

(1)Could the authors clarify the main additional insight or capability of WorldCup relative to MirrorMark, particularly given their shared use of multi-bit tournament sampling and complementary message hypotheses? Which results most directly illustrate the added value of WorldCup’s distinctive components?

(2)How much of the observed improvement can be attributed to the proposed encoding mechanisms, as opposed to a different watermark-strength setting? It would be helpful to explain where entropy-aware modulation offers an advantage over adjusting the overall strength alone.

(3) Tables 5 and 11 provide useful sensitivity analyses. Since Section 5.1 defines “w/o entropy” using fixed $\lambda =1 $ , the comparison changes both adaptivity and effective strength. Could the authors comment on whether the recovery–quality benefit would remain against a tuned constant-λ\lambda baseline or a strength-matched control? Such a comparison would help clarify the role of adaptivity.

(4) Could the authors discuss the conditions under which the gains in bit accuracy also lead to reliable whole-message recovery while maintaining downstream-task quality? In particular, a clearer account of the benefits and limitations at high payloads and under strong rewriting would be helpful.

5 comments
AuthorsOpenReview ↗

Response to Reviewer mode (1/2)

We thank the reviewer for the careful assessment and for recognizing the empirical improvements and the controlled decoder comparison. We address each concern below.

W1 & Q1: Comparison with MirrorMark

We thank the reviewer for pointing to MirrorMark, a concurrent work that draws on the intuition of complementarity to develop a common mirroring rule for SynthID and AA. WorldCup focuses on how complementary tournament distributions can encode several bits jointly, adapt watermark strength to token entropy, and support bitwise confidence-aware decoding. We analyze this complementarity both theoretically and empirically.

  • Multi-group distribution composition. MirrorMark assigns a distinct reflection of the sampling randomness to each kk-bit symbol. WorldCup uses one complementary pair of gg-function families per bit. It adds the tournament distributions for the intended bit values and subtracts a weighted sum of those for the opposite values. On C4 with k=2k=2 and a 48-bit payload, although subtraction slightly increases PPL, it improves bit accuracy by 12.8 percentage points, from 0.802 to 0.930 (Table 5).

  • Entropy-aware modulation. MirrorMark’s reflections preserve the base sampler’s distribution. WorldCup applies stronger subtraction when next-token entropy is high and weaker subtraction when it is low. In the same setting, replacing fixed subtraction (λ=1\lambda=1) with entropy-aware subtraction lowers PPL from 27.00 to 11.06 (Table 5).

  • Bitwise confidence-aware decoding. MirrorMark aggregates soft evidence for each of the 2k2^k candidate symbols and selects the highest-scoring one. WorldCup aggregates soft evidence across tournament layers and tokens separately for each bit, so decoding requires only kk binary comparisons. With the same WorldCup encoder, retaining soft evidence raises bit accuracy from 0.886 with hard counting to 0.925, with PPL unchanged (Table 16).

These ablations quantify the contribution of WorldCup’s components. We will include this comparison in the revision and conduct a matched evaluation when MirrorMark’s implementation becomes publicly available.

W2 & Q2: Encoding mechanisms and watermark strength

(1) Contribution of the encoding mechanism. Complementary scoring functions make the two bit values easier to distinguish, and entropy-aware modulation helps preserve text quality. In the 48-bit, k=1k=1 ablation, replacing random functions with complementary ones at the same overall watermark strength setting raises bit accuracy from 79.7% to 88.0% (+8.3 percentage points) without increasing PPL. At k=2k=2, replacing fixed subtraction (λ=1\lambda=1) with entropy-aware subtraction lowers PPL from 27.00 to 11.06 while retaining 93.0% bit accuracy (Table 5).

(2) When entropy-aware modulation helps. A larger payload or a shorter text leaves fewer tokens to carry evidence for each bit. Table 1 illustrates this on LLaMA3-8B. With a 128-token limit, increasing the payload from 16 to 48 bits lowers WorldCup’s bit accuracy from 98.2% to 86.2%. Extending the limit to 256 tokens brings 48-bit accuracy back to 91.6%. When many bits must fit into a short text and entropy varies across positions, increasing the overall strength also perturbs highly predictable tokens. Entropy-aware modulation keeps subtraction weak at those positions and strengthens it where several continuations are plausible, so those positions can carry more watermark evidence.

AuthorsOpenReview ↗

Response to Reviewer mode (2/2)

W2 & Q3: Entropy-aware modulation

We fixed λ=1\lambda=1 in Table 5 to ablate entropy-aware modulation, which adjusts watermark strength to the model’s uncertainty at each token. For 16-bit messages on Llama3-8B, enabling modulation reduced PPL from 28.38 to 11.25 while retaining 99.0% bit accuracy (99.6% with fixed λ=1\lambda=1). The same entropy-aware rule can adapt to each context without retuning its global scale. In Table 11, it achieved 91.8% bit accuracy on Llama3-8B and 86.9% on Gemma2-9B at 48 bits, with PPLs of 10.94 and 8.50, respectively. We also compared the adaptive rule with a quality-tuned constant (λQ\lambda_Q) and a constant calibrated to match the adaptive rule’s average watermark strength (λS\lambda_S).

BackboneMethodBit Acc (%) ↑PPL ↓
Llama3-8BAdaptive87.975.922
λQ=0.5\lambda_Q=0.581.045.062
λS=0.691\lambda_S=0.69186.465.532
λ=1\lambda=193.7513.82
Gemma2-9BAdaptive83.805.164
λQ=0.5\lambda_Q=0.582.604.393
λS=0.627\lambda_S=0.62782.604.260
λ=1\lambda=194.6912.30

A tuned constant may achieve a similar trade-off on a given model and dataset, but its best value may change across settings. Under the same strength matching criterion, the selected coefficient was 0.691 on Llama3-8B and 0.627 on Gemma2-9B. Using a single global scale across both models therefore illustrates the practical advantage of entropy-aware modulation.

W2 & Q4: Operating conditions

Whole-message recovery. At a fixed generation length, shorter payloads leave more tokens to support each message position. With a 256-token limit and a 16-bit codeword, WorldCup recovers all 11 information bits after Hamming decoding in 99.0% of cases on LLaMA3 and 96.5% on Gemma2 (Table 4). With a 48-bit codeword carrying 42 information bits, exact recovery falls to 31.0% and 13.0%, respectively. At this higher payload, even a few residual bit errors can prevent recovery of the full message.

Downstream quality. We expect a more favorable trade-off on tasks with more high-entropy positions, where the encoder can embed a stronger signal while choosing among several plausible next tokens. WorldCup achieves the highest bit accuracy in all 12 model and task combinations in Table 2. On Ministral long-form QA, it reaches 90.5% bit accuracy with a QA score of 5.21, close to the unwatermarked score of 5.25. Quality still varies by task. On LLaMA3.1 math reasoning, for example, bit accuracy reaches 94.7%, while Pass@1 falls from 0.500 to 0.445.

Strong rewriting. Under word-level edits and copy-paste attacks, surviving passages retain local context that the decoder can use to recover individual bits. Strong paraphrasing and back-translation can change that context even when the meaning is preserved. The decoder may then assign tokens to different message positions, making whole-message recovery unreliable.

As the rebuttal deadline approaches, we welcome further discussion, which would help us promptly address any remaining questions. If our responses have addressed your concerns, we hope you will take them into account in your final assessment, and we sincerely appreciate your time and feedback.

Reviewer_7vpzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors propose WorldCup, a new watermarking framework based on multi-bit watermarking. Unlike conventional zero-bit schemes that implicitly embed the message into the generation process, WorldCup incorporates the message into tournament sampling to guide competition among candidate tokens. Based on the proposed approach, the authors expect to solve the problem caused by static logit perturbations used to inject watermark signals, which harm the original model distribution. Experimental results show that WorldCup improves watermark detectability while maintaining the balance of other factors.

Critical strengths
  • This work tries to solve the problem caused by static logit perturbations used to inject watermark signals, which harms the original model distribution.
  • The authors extend tournament sampling to achieve this objective.
  • Additionally, the proposed entropy-aware modulation mechanism can balance detectability and fluency.
  • Using soft confidence-aware aggregation that uses agreement across tournament layers as token-level evidence can mitigate the disproportionate influence of low-entropy tokens.
  • The proposed method can mitigate information allocation when the message length grows.
  • Experimental results across LLMs and downstream tasks show that the proposed method, WorldCup, outperforms baselines in bit accuracy while keeping the balance among multi-bit capacity, watermark detectability, robustness, text quality, and decoding efficiency.
  • In addition to the intuitive approaches, the authors provide theoretical background for their approach.
Critical weaknesses
  • In text summarization and machine translation, traditional surface-based matching metrics like BLEU and ROUGE-L are used. These metrics are difficult to use for assessing contextual appropriateness compared with recent neural-based metrics like COMET and BERTScore.
  • Thinking about the fact that tournament-based sampling samples text more times than the baselines, the authors need to compare their approach to the baselines with the same sampling size or beam decoding.
Questions for authors
  • How did you calculate the PPL for each LLM output? To avoid bias from using the same models to calculate PPL, you should use an LLM other than LLaMA3 or Gemma2.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 7vpz

We thank the reviewer for the constructive feedback and for recognizing the contributions of our method. We address the concerns about evaluation metrics, sampling budgets, and perplexity computation below.

W1: Additional metrics

To complement BLEU and ROUGE-L, we evaluated translation quality on WMT using BERTScore and COMET, and summarization quality on CNN/DailyMail using BERTScore. We also report watermark detection F1 and bit accuracy.

ModelMethodWMT F1WMT Bit AccWMT BERTScoreWMT COMETCNN/DM F1CNN/DM Bit AccCNN/DM BERTScore
LLaMA3.1-8B-ITUnwatermarked--0.6980.827--0.227
MPAC0.6920.7880.6510.8020.8300.8350.215
BiMark0.6980.5350.6830.8190.9360.7640.219
WorldCup (ours)0.8000.8100.6850.8260.9610.8780.221
Gemma2-9B-ITUnwatermarked--0.8000.875--0.245
MPAC0.6680.6830.7910.8730.6820.5890.230
BiMark0.6740.6680.7930.8740.6720.4450.231
WorldCup (ours)0.6880.6930.7970.8740.7660.6020.233

The results show that WorldCup achieves higher BERTScore, detection F1, and bit accuracy than baselines in all four settings. We will include these results and the metric configurations in the revised manuscript.

W2: Sampling budget

We appreciate the opportunity to clarify this point. WorldCup and the baselines use the same sampling budget. Within each experimental setting, all methods use the same prompts, number of continuations per prompt, and output length limits.

The tournament diagram shows how candidate tokens compete at a single decoding step. For a given prefix, Eq. (3) defines the probability that each token wins. We compute these probabilities directly from the model’s next token distribution using the vectorized formulation in Appendix G.2 (Eq. (21) for one layer and Eq. (30) for multiple layers).

For generalized WorldCup, we combine the component distributions for the same prefix using Eq. (5), then clip and normalize the resulting scores as described in Section 3.3. We draw one next token from the final distribution. This requires no additional candidate continuations, beam search, or reranking.

Q1: PPL computation

As stated in Section 4.1 (lines 360–361), we use Vicuna-13B-v1.5 as a fixed external evaluator for outputs from both LLaMA3 and Gemma2. For each output, we exclude the original prompt and tokenize the generated continuation with the Vicuna tokenizer, without adding special tokens. For continuation tokens y1,…,yLy_1,\ldots,y_L, Appendix D defines perplexity as

PPLϕ(y)=exp⁡(−1L∑t=1Llog⁡pϕ(yt∣y<t))PPL_{\phi}(y)=\exp\left(-\frac{1}{L}\sum_{t=1}^{L}\log p_{\phi}(y_t\mid y_{<t})\right)

where pϕp_{\phi} is Vicuna’s next token distribution. We apply the same procedure to all methods and report PPL across continuations in each experimental setting.

As the rebuttal deadline approaches, we welcome further discussion, which would help us promptly address any remaining questions. If our responses have addressed your concerns, we hope you will take them into account in your final assessment, and we sincerely appreciate your time and feedback.

Dear Reviewer,

With about 12 hours remaining in the rebuttal period, we would be grateful if you could take a moment to review our responses. Please let us know if any concerns remain, as we would be happy to provide further clarification before the deadline. Thank you again for your time and thoughtful feedback.

Best wishes,

The Authors

Dear Reviewer,

With about 1 hours remaining in the rebuttal period, we would be grateful if you could take a moment to review our responses. Thank you again for your time and thoughtful feedback.

Best wishes,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.