--- pretty_name: LLM Watermark Evaluation — Inputs and Generations license: [cc-by-4.0, cc-by-sa-4.0] language: [en, ar, zh, hi, ja] task_categories: [text-generation] tags: [watermarking, llm-watermarking, ai-generated-text, text-detection, robustness] size_categories: [100K__` split each in the ELI5 configs: the same questions, sampling settings and seeds; every prompt re-rendered from `question` with the generator's own chat template (Gemma has no system role, so the multilingual system prompt opens the user turn); generators without a thinking mode run `multilingual` and `benchmarks` without a reasoning budget (the whole generation is the answer, `reasoning` is empty). | scheme (split) | generator | scheme and settings | main | dilution_answers | multilingual | benchmarks | attacks | documents | |---|---|---|---|---|---|---|---|---| | `thicket__llama31` | `meta-llama/Llama-3.1-8B-Instruct` | Thicket (ours): keyed token-pair circle tilt with a likelihood price; κ = 6, price γ = 2 in answers (4 / 1 inside the reasoning delimiter), top-128 candidate shortlist, 256-label circle, first token and every 512-token restart neutral | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `textseal__llama31` | `meta-llama/Llama-3.1-8B-Instruct` | TextSeal (official code), mixing α = 0.1 | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `none__llama31` | `meta-llama/Llama-3.1-8B-Instruct` | No watermark (plain sampling; the unwatermarked reference) | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `kgw_delta0.7__llama31` | `meta-llama/Llama-3.1-8B-Instruct` | KGW, γ = 0.25, δ = 0.7 (perplexity cost matched to Thicket's) | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `kgw_delta0.9__llama31` | `meta-llama/Llama-3.1-8B-Instruct` | KGW, γ = 0.25, δ = 0.9 | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `thicket__gemma2` | `google/gemma-2-9b-it` | Thicket (ours): keyed token-pair circle tilt with a likelihood price; κ = 6, price γ = 2 in answers (4 / 1 inside the reasoning delimiter), top-128 candidate shortlist, 256-label circle, first token and every 512-token restart neutral | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `textseal__gemma2` | `google/gemma-2-9b-it` | TextSeal (official code), mixing α = 0.1 | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `none__gemma2` | `google/gemma-2-9b-it` | No watermark (plain sampling; the unwatermarked reference) | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `kgw_delta0.7__gemma2` | `google/gemma-2-9b-it` | KGW, γ = 0.25, δ = 0.7 (perplexity cost matched to Thicket's) | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `thicket__mistral7b` | `mistralai/Mistral-7B-Instruct-v0.3` | Thicket (ours): keyed token-pair circle tilt with a likelihood price; κ = 6, price γ = 2 in answers (4 / 1 inside the reasoning delimiter), top-128 candidate shortlist, 256-label circle, first token and every 512-token restart neutral | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `textseal__mistral7b` | `mistralai/Mistral-7B-Instruct-v0.3` | TextSeal (official code), mixing α = 0.1 | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `none__mistral7b` | `mistralai/Mistral-7B-Instruct-v0.3` | No watermark (plain sampling; the unwatermarked reference) | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | | `kgw_delta1.0__mistral7b` | `mistralai/Mistral-7B-Instruct-v0.3` | KGW, γ = 0.25, δ = 1.0 | 5,000 | 200 | 500 | 300 | 1,600 | – | | `kgw_delta0.7__mistral7b` | `mistralai/Mistral-7B-Instruct-v0.3` | KGW, γ = 0.25, δ = 0.7 (perplexity cost matched to Thicket's) | 5,000 | 200 | 500 | 300 | 1,600 | 1,400 | Each scheme used one fixed key; the keys are released in [`keys.json`](keys.json) and explained in [`KEYS.md`](KEYS.md), so the detection results can be reproduced with the project's code. ## Splits | split | inputs | outputs | sampling | |---|---|---|---| | `main` | 1,000 English questions (ELI5, test split) | 5 independent answers per question (`sample` 0–4), at most 400 tokens | temperature 0.8, top-p 0.9; thinking disabled | | `dilution_answers` | the first 200 of those questions | 1 answer each; the evaluation embeds these answers into 2k–12k-token Wikipedia documents to test detection inside longer text (see `documents`) | temperature 1.0, top-p 0.95 | | `multilingual` | 100 questions in each of English, Arabic, Chinese, Hindi, Japanese | 1 answer each, with its reasoning trace (`reasoning`) and final answer (`answer`); the watermark covers both | temperature 0.7, top-p 0.95; at most 2,000 reasoning tokens, 4,096 tokens in total | | `benchmarks` | 300 questions from GSM8K, MATH-500, AIME, MMLU, ARC-Easy, ARC-Challenge, HellaSwag, WinoGrande, HumanEval and MBPP (30 each), with their reference answers | 1 answer each (reasoning trace and final answer), used to measure the effect of each watermark on task accuracy | temperature 0.6, top-p 0.95; at most 3,000 reasoning tokens, 4,096 tokens in total | | `attacks` | the first 200 `main` answers of `sample` 0 | each answer attacked 8 ways (`attack`, `setting`): rephrased by Llama-3.2-3B-Instruct with the TextSeal repository's four prompts (mild, moderate, aggressive, extreme), 30% word deletion, 50% WordNet synonym substitution, DIPPER-XXL paraphrase, English→French→English back-translation (NLLB-200-distilled-600M) | rephrasing: temperature 0.5 / 0.8 / 1.0 / 1.2 (mild … extreme), top-p 0.95, at most 1,024 new tokens; DIPPER: lexical diversity 60, order diversity 0, top-p 0.75; edits and sampling with fixed seeds | | `documents` | the 200 `dilution_answers` texts and unwatermarked English Wikipedia host documents | 1,400 documents: each answer embedded whole in a 2k, 4k, 8k or 12k-token document (`dilute_*`) or cut into k = 2, 3 or 5 pieces scattered over an 8k-token document (`fragment_8k_k*`); `true_spans` marks the watermarked text | insertion points drawn with a fixed seed; lengths in generator tokens | ## Fields | field | meaning | |---|---| | `method` | watermark scheme (the config name, without the `__` suffix of the other-generator configs) | | `model` | the generator that produced the text (in `attacks` and `documents`: the generator of the source answer) | | `sample` | sample index 0–4 (the five samples use different sampling seeds); in `attacks`, the sample of the attacked answer | | `prompt_index` | position of the question in the evaluation's list (in `attacks`: of the attacked answer, which `main` holds under the same `prompt_id` and `sample`; in `documents`: of the embedded answer, which `dilution_answers` holds under the same `prompt_id`) | | `prompt_id` | identifier of the question in its source dataset | | `question` | the question as written in the source dataset | | `prompt` | the exact string given to the model: the question inside the generator's chat template | | `language` | multilingual splits only | | `benchmark`, `kind`, `reference_answer` | `benchmarks` only: source benchmark, answer type and gold answer | | `text` | the generated text; in `attacks` the attacked text; in `documents` the whole document | | `reasoning`, `answer` | multilingual splits and `benchmarks`: `text` split at its first `` (the answer stripped of surrounding whitespace); empty `reasoning` when the generator has no thinking mode | | `attack`, `setting` | attacks only: attack family (rephrasing, word_deletion, synonym_substitution, dipper, back_translation) and its strength or parameters | | `condition`, `length`, `k` | documents only: dilution / fragmentation condition, target document length in generator tokens, number of watermarked pieces | | `host_id` | documents only: the Wikipedia host document (consecutive English Wikipedia articles concatenated to the target length) | | `true_spans` | documents only: character offsets (`start`, `end`, Python string indices) of the watermarked pieces in `text` | | `n_tokens` | length of `text` in tokens of the generator's tokenizer | ## Generation Model `Qwen/Qwen3.5-9B`, fp16, PyTorch SDPA attention; batched sampling with fixed per-batch seeds; each watermark applied inside the sampler by its official or reference implementation (TextSeal, Gumbel-max and SynthID-Text from the TextSeal codebase; KGW, Unigram, SWEET and EWD through the Hugging Face `transformers` watermark logits processor or re-implementations of their papers; Thicket from its authors' code). The unwatermarked reference (`none`, when present) uses the same sampler without a watermark. The `__` configs use `meta-llama/Llama-3.1-8B-Instruct`, `google/gemma-2-9b-it`, `mistralai/Mistral-7B-Instruct-v0.3` the same way (fp16), with the prompt handling described above. ## Not included - GPQA: its 30 questions and the corresponding answers are left out because GPQA's authors ask that its examples not be published online (benchmark-leakage prevention). All other benchmarks are included. - The negatives of the false-positive tests (Wikipedia passages and unwatermarked same-model answers to other ELI5 questions). - Copy-paste attack documents (watermarked fragments pasted into Wikipedia text). - `attacks` and `documents` of the runs that finished before per-text records were kept (their attacked texts and documents were not stored; see the tables above). - Evaluation scores: no p-values, detector outputs or metrics. ## Sources and licenses Generated texts: CC BY 4.0. They are outputs of the generator named in `model`, and their use is also subject to that model's terms: - `Qwen/Qwen3.5-9B`: subject to the Qwen3.5 model license (see the model card). - `meta-llama/Llama-3.1-8B-Instruct`: **Built with Llama.** Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved; outputs are also subject to the Llama 3.1 Acceptable Use Policy. - `google/gemma-2-9b-it`: Gemma Terms of Use (ai.google.dev/gemma/terms) and the Gemma Prohibited Use Policy. - `mistralai/Mistral-7B-Instruct-v0.3`: Apache 2.0. Attacked texts (`attacks`, `c4_attacks`): the rephrasings are outputs of `meta-llama/Llama-3.2-3B-Instruct` (**Built with Llama**; Llama 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved), the paraphrases of `kalpeshk2011/dipper-paraphraser-xxl`, the back-translations of `facebook/nllb-200-distilled-600M` (model license CC BY-NC 4.0). **`documents` / `c4_documents`: CC BY-SA 4.0.** These splits contain text from English Wikipedia (`wikimedia/wikipedia`, dump 20231101.en; © Wikipedia contributors, https://en.wikipedia.org), redistributed under CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/); the generated passages they embed are marked by `true_spans`. All other splits: CC BY 4.0. Inputs keep the licenses of their sources: ELI5 (Reddit questions, as distributed in the `Pavithree/eli5` Hugging Face dataset; also the English multilingual questions), the multilingual question pool (`shanearora/CaLMQA`, CC BY 4.0), GSM8K (MIT), MATH-500 (MIT), AIME (public competition problems), MMLU (MIT), ARC (CC BY-SA 4.0), HellaSwag (MIT), WinoGrande (CC BY), HumanEval (MIT), MBPP (CC BY 4.0). ## Citation Paper under review; citation to be added.