AI-FLow-OpticsReview-9B

OpticsReviewer-9B

OpticsReviewer-9B is a multimodal model for six-dimensional quality assessment and structured peer-review feedback on optics manuscripts. It is developed by the Institute of Artificial Intelligence (TeleAI), China Telecom, and described in A learnable paradigm for scientific evaluation: expert judgment encoded in an optics peer-review model.

Given manuscript text and figures, the model generates a quantitative JSON score block followed by Chinese review comments. It is intended to assist researchers and expert reviewers with manuscript assessment and revision.

This release contains the SRRO policy checkpoint at global step 350, exported as a complete Hugging Face BF16 model. It includes the language model, vision encoder, tokenizer and multimodal processor. It does not require a separately downloaded base model or LoRA adapter.

Model details

The source training checkpoint stores FP32 parameters. Export explicitly rounds them to BF16; it is not an FP8 or integer-quantized model. The hardware-specific FlashAttention 3 configuration used during training is omitted so users can select a backend compatible with their installation.

Export validation on 2026-09-23 checked all 760 tensor names and shapes against the architecture, verified that every serialized tensor equals its BF16 merged source, and confirmed finite weights. A Transformers 5.12.1 BF16/SDPA load reported no missing, unexpected or mismatched weights. A synthetic-image generation smoke test with thinking enabled also passed. This short loading test is not an evaluation of full-review quality. See EXPORT_INFO.json and SHA256SUMS for release integrity information.

Output schema

The six dimensions are:

Each score is an integer on a 1–5 scale. A schematic output, with illustrative values rather than a model prediction, is:

# 审稿意见

## 审稿定量指标
```json
{
  "stage": "外审",
  "scores": {
    "novelty": 3,
    "reliability": 2,
    "theory": 3,
    "structure": 4,
    "logic_terms": 4,
    "overall": 3
  },
  "recommendation": "修改后再审",
  "recommendation_id": 1,
  "similar_work": "需结合文献检索进一步核实"
}
```

## 审稿阶段
外审

## 审稿建议
<Summary of contribution, strengths, weaknesses and recommendation.>

## 给作者意见
<Specific, actionable comments grounded in the manuscript.>

The accompanying review instruction uses four recommendation IDs: 0 = 退稿 (reject), 1 = 修改后再审 (revise and resubmit), 2 = 修改后常规发表 (revise for regular publication), and 3 = 修改后推优发表 (revise for priority publication). These IDs are categorical labels, not the numerical recommendation values used in evaluation.

Training data and method

The paper introduces OpticsReview, curated from authentic optics-journal submissions and expert review records. Text and figures are aligned with the corresponding reviews. The six-dimensional labels are extracted from human review reports using field matching; they are not inferred or imputed by a model.

The paper reports 88,167 initially processed samples, reduced to 25,452 training samples and 4,099 test samples after length control and quality filtering. The test set contains 2,132 independent manuscripts. Samples and manuscripts are different units because a manuscript can have multiple reviews. Cleaning includes score-completeness checks, redundancy filtering and down-sampling of reviews with identical scores across all six dimensions. Scores are concentrated around 3 and 4, with fewer examples at the extremes.

OpticsReviewer-9B follows domain SFT with SRRO, which separates quantitative, per-field, format, textual-quality and calibration reward channels during group-relative normalization. It assigns channel advantages to relevant output regions and masks invalid or non-discriminative feedback. A Qwen3.5-9B evaluator scores review text against a fixed ten-criterion rubric. The model weights do not bundle that evaluator.

Selected SRRO settings reported in the paper:

The architecture retains its base-model positional configuration, but this is not evidence that optics-review quality has been validated at the full architectural context limit.

Evaluation: results reported in the paper

The table below reproduces the paper's main results, rather than a new evaluation of this exported directory. Evaluation is retrospective, on the optics peer-review test set, with deterministic decoding at temperature zero. Export integrity and loading checks are separate from benchmark reproduction.

• EM-6d: A prediction matches if it equals any human reviewer's score for that manuscript and dimension; results are averaged across manuscripts and dimensions.
• MAE-6d: First average human scores within each manuscript and dimension, then average absolute prediction errors. 52.60 in the table means 0.5260 score units, not 52.60 score units.
• Decision consensus-only: Binary accept/reject agreement on manuscripts with unanimous human binary decisions. Human recommendations are mapped to 1, 4, 7, 9 and 10, with threshold 5.5; these values are distinct from the output recommendation IDs.
• Spearman-6d and Pairwise-6d: Dimension-averaged measures of cross-manuscript ranking agreement, as defined in the paper.

Falsifiable Review is an additional inference-time agent workflow, not part of these weights. It decomposes criticisms, verifies them against manuscript evidence, applies conservative dimension-aware corrections and regenerates the recommendation. Loading this model alone produces direct reviews and does not reproduce the agent-augmented row. The paper's separate ICLR/DeepReviewer-7B agent results are not benchmark results for this checkpoint.

Usage

Use a recent Transformers version with native Qwen3.5 support. The export tooling uses PyTorch 2.11.0 and Transformers 5.12.1. An SDPA backend avoids a mandatory FlashAttention 3 installation. BF16 weights alone occupy approximately 17.53 GiB; actual inference memory is higher because of activations, image tokens and the KV/recurrent cache.

Install dependencies in your chosen environment:

pip install 'transformers==5.12.1' 'accelerate>=1' safetensors pillow torchvision

Download this release from Hugging Face:

hf download TeleAI-AI-Flow/AI-FLow-OpticsReview-9B --local-dir ./OpticsReviewer-9B

Install a PyTorch build appropriate for your hardware separately. The following example uses local manuscript text and a local figure; the model does not directly parse PDF or Word files.

from pathlib import Path
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_dir = Path("./OpticsReviewer-9B")  # Your downloaded model directory
processor = AutoProcessor.from_pretrained(model_dir)
model = AutoModelForImageTextToText.from_pretrained(
    model_dir,
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="sdpa",
).eval()

messages = [
    {"role": "system", "content": [
        {"type": "text", "text": (model_dir / "review_prompt.txt").read_text()}
    ]},
    {"role": "user", "content": [
        {"type": "text", "text": Path("manuscript.txt").read_text()},
        {"type": "image", "url": str(Path("figure_1.png").resolve())},
    ]},
]
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
answer = processor.batch_decode(
    output[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True
)[0]
print(answer)

The example enables thinking and allows a larger output budget; it is a usage example, not a claim of exact reproduction of the paper's evaluation settings. A backend may expose reasoning separately from final content. Parse the final score JSON and validate all six fields; do not silently invent missing scores when generation is truncated. Preserve equations, tables, captions and figure order during document extraction, and budget input length before generation.

Intended use and limitations

The intended use is non-commercial research and assistance with optics manuscript assessment, pre-submission feedback and analysis of expert-review agreement. Scores and comments are advisory; scientific and editorial decisions remain with qualified humans.

The training records reflect the source journals and reviewer populations, including reviewer disagreement and preferences. Transfer to other disciplines, languages and journals is not established by the reported optics results. Generated criticisms, literature comparisons and recommendations can be unsupported or incorrect. Verify them against the manuscript and relevant literature. Retrospective score agreement does not establish that deployment improves author revisions, editorial outcomes or reviewer workload.

The release does not include confidential manuscripts, human review reports or the full training corpus. As stated in the paper, redistribution of OpticsReview is subject to the confidentiality and licensing conditions of the source records; de-identified annotations and derived statistics may be shared where permitted. Video-related processor configuration is inherited for compatibility; video peer-review performance has not been evaluated here.

License and attribution

OpticsReviewer-9B adapted weights and accompanying original release documentation are offered under CC BY-NC 4.0. Non-commercial sharing and adaptation require attribution, a license link and identification of changes. Commercial use is not granted by this license. See LICENSE for the full terms.

The underlying Qwen3.5-9B model is distributed under Apache-2.0. Its upstream license is retained in LICENSE-QWEN; see NOTICE for attribution and the modifications in this release. This release does not change the upstream license on unmodified Qwen components. The weights are a non-commercial release, rather than an unrestricted open-source license.

Citation

The following cites the supplied 2026 manuscript; no publication venue, DOI or arXiv identifier is asserted.

@misc{song2026opticsreviewer,
  title={A learnable paradigm for scientific evaluation: expert judgment encoded in an optics peer-review model},
  author={Song, Huan and Luo, Jiahao and Lu, Dakuan and Tian, Shuyu and Shao, JiaWei and Li, Xuelong},
  year={2026},
  note={Manuscript. Institute of Artificial Intelligence (TeleAI), China Telecom}
}