RedOne: Dual Systems in Radiology & SNS

• RedOne is a polysemous label referring to both CXR-ReDonE, a radiology report generation system, and a SNS-specific large language model developed by Xiaohongshu.
• CXR-ReDonE leverages data-centric pipelines, using GPT-3 and BioBERT-based methods to remove prior-reference hallucinations from chest radiograph reports.
• In the SNS domain, RedOne and its successor RedOne 2.0 evolve from SFT-dominated models to RL-prioritized post-training pipelines, enhancing robustness and domain alignment.

Searching arXiv for the provided RedOne-related papers to ground the article in the latest records. RedOne is a polysemous label in recent machine-learning literature. In one line of work, it denotes CXR-ReDonE, a chest X-ray radiology report generation system whose central aim is to remove hallucinated references to non-existent prior examinations from generated impressions (Ramesh et al., 2022). In another, it denotes a family of domain-specific LLMs for Social Networking Services (SNS) developed by Xiaohongshu, first as RedOne and later as RedOne 2.0, with post-training recipes designed for heterogeneous SNS tasks, bilingual content, and robustness under distribution shift (Zhao et al., 13 Jul 2025, Zhao et al., 10 Nov 2025). The shared name does not imply a shared technical lineage; the cited works concern separate application domains, model classes, datasets, and evaluation protocols.

1. Terminological scope and disambiguation

In the cited literature, “RedOne” refers to two unrelated systems.

Usage of “RedOne”	Domain	Core characterization
CXR-ReDonE	Radiology report generation	“Contrastive X-Ray Report Determination Employing Prior Reference Removal”
RedOne	SNS-domain LLM	Xiaohongshu’s original SNS-oriented LLM
RedOne 2.0	SNS-domain LLM	RL-prioritized successor to the original RedOne

For radiology, the shorthand “RedOne” arises informally from CXR-ReDonE (“Re-Don-E → RedOne”), while the paper explicitly introduces CXR-ReDonE as the main new system (Ramesh et al., 2022). For SNS, RedOne is the formal model name, and RedOne 2.0 is a redesigned successor that replaces a classical SFT-dominated recipe with a progressive, RL-prioritized post-training paradigm (Zhao et al., 13 Jul 2025, Zhao et al., 10 Nov 2025).

This dual usage is important because the two research programs address different failure modes. CXR-ReDonE targets hallucinated temporal comparison language in medical reporting. RedOne and RedOne 2.0 target domain adaptation for social-network workloads, including moderation, retrieval, dialogue, user-behavior modeling, and SNS-style translation.

2. CXR-ReDonE: RedOne in chest radiography

CXR-ReDonE is a radiology report generation system for chest X-rays that is designed specifically to avoid hallucinated references to non-existent prior exams (Ramesh et al., 2022). Its task is standard report generation from a chest X-ray image or images, producing the free-text radiology report’s impression section. The system is built by retraining CXR-RePaiR, a retrieval-based report generator, on a cleaned version of MIMIC-CXR in which references to priors are removed or rewritten; this adapted dataset is called MIMIC-PRO.

The motivating problem is that radiology reports in routine practice often refer to priors, meaning previous imaging studies and their associated reports for the same patient. In the source data, phrases such as “Compared to the prior radiograph,” “No interval change from prior study,” “Heart size is stable,” and “Again seen is” encode comparative information over time rather than pure description of the current study. In MIMIC-CXR, 173,822 of 226,759 reports (76.3%) contain at least one of 18 “prior-like” keywords, including change, unchanged, prior, stable, interval, previous, and again (Ramesh et al., 2022).

The paper’s core claim is that models trained on such text internalize the statistical regularity that good radiology reports often contain prior-related language. At inference time, however, deployed systems frequently do not have access to previous exams or reports. The result is a clinically important hallucination mode: models generate statements such as “Compared to the prior study…” or “Heart size is unchanged from the prior radiograph” even when only a single current X-ray is available. The paper shows that this failure mode occurs in systems such as CXR-RePaiR, R2Gen, and M² Trans, and argues that such hallucinations misrepresent the evidential basis of the report and can undermine clinical trust (Ramesh et al., 2022).

The dataset context is equally central. MIMIC-CXR contains 377,110 images from 227,835 studies, and the paper uses a curated subset of 226,759 reports, referred to in the paper simply as “MIMIC-CXR.” Reports are semi-structured, typically including Findings and Impression, and the benchmark role of MIMIC-CXR in chest X-ray report generation makes prior-reference contamination a systemic issue rather than a narrow artifact of one model.

3. Prior-reference removal and construction of MIMIC-PRO

The central contribution behind CXR-ReDonE is a data-centric pipeline for removing prior-related language from the training corpus (Ramesh et al., 2022). The paper proposes two methods: a GPT-3-based few-shot rewriting pipeline named FilBERT+GPT-3, and a BioBERT-based token classification pipeline named GILBERT. Both are developed using a manually curated shared corpus of 103 report pairs, where each pair contains an original report and a ground-truth edited version with prior references removed or rewritten.

The first method, FilBERT+GPT-3, is sentence-centric. FilBERT is a BioBERT base model with a sequence classification head that predicts whether a sentence contains any reference to a prior. It is initialized from dmis-lab/biobert-base-cased-v1.2, trained for 10 epochs with Adam at learning rate $2 \times 10^{-5}$ and $\epsilon = 10^{-8}$, with batch size 16 on one Tesla P100 GPU. On the held-out test set it achieves accuracy $= 0.907$, with false positive rate $= 5.55\%$ and false negative rate $= 3.70\%$ at threshold $0.5$ (Ramesh et al., 2022). This filter is used because running GPT-3 on every report sentence in MIMIC-CXR is estimated to cost $92,000**, whereas filtering reduces the estimate to **$16,560.

Flagged sentences are then rewritten by GPT-3 (DaVinci) using few-shot prompting. The paper reports that DaVinci outperforms Curie, Babbage, and Ada, with $F_1 = 0.5569$ versus 0.37, 0.33, and 0.23. A temperature sweep finds the best result at temperature $= 0.3$, and experiments over subreports of $n \in \{1,2,3,4\}$ sentences show that $\epsilon = 10^{-8}$0 performs best, indicating that GPT-3 is more reliable as a sentence-level rewriter than as a whole-report editor (Ramesh et al., 2022).

The second method, GILBERT, formulates prior removal as a token-classification problem. Each token is labeled as either KEEP or REMOVE, rather than with BIO tags. Because BioBERT uses WordPiece tokenization, the tokenizer is modified so that every subword receives a label. The model uses BioBERT with a token-classification head and standard token-level cross-entropy,

$\epsilon = 10^{-8}$1

Training is implemented in PyTorch for 10 epochs, each with 100 steps, with training batch size 4, test batch size 2, gradient clip norm 10, Adam learning rate $\epsilon = 10^{-8}$2, on a single Tesla T4 GPU (Ramesh et al., 2022).

On the shared test corpus, GILBERT achieves $\epsilon = 10^{-8}$3, substantially higher than FilBERT+GPT-3, which achieves $\epsilon = 10^{-8}$4. When GILBERT is applied to all 226,759 reports in MIMIC-CXR, the total number of prior keyword instances drops from 259,376 to 82,074, a >68.3% reduction. Because of this higher $\epsilon = 10^{-8}$5, lower cost, and better scalability, GILBERT is selected as the primary method for constructing MIMIC-PRO (Ramesh et al., 2022).

The paper also reports a manually adapted evaluation set: from the standardized MIMIC-CXR test set of 2,192 images, a board-certified radiologist and two senior medical students edit reports to remove prior phrases or rewrite them as current-state descriptions. This expert-edited set serves as the ground-truth test set for evaluating report-generation systems that are not supposed to reference priors.

4. Retrieval architecture, evaluation, and limitations of CXR-ReDonE

CXR-ReDonE inherits the retrieval formulation of CXR-RePaiR but changes the training data and the vision-language backbone (Ramesh et al., 2022). CXR-RePaiR treats report generation as retrieval from a corpus $\epsilon = 10^{-8}$6, encoding reports with a text encoder $\epsilon = 10^{-8}$7 and images with an image encoder $\epsilon = 10^{-8}$8, then scoring image-report compatibility by dot product:

$\epsilon = 10^{-8}$9

Prediction can operate at the report level, returning the report $= 0.907$0 that maximizes $= 0.907$1, or at the sentence level, composing a synthetic report from the top $= 0.907$2 retrieved sentences.

CXR-ReDonE retrains this framework on MIMIC-PRO and replaces the original CLIP backbone with ALBEF (“Align Before Fuse”), while preserving the core retrieval logic. The paper emphasizes that the main innovation is not a new generator architecture but a combination of prior-cleaned training text and a stronger image-text aligner. Training uses 60 epochs, 100 steps per epoch, on 4 Quadro RTX-8000 GPUs (Ramesh et al., 2022).

Evaluation uses BERTScore, $= 0.907$3 based on CheXbert embedding similarity, and RadGraph $= 0.907$4, rather than BLEU, METEOR, or CIDEr, which the paper argues are not well suited to radiology. On the expert-edited test set, CXR-ReDonE improves over the baseline trained on original MIMIC-CXR in every evaluated configuration. At report-level retrieval, the baseline scores BERTScore $= 0.907$5, $= 0.907$6, and RadGraph $= 0.907$7, whereas CXR-ReDonE scores 0.2160, 0.3601, and 0.0925. At sentence-level retrieval with $= 0.907$8, where CXR-ReDonE achieves its best BERTScore, the baseline scores 0.2292 / 0.3822 / 0.1045, whereas CXR-ReDonE reaches BERTScore $= 0.907$9, $= 5.55\%$0, and RadGraph $= 5.55\%$1, corresponding to a 2.57% absolute improvement in BERTScore (Ramesh et al., 2022). At $= 5.55\%$2, CXR-ReDonE achieves the best reported RadGraph $= 5.55\%$3, a +2.68% improvement over the baseline.

Qualitative examples show the same pattern. Systems trained on the original corpus often generate “stable,” “again,” “unchanged,” or “decreased since previous,” whereas CXR-ReDonE tends to produce descriptions of current device positioning and lung findings without unsubstantiated comparative language. The paper interprets this as evidence that removing prior references from training data improves clinical content fidelity rather than degrading it (Ramesh et al., 2022).

The limitations are explicit. Residual hallucinations remain because prior keyword counts are reduced by about 68%, not eliminated. Grammar and fluency can degrade because GILBERT deletes tokens without grammatical repair; the paper mentions a future grammatical error correction layer such as GECToR. Temporal information is also genuinely useful in practice, so removing all prior references is appropriate only for the setting where priors are unavailable. Additional constraints include dependence on BioBERT and GPT-3 behavior, restriction to chest radiography and MIMIC-CXR, and the absence of a prospective user study or radiologist satisfaction metric.

5. RedOne as a domain-specific LLM for Social Networking Services

In the SNS literature, RedOne is a domain-specific LLM developed by Xiaohongshu for Social Networking Services, rather than a radiology system (Zhao et al., 13 Jul 2025). The motivation is that many existing SNS applications rely on separate single-task finetuned models for sentiment analysis, rumor detection, content moderation, search relevance, or other scenarios. The paper argues that such specialization causes performance plateaus, fragmented deployment across scenarios, and a mismatch with SNS language, which is informal, slang-heavy, emoji-rich, and behaviorally grounded.

The original RedOne is a post-trained variant of Qwen2.5. It introduces no new architecture, tokenizer, or special SNS-specific modules; domain specificity comes from data composition and a three-stage training strategy: Continue Pretraining (CPT), Supervised Fine-Tuning (SFT), and Preference Optimization (PO). The reported model sizes are RedOne-7B, built from Qwen2.5-7B-Instruct, and RedOne-32B, built from Qwen2.5-32B (Zhao et al., 13 Jul 2025).

During CPT, the authors collect more than 100B tokens from general high-quality corpora and SNS-specific domain data, then apply task-oriented rule filtering, SLM filtering, and RegMix, reducing the corpus to about 20B high-quality tokens. Continue pretraining then uses standard causal language modeling with sequence length 4,096, learning rate $= 5.55\%$4, for 1 epoch (Zhao et al., 13 Jul 2025).

The SFT stage constructs a large multi-task instruction suite from SNS notes, comments, queries, and interaction logs, while preserving SNS style such as informal wording, sarcasm, sentiment, and topic drift. Tasks are organized into six capability areas: Content Understanding, Information Extraction, Semantic Matching, User Behavior Modeling, Dialogue and Persona Simulation, and Translation. The paper explicitly names tasks such as Note Taxonomy, Query Classification, Query Intent Recognition, Hashtag Prediction, Machine Reading Comprehension, Highlight Word Detection, Query-Note Relevance, Query-Note Retrieval, Post-View Search, Emotional Companion Dialogue, Role-playing Dialogue, and SNS Domain Translation. General instruction data such as LIMA, Infinity Instruct, WildChat, OpenThoughts, and LIMO are mixed in to maintain general capabilities. SFT is performed in two steps: 3 epochs on balanced SNS and general data, then 2 more epochs with an SNS-heavier ratio, using sequence length up to 16,384, batch size 128, AdamW with $= 5.55\%$5, $= 5.55\%$6, $= 5.55\%$7, learning rate $= 5.55\%$8, and linear warmup ratio 0.1 (Zhao et al., 13 Jul 2025).

The PO stage uses Direct Preference Optimization (DPO) on both subjective and objective SNS tasks. Preference pairs come from human expert judgments, judge-model annotations, explicit correct-versus-incorrect labels, and model-error constructions. Implementation details are learning rate $= 5.55\%$9, batch size $= 3.70\%$0, sequence length $= 3.70\%$1, epochs $= 3.70\%$2, and an auxiliary SFT loss coefficient $= 3.70\%$3 (Zhao et al., 13 Jul 2025).

Evaluation covers SNS-Bench, containing 8 tasks, and SNS-TransBench, a bilingual benchmark for ZH→EN and EN→ZH SNS translation. For RedOne-7B, the paper reports General-Bench $= 3.70\%$4, SNS-Bench $= 3.70\%$5, and SNS-TransBench $= 3.70\%$6, compared with Qwen2.5-7B at 63.01 / 52.86 / 40.55. The paper summarizes this as up to +14.02% average improvement on SNS-Bench and +7.56% on SNS-TransBench, with a slight +0.82% gain on general benchmarks (Zhao et al., 13 Jul 2025). For RedOne-32B, the reported numbers are 73.72 / 69.03 / 49.84, versus 71.68 / 58.89 / 43.44 for Qwen2.5-32B, summarized as +10.14% on SNS-Bench, +6.40% on SNS-TransBench, and +2.04% on general benchmarks.

The paper also reports online A/B tests in two internal pipelines. In Harmful Content Detection, RedOne reduces Exposure Rate by 11.23% relative to an existing single-task finetuned baseline. In Post-View Search, it improves Click Page Rate by 14.95% (Zhao et al., 13 Jul 2025). Ablation results show that the full CPT + SFT + PO pipeline outperforms partial configurations, and comparisons against task-specific finetuning show that a domain-adapted RedOne provides a stronger base for further task specialization than the generic Qwen2.5 model.

6. RedOne 2.0: RL-prioritized post-training for SNS

RedOne 2.0 is a redesigned successor to the original SNS RedOne, built by post-training the Qwen3 family at 4B, 8B, 30B-A3B, and 32B scales with a progressive, RL-prioritized three-stage pipeline (Zhao et al., 10 Nov 2025). The redesign responds to a specific problem identified in the original RedOne recipe: an SFT-dominated pipeline can produce a “seesaw” between in-distribution SNS gains and out-of-distribution robustness, especially in smaller models.

The new pipeline consists of Exploratory Learning, Targeted Fine-Tuning, and Refinement Learning. The training corpus is normalized to $= 3.70\%$7 format and combines SNS domain data covering more than 75 tasks with curated general-domain data containing rationales. Stage 1, Exploratory Learning, uses DAPO-based RL on approximately 750K SNS entries and 50K general entries. A defining feature is the use of task-type-specific reward functions rather than a single scalar signal. The paper specifies four types: Exact Match for closed-ended classification, metrics-based rewards such as BLEU or chrF++ for open-ended tasks like translation, sandbox rewards for executable tasks, and pattern rewards for format or instruction-following tasks (Zhao et al., 10 Nov 2025).

The RL objective is a clipped-ratio DAPO loss over grouped candidates sampled from the old policy. The paper defines normalized token-level advantages from the sample rewards and optimizes a trust-region-style objective with clipping parameters $= 3.70\%$8 and $= 3.70\%$9. This stage is intended to establish initial SNS alignment while also diagnosing hard sub-tasks that remain weak after exploratory training (Zhao et al., 10 Nov 2025).

Stage 2, Targeted Fine-Tuning, constructs a dataset of 1.8M examples, comprising 1.7M SNS examples focused on Stage-1 failure buckets and 100K general examples with soft labels. For these general examples, the Stage-1 model generates 8 candidate outputs per prompt, and a separate judge model selects the best candidate as the soft target. The stated purpose is to reduce distributional mismatch between SFT targets and the model’s own prior while regularizing against catastrophic forgetting. SFT uses standard next-token cross-entropy for 2 epochs, batch size 64, maximum sequence length 16,384, optimizer AdamW, learning rate $0.5$0, warmup ratio 0.1, and cosine decay (Zhao et al., 10 Nov 2025).

Stage 3, Refinement Learning, applies DAPO RL again on about 400K examples drawn from SNS and general sources, emphasizing difficult subsets and increasing the share of rationale-rich samples to 57.18%. The RL stages use 500 steps, maximum prompt length 10,000 tokens, maximum response length 8,192 tokens, an additional 4,096-token overlong buffer with penalty factor 1.0, prompt batch size 1,024, 16 responses per prompt, mini-batch size 256 prompts, 4 gradient updates per rollout, AdamW, constant learning rate $0.5$1, weight decay 0.1, and linear warmup for 10 rollout steps (Zhao et al., 10 Nov 2025).

The quantitative comparisons emphasize compact-model efficiency. The abstract reports that the 4B RedOne 2.0 model delivers average improvements of about 2.41 over the 7B sub-optimal baseline and average lift of about 8.74 from the Qwen3-4B base model, while using less than half the domain data required by the SFT-centric original RedOne (Zhao et al., 10 Nov 2025). In the stage ablation for Qwen3-4B, the base model scores 69.80 / 51.81 / 38.22 on General-Bench / SNS-Bench / SNS-TransBench; Exploratory Learning only yields 71.25 / 62.27 / 43.35; RL→SFT yields 70.04 / 65.67 / 47.72; and the full RL→SFT→RL pipeline reaches 70.80 / 67.57 / 47.67 (Zhao et al., 10 Nov 2025). The paper also reports that a naive SFT→RL baseline ends below the full RL-prioritized curriculum on all three aggregated benchmarks.

Task-level ablations reinforce the claim that one unified domain model can approach or exceed task-specific finetuning. For Qwen3-4B, RedOne 2.0 scores 60.92 on QCorr versus 60.11 for task-specific fine-tuning, 66.54 on MRC versus 57.54, 79.11 on CHLW versus 67.24, and 47.67 on SNS-TransBench versus 44.25 (Zhao et al., 10 Nov 2025). The exceptions are also reported explicitly: for Hashtag, task-specific fine-tuning reaches 90.12 versus 89.05 for RedOne 2.0, and for QGen, task-specific fine-tuning reaches 49.24 versus 47.17.

The paper also includes an online A/B experiment on personalized re-creation of post titles for a platform with more than 3M users. It reports Advertiser Value (AdvV) ↑ 0.43%, Vague titles ratio ↓ 11.9%, Practical titles ratio ↑ 7.1%, Authentic titles ratio ↑ 12.9%, and Interactive titles ratio ↑ 25.8% (Zhao et al., 10 Nov 2025). At the same time, a negative case study shows that style optimization can omit critical factual information, specifically a case where “transportation card” is lost, and the paper identifies this as a faithfulness vs engagement trade-off.

The limitations of RedOne 2.0 are correspondingly centered on reward design, safety, and operational cost. The paper notes dependence on curated task-specific rewards, limited treatment of bias and adversarial safety, incomplete coverage of evolving SNS task families, and the compute cost of RL with long context and large rollout batches (Zhao et al., 10 Nov 2025). A plausible implication is that RedOne 2.0 is best understood not as a finished universal SNS model but as a post-training template in which RL is placed at the center of domain adaptation rather than at the end of an SFT-heavy pipeline.