ProseLens

ProseLens detects who wrote the words of a document. It is IdeaLens's counterpart in the paper: the same backbone, training documents and labels, but it reads the raw document text instead of an outline, so it learns word-level provenance. It returns P(human), the probability that the document was written by a person.

ProseLens is nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 fine-tuned with LoRA (rank 64) on 1M English web documents (WildOutlines).

Try it in your browser: the IdeaLens & ProseLens demo scores your own text with both detectors, no installation or keys needed.

Results

From the paper: ProseLens is accurate when a document's ideas and words come from the same source (99.1%), but when they come from different sources it tracks the words and reaches 25.4% on idea provenance. IdeaLens, trained identically on outlines, reaches 95.3% and 81.3%; Pangram 4 reaches 98.5% and 25.9%.

Usage

Pass the document text as is.

Quick start with the idealens package

idealens (PyPI) scores documents with ProseLens on vLLM and applies the thresholds in this repo. No outline and no LLM call is needed:

pip install "idealens[vllm]"
idealens score docs.jsonl -o scores.jsonl --model ProseLens

Input is JSONL with a text field per document. In Python:

import idealens as il

with il.Detector("ProseLens") as det:         # vLLM, with this repo's thresholds
    records = det.score_documents([open("document.txt").read()])
print(records[0]["p_human"], records[0]["verdict"]["ai"])

To compare with idea-level detection on the same documents, idealens run docs.jsonl -o ideas.jsonl --model IdeaLens extracts their outlines and scores them with IdeaLens. The rest of this section runs the model directly.

Without a GPU: Tinker

ProseLens is also published on Tinker, which runs it on its own servers, so any laptop works. Sign up at auth.thinkingmachines.ai/sign-up, create an API key in the Tinker console, and run:

pip install "idealens[tinker]"
export TINKER_API_KEY=...
idealens score docs.jsonl -o scores.jsonl --model ProseLens --backend tinker

Tinker bills your account per token: under $1 per 1,000 typical web documents at list price (--dry-run gives an estimate). The checkpoint is tinker://f21bf499-021e-5735-85f5-f5e1183aad78:train:0/sampler_weights/nemo_docs_v1m_docs_r64_ep1_final; on our checks its scores agree with the weights in this repo to about 0.002 in P(human).

Load the merged model (66 GB download)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("rishanthrajendhran/ProseLens")
model = AutoModelForCausalLM.from_pretrained("rishanthrajendhran/ProseLens", dtype=torch.bfloat16, device_map="auto").eval()

The weights take 59 GiB of GPU memory, and each input adds more; see Hardware requirements.

Or apply the adapter to the base model (3 GB download)

adapter/ holds the LoRA adapter as trained, in the layout of the Tinker training service. If you already have the base model, load_adapter.py merges the adapter into it in memory. The resulting weights are bit-identical to the merged model's:

import importlib.util
from huggingface_hub import hf_hub_download

path = hf_hub_download("rishanthrajendhran/ProseLens", "load_adapter.py")
spec = importlib.util.spec_from_file_location("load_adapter", path)
la = importlib.util.module_from_spec(spec); spec.loader.exec_module(la)
model, tok = la.load_model()   # base model + adapter/, then la.p_human(model, tok, text)

Do not load adapter/ with peft.PeftModel. In transformers, Nemotron fuses the Mamba gate and x projections into one in_proj and stores each layer's 128 routed experts as a single 3D tensor, so PEFT has nowhere to attach most of the adapter and skips it without a warning; the model then scores close to the base model. tinker-cookbook's weights.build_hf_model can also merge the adapter into full weights.

Score a document

ProseLens compares the next-token probabilities of human and ai:

SYSTEM = "Given a document, answer with one word: human if the document was human-written, ai if it was AI-generated."
SUFFIX = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
HUMAN, AI = 50755, 2464   # token ids of "human" and "ai"

@torch.no_grad()
def p_human(text):
    ids = tok.encode(f"<|im_start|>system\n{SYSTEM}<|im_end|>\n<|im_start|>user\n{text}{SUFFIX}",
                     add_special_tokens=False)
    logits = model(torch.tensor([ids], device=model.device)).logits[0, -1].float()
    return torch.softmax(logits[[HUMAN, AI]], -1)[0].item()

text = open("document.txt").read()
print(p_human(text))

Build the prompt string exactly as above rather than through the chat template.

Score with vLLM

For many inputs, vLLM is about 15 times faster than the code above and fits much longer inputs on one 80 GB GPU. With vLLM 0.21 (install xgrammar==0.2.1; later releases require transformers < 5), reusing SYSTEM, SUFFIX, HUMAN and AI from above:

import math, os
os.environ.setdefault("VLLM_USE_FLASHINFER_SAMPLER", "0")   # FlashInfer kernels compile CUDA code and need nvcc
os.environ.setdefault("VLLM_USE_FLASHINFER_MOE_FP16", "0")
os.environ.setdefault("VLLM_USE_DEEP_GEMM", "0")            # H100 warmup crashes when DeepGEMM is not installed
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

tok = AutoTokenizer.from_pretrained("rishanthrajendhran/ProseLens")
llm = LLM(model="rishanthrajendhran/ProseLens", dtype="bfloat16", max_num_seqs=256, enable_prefix_caching=False,
          max_logprobs=20, enable_flashinfer_autotune=False, seed=0)
sp = SamplingParams(max_tokens=1, temperature=0.0, logprobs=20)

def p_human_batch(texts):
    prompts = [{"prompt_token_ids": tok.encode(f"<|im_start|>system\n{SYSTEM}<|im_end|>\n<|im_start|>user\n{x}{SUFFIX}",
                                               add_special_tokens=False)} for x in texts]
    out = []
    for r in llm.generate(prompts, sp, use_tqdm=False):
        lp = r.outputs[0].logprobs[0]           # the top 20 next-token log-probabilities
        out.append(1 / (1 + math.exp(lp[AI].logprob - lp[HUMAN].logprob)))
    return out

# when done: without this, vLLM 0.21 keeps a script running after its last line
llm.llm_engine.engine_core.shutdown()

max_num_seqs=256 keeps every running sequence's Mamba state in memory; vLLM's H100 default (1,024) does not fit beside the weights. If human or ai is missing from the top 20 (rare), score the prompt followed by each label token with SamplingParams(max_tokens=1, prompt_logprobs=0) and read the last prompt log-probability of each. Scores agree with the training-time scores to about 0.001 in P(human) on average; A100 and H100 GPUs differ by as much.

Thresholds

ProseLens flags a document as AI-written when P(human) is below a cut. Each cut is set so that a given share of human documents is flagged (the false-positive rate, FPR), measured on the 80,000 human documents in WildOutlines's calibration split (10,000 per format). The paper's operating point is the global cut at 1% FPR.

Per-format cuts give each format its own operating point. They need the document's format, which the paper assigns with WebOrganizer's annotation prompt run on Gemini 3.7 Flash; the calibration documents use the formats recorded in WildOutlines. Each is the format's own quantile, shrunk toward the global cut with weight n / (n + 2500); at 0.1% FPR 10,000 documents per format are too few, so there is no per-format cut. A document outside these eight formats has no per-format cut; do not fall back to the global cut for it.

thresholds.json holds every cut at full precision.

These rates hold for English web documents like the training data. For another domain, fit the cut on human documents from that domain.

ProseLens scores 97% of human calibration documents above 0.99, so its cuts at higher FPRs sit close to 1 and small shifts in score move the realised FPR a long way.

Hardware requirements

Measured with transformers 5.15 in bf16 on NVIDIA H100 80GB GPUs (our other runs used A100 80GB), with transformers' PyTorch implementation of the Mamba layers (no fused Mamba kernels installed). We have not tried CPU-only inference.

GPU memory then grows with the length of the input, by about 4.2 MiB per token at typical lengths, scoring one input at a time:

Inputs of 6,000 tokens do not fit on one 80 GB GPU and 12,000 do not fit on two; lowering the Mamba chunk size from 128 to 64 did not change either limit.

ProseLens reads whole documents, so their length decides the hardware. One 80 GB GPU handles documents up to about 4,000 tokens (about 3,000 words); 85% of WildOutlines's test documents are that short. Two 80 GB GPUs handle up to 8,000 tokens (about 6,000 words), which covers all but about 1 in 1,000 test documents. We have not measured longer inputs or more GPUs.

Intended use and limitations

• ProseLens estimates who wrote a document's words. It should not be the sole basis for decisions about a person's work.
• It was trained on English web documents of at least 500 words in eight long-form formats (Nonfiction Writing, Knowledge Article, Personal Blog, News Article, Academic Writing, User Reviews, Personal About Page, Creative Writing).
• Its training labels come from the Pangram prose detector, applied to whole documents.

Related models

Training data: WildOutlines.

All IdeaLens models and datasets are in the IdeaLens collection.

License

OpenMDW-1.1, the license of the base model (see LICENSE).

Citation

@article{idealens2026,
  title         = {IdeaLens: Detecting AI Ideas in Long-form Writing},
  author        = {Rajendhran, Rishanth and Choi, Minjoon and Russell, Jenna and Namuduri, Ramya and B{\"o}l{\"o}ni-Turgut, Deniz and Karpinska, Marzena and Wieting, John and Iyyer, Mohit},
  journal       = {arXiv preprint arXiv:2610.06778},
  year          = {2026},
  eprint        = {2610.06778},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.06778}
}