← All posts

Green-List Red-List Evasion Techniques Ranked by Effectiveness

Ranked by z-score impact, BIRA leads single-rewriter methods for beating KGW.

November 16, 2026

Cover illustration for “Green-List Red-List Evasion Techniques Ranked by Effectiveness”

KGW watermark evasion is best ranked by its effect on the detector’s z-score. Lettershred’s multi-model dispersal addresses the generation architecture, while BIRA is the strongest single-rewriter technique covered here, followed by entropy-aware substitution and generic paraphrasing. The approaches differ in how directly they disrupt green-token overrepresentation, but tests run with different configurations cannot show which works best in every case.

How KGW watermarks work at the token level

KGW watermarking partitions the vocabulary into green and red lists before each token is sampled. The previous token is hashed with a secret key to determine the partition, with proportion γ assigned to the green list. A logit bias δ then increases the probability of selecting green tokens. The dependency is local, but the evidence accumulates. Each selected token also becomes the context used to determine the next partition.

Detection counts the excess of green tokens over the expected baseline. The detector computes z = (|S|_G − γ|T|) / √(|T|γ(1−γ)) and flags text when that score exceeds its threshold. Detection is model-free and requires no rerun of the generator. LLM Watermark Evasion via Bias Inversion explains the predecessor dependency that connects this count to the token sequence.

KGW-Soft favors green tokens, while KGW-Hard blocks all red tokens, which makes the signal stronger but harms text quality more. Invisible Entropy: Towards Safe and Efficient Low-Entropy LLM Watermarking explains why predictable positions strain this mechanism. When a fixed phrase strongly determines the next token, forcing a different choice can disrupt fluency and yield unreliable evidence of watermarking.

Why evasion effectiveness must be measured in z-score impact, not word-change rate

The operative benchmark is the detection threshold. A rewrite that changes many words but leaves z above that threshold has failed against that detector. A smaller change that moves z below it has succeeded for that tested text and configuration.

Lettershred reports a KGW z-score reduction from 15.46 to -0.32, below a threshold of 4, with 177 tokens, a substantial share of the text, rewritten through cross-provider dispersal. That is a self-reported result for the tested configuration and detector.

The same distinction governs the research comparisons that follow. The BIRA paper targets the conditional probability of green-token sampling, while entropy-aware approaches focus on positions where the watermark bias can change which token is chosen. Both address the statistic more directly than edit volume does. Their results still require context. A z-score, an attack-success rate and a theoretical bound describe different evidence and should not be treated as interchangeable benchmarks.

Paraphrasing: the entry-level technique and why it underperforms

Generic paraphrasing is easy to use because it needs neither the watermark key nor knowledge of green-list assignments. A language model rewrites the passage, disrupting some predecessor-token relationships as it changes wording and sentence structure. Some watermark signal can disappear in the process. However, the rewrite does not specifically target the green-token excess that determines the z-score.

Single-pass paraphrasing can meaningfully reduce true-positive rates in benchmarks, so describing it as useless would misrepresent its effect. The limitation is residual signal. Semantically faithful changes can leave enough green-token overrepresentation for detection, even when the passage looks substantially different.

Document length makes that limitation more pronounced. For a sustained excess of green tokens, the z-score grows with the square root of text length, giving the detector more statistical power. A rewrite that escapes detection on a short passage may therefore remain detectable in a longer article. The weakness is not a lack of visible change, but inconsistent disruption of the positions that contribute to the signal.

Entropy-aware token substitution: targeting the high-value positions the watermark relies on

Entropy-aware substitution improves on generic paraphrasing by distinguishing between token positions. At a high-entropy position, several continuations are plausible, so the green-list bias can tip the selection toward one of them. At a low-entropy position, a dominant continuation leaves less freedom.

The Invisible Entropy paper describes how to watermark only high-entropy positions and offers a lightweight way to find them without access to the original language model. For evasion, the corresponding logic is to concentrate substitutions where that selection pressure operates. Replacing a high-entropy green token with a semantically suitable red token directly reduces the relevant excess. Entropy alone, however, is not knowledge of the secret partition.

A Linguistics-Aware LLM Watermarking via Syntactic Predictability introduces STELA and distinguishes semantic fixedness from grammatical necessity as sources of low entropy. That distinction supports finer targeting than uniform rewriting. Substitution nevertheless remains tied to an existing sequence. It changes selected choices rather than generating the passage from a fresh distribution, which limits how far it can move beyond the original output’s statistical structure.

Bias Inversion (BIRA): the highest-performing single-model evasion technique

The Bias-Inversion Rewriting Attack, or BIRA, formalizes the move from changing words to suppressing the watermark’s statistical advantage. The BIRA paper derives a sufficient condition under which reducing the average conditional probability of green-token sampling makes detection probability decay exponentially. Its result depends on a positive suppression margin, expressed as δ, with exponential dependence on δ². Consistent suppression is the central lever.

BIRA implements that principle through a proxy suppression set identified using token self-information, or surprisal. During rewriting, it applies a negative logit bias to those tokens. This reverses the direction of the promotion on which green-list watermarking depends, without requiring the original model’s internal parameters. The attack uses the generated text and a proxy language model.

The paper reports evasion rates exceeding 99% across diverse watermarking schemes, with particularly high KGW attack success and better semantic fidelity than prior baselines. These are the paper’s experimental results, whose generalizability across detector configurations remains uncertain. The theoretical result explains the mechanism. It does not make every practical proxy equally effective.

BIRA is consequently the strongest single-rewriter approach in this comparison, but it remains a rewriting method within one proxy model’s distribution. That proxy may favor tokens much like those the original generator favors, especially when their training data are similar. This possibility motivates the architectural tier without proving that every multi-model system will outperform BIRA.

Learnability attacks: the long-horizon threat that exposes watermark key inference

Learnability attacks pursue a different objective from rewriting. Instead of removing a signal from an existing passage, they attempt to infer green-list assignments from accumulated watermarked output. On the Learnability of Watermarks for Language Models investigates whether neural networks can learn those assignments. The underlying exposure is deterministic repetition.

If the same predecessor and secret key determine the same partition, enough observations may reveal which tokens are likely to be green without revealing the key itself. Successful inference would allow future generation to avoid green tokens directly. That would compromise the practical secrecy of the assignments. The threat is therefore distinct from an immediate editing tool.

Kirchenbauer et al. find that hashing fewer previous tokens (lower values of k) produces more repetitive output, while higher k resists learnability and improves naturalness. Learnability attacks still require substantial observed watermarked text for training. Their practical relevance is greatest over a longer horizon, particularly for high-volume production, rather than as a directly comparable entry in a rewrite-success benchmark.

Why multi-model architectures disrupt watermarks structurally, not symptomatically

Rewriting attacks fight a signal that already exists. Multi-model dispersal instead changes how the text is assembled, so that one model’s watermarked distribution need not govern the whole passage. Each hand-off brings in choices from another token distribution.

For this architecture to work as argued, the generators must differ in real ways. A watermark assignment associated with one model’s output does not automatically govern another model’s generation, so independent contributions can dilute the original signal. Single-instance watermarking assumes that the text comes from a more consistent source than several models working together to build a passage. Provenance becomes harder to represent as one uninterrupted statistical imprint.

That distinction also separates dispersal from BIRA’s single proxy. Different providers can bring different training data, tokenizers and sampling logic, rather than another pass through one rewriting distribution. Provider diversity must nevertheless be real. Two endpoints from the same underlying provider may still make correlated choices or share green-list logic.

This places architectural dispersal above surface editing in the depth of its intervention. The mechanism offers a reason to expect disruption of a single-source signal. A detector result is still needed to establish whether a particular implementation achieved it.

How Lettershred implements multi-model dispersal and what the z-score benchmarks show

Lettershred implements that architectural approach through parallel drafting, line-by-line selection and a final enhance pass. A brief enters the system, and a dozen model kernels draft in parallel across different models. A judge panel selects the best lines. The enhance pass then unifies the voice, so the assembled article does not depend on one model drafting the entire text.

Unlike a paraphraser, Lettershred disperses token-level authorship during composition rather than relying only on a rewrite of a completed watermarked draft. That is the relevant capability for this comparison. The architecture is intended to prevent one model’s green-list signature from dominating the whole article.

The before-and-after scores reported earlier move from clearly above the tested KGW threshold to below it, while the token count identifies how much of the processed sequence changed. The negative final score shows that, in this test, the positive green-token excess the detector was looking for was gone. It does not establish performance against other detector configurations.

The evidence therefore supports a narrower conclusion than universal watermark removal. Lettershred reports successful disruption in its own KGW test, with an architectural explanation for the result. The supplied benchmark supports comparison with BIRA, entropy-aware substitution or paraphrasing by mechanism alone.

The operational distinction is that the surrounding content process can remain intact. Lettershred takes a brief and outline and turns them into one on-brand article, so teams can change how text is generated while keeping their planning process. Architectural intervention need not mean workflow overhaul. Here, the proposed advantage is the combination of dispersed generation and a familiar article-production interface.

What the ranking reveals about where compliance must be built

The ranking separates three kinds of intervention. Paraphrasing and substitution alter existing token choices, BIRA changes the probabilities used by a single rewriter, and multi-model dispersal changes the generation structure itself. Learnability sits outside that progression because it seeks reusable knowledge of secret assignments.

The strategic implication is to locate the intervention where the unwanted statistical dependency originates. For teams pursuing watermark-related compliance, that means addressing generation rather than treating every completed draft as a remediation task. The number of models or edits alone cannot show how well a pipeline works.

Any watermark-removal claim should name the detector, show z-scores before and after, and account for token-level changes. That standard applies equally to rewriting methods and architectural systems. It distinguishes an explained mechanism from a demonstrated outcome, without assuming that one test settles every configuration.

Content teams can keep clear briefs, approved outlines and the review process around them while changing the generation step. The durable objective is compliance by construction, replacing a bolt-on editing stage whose success depends on how much residual signal survives.

AI Watermark Detection

Sources