ParaWriter-sft-stage1-4B API & Inference Endpoint

它能做什么 / What It Can Do

给它一篇论文，它可以告诉你：

Table with columns: 能力, 说明
能力	说明
识别论文核心认知更新	这篇论文让审稿人觉得"原来是这样"还是"原来还能这样"？
提取问题-动机-洞察	论文解决什么问题、为什么重要、核心 idea 是什么？
分析 Introduction 段落结构	每段的功能：铺背景→指问题→分析 gap→提洞察→预览方法→预览结果→总结贡献
检测承诺-兑现闭环	Introduction 里的每个 claim 在实验部分有没有被证据支撑？
推理方法设计的必然性	为什么模型要这样设计？去掉某个模块会怎样？
诊断叙事弱点	Introduction 写作的强项和弱项分别在哪？

快速体验 / Quick Demo

下面用一篇真实的 ICLR 2024 论文做演示。

输入：一篇完整论文

text

论文标题：The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning Abstract:Alignment tuning has become the de facto standard practice for enabling baselarge language models (LLMs) to serve as open-domain AI assistants. The alignmenttuning process typically involves instruction learning through supervisedfine-tuning (SFT) and preference tuning via reinforcement learning from humanfeedback (RLHF). A recent study, LIMA, shows that using merely 1K examples forSFT can achieve significant alignment performance, suggesting that the effect ofalignment tuning might be "superficial." This raises questions about how exactlythe alignment tuning transforms a base LLM. We analyze the effect of alignment tuning by examining the token distributionshift between base LLMs and their aligned counterparts. Our findings reveal thatbase and aligned models perform nearly identically on 77.7% of token positions,with the shifts occurring primarily on stylistic tokens. Based on these findings,we introduce URIAL (Untuned LLMs with Restyled In-context ALignment), whichachieves effective alignment purely through in-context learning with as few asthree constant stylistic examples. Results demonstrate that base LLMs with URIALcan match or even surpass the performance of SFT/RLHF models. Introduction:Base large language models (LLMs) that are only pre-trained on unsupervised textcorpora typically cannot directly serve as open-domain AI assistants. In order toalign these base LLMs as helpful and harmless assistants, recent research hasfocused on alignment tuning through SFT and RLHF... On the other hand, LIMA proposes the "Superficial Alignment Hypothesis," whichargues that alignment tuning might simply teach base LLMs to select asubdistribution of data formats for interacting with users... To this end, we investigate the effects of alignment tuning by directly comparingtoken distributions between base LLMs and their aligned versions. Surprisingly,we find that 77.7% tokens are unshifted, and shifted tokens are predominantlystylistic (discourse markers, safety disclaimers)... Based on our findings, we pose the research question: how effectively can wealign base LLMs without any tuning? We propose URIAL, which achieves alignmentthrough in-context learning with only three constant stylistic examples... [后续正文包含 method、experiments 等]

任务示例 1：识别论文核心认知更新 / Central Realization

问模型：这篇论文让审稿人产生了什么核心认知更新？是"原来是这样"还是"原来还能这样"？

模型回答：

json

{  "realization_text": "Alignment tuning 并没有真正改变 base LLM 的知识，它只是在极少数 token 位置上做了风格调整（77.7% 的 token 完全不变）。因此不需要微调，仅靠 in-context learning 就能实现对齐。",  "realization_type": "原来是这样",  "why_this_type": "这篇论文揭示了 alignment tuning 的本质机制——表面看很复杂的 SFT+RLHF 流程，实际只改变了风格 token。这是对已知现象（alignment 有效）的深层解释，属于'原来是这样'类型。"}

任务示例 2：分析 Introduction 段落结构 / Paragraph Role Analysis

问模型：这篇论文的 Introduction 每段分别承担什么功能？

模型回答：

json

[  {    "paragraph_id": "intro_p1",    "role": "scene_setting",    "function": "建立研究背景：base LLM 需要通过 alignment tuning 才能用作 AI 助手，这是当前标准做法。"  },  {    "paragraph_id": "intro_p2",    "role": "gap_analysis",    "function": "引入 LIMA 的'浅层对齐假说'，指出缺乏直接的定量证据，提出需要深入分析的动机。"  },  {    "paragraph_id": "intro_p3",    "role": "insight_delivery",    "function": "报告 token distribution shift 分析的核心发现：77.7% token 不变，shifted 的主要是风格 token。"  },  {    "paragraph_id": "intro_p4",    "role": "method_preview",    "function": "基于发现提出研究问题和 URIAL 方法：不微调，仅用 3 个 ICL 示例实现对齐。"  },  {    "paragraph_id": "intro_p5",    "role": "evidence_preview",    "function": "预览评估方案和主要结果：URIAL 在多维度上达到甚至超过 SFT/RLHF 模型的表现。"  },  {    "paragraph_id": "intro_p6",    "role": "contribution_summary",    "function": "总结三点贡献并强调对未来 alignment 研究方向的意义。"  }]

任务示例 3：检测承诺-兑现闭环 / Commitment-Evidence Mapping

问模型：Introduction 里做了哪些 claim？这些 claim 在后文实验中被兑现了吗？

模型回答：

json

{  "explicit_commitments": [    {      "claim": "Alignment tuning 主要改变风格 token，知识来自 base LLM 本身",      "source_paragraph": "intro_p3",      "fulfilled_in": "Section 2.2, Figure 2（token distribution shift 分析）",      "fulfillment_strength": "strong"    },    {      "claim": "URIAL 仅用 3 个常量 ICL 示例即可实现有效对齐",      "source_paragraph": "intro_p4",      "fulfilled_in": "Section 3.3, Table 1（多维度评估对比）",      "fulfillment_strength": "strong"    },    {      "claim": "Base LLM + URIAL 可以匹配甚至超过 SFT/RLHF 模型",      "source_paragraph": "intro_p5",      "fulfilled_in": "Table 1, Figure 1（Mistral-7b-Instruct、Llama-2-70b-chat 对比）",      "fulfillment_strength": "strong"    }  ],  "implicit_promises": [    "方法具有跨模型泛化性（实际在 Llama-2、Mistral 上都测了）"  ]}

使用方法 / Usage

Python (Transformers)

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_name = "bhxdianzhang/ParaWriter-sft-stage1-4B"tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(    model_name, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True) # 把论文的 abstract + introduction + method + experiments 拼成输入paper_content = """<PAPER_BODY>... 你的论文正文 ...</PAPER_BODY>""" instruction = """You are an expert academic paper analyst. Given the body of a research paper,identify the paper's CENTRAL REALIZATION — the single most important cognitive update itdelivers to its reader. Determine whether it is '原来是这样' or '原来还能这样'.""" messages = [{"role": "user", "content": f"{instruction}\n{paper_content}"}]text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer([text], return_tensors="pt").to(model.device) with torch.no_grad():    outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.3, do_sample=True) response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)print(response)

vLLM 部署

bash

vllm serve bhxdianzhang/ParaWriter-sft-stage1-4B --trust-remote-code

python

from openai import OpenAIclient = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy") response = client.chat.completions.create(    model="bhxdianzhang/ParaWriter-sft-stage1-4B",    messages=[{"role": "user", "content": "...论文内容 + 任务指令..."}],    temperature=0.3,)print(response.choices[0].message.content)

训练详情 / Training Details

Table with columns: 参数, 值
参数	值
Base Model	Qwen/Qwen3.5-4B
Framework	LLaMA-Factory
Data	65,678 samples (8,020 NeurIPS/ICML/ICLR papers, 2023-2026)
Tasks	9 types: realization, motivation, decomposition, roles, commitments, necessity, closure, critique, revision
Epochs	3
LR	3e-6
Batch Size	32 (effective)
Max Length	16,384 tokens
Hardware

这是 Stage 1 / This is Stage 1

当前模型是第一阶段训练的结果，专注于理解和诊断——看懂论文的叙事结构。

第二阶段（Stage 2）将加入生成能力——根据论证结构直接生成 abstract/introduction 的段落和大纲。

This is the Stage 1 model, focused on understanding and diagnosis — comprehending paper narrative structure. Stage 2 will add generation capabilities — directly producing abstract/introduction paragraphs based on argument structure.

ParadoxGPT 模型家族 / Model Family

Table with columns: 模型, 用途, 状态
模型	用途	状态
ParaWriter-4B (Stage 1)	论文叙事理解与诊断	✅ 已发布
ParaWriter-9B (Stage 1)	同上，更强的理解能力	🔄 训练中
ParaWriter (Stage 2)	叙事理解 + 控制生成	📋 计划中
ParaReviewer-4B/9B	审稿意见分析与 concern grounding	📋 计划中

Citation

bibtex

@misc{paradoxgpt2026,  title={ParadoxGPT: Learning the Reviewer's "Aha Moment" for Better Scientific Writing},  author={Heng Zhang},  year={2026},  url={https://huggingface.co/bhxdianzhang/ParaWriter-sft-stage1-4B}}

License

Apache 2.0

它能做什么 / What It Can Do

给它一篇论文，它可以告诉你：

Table with columns: 能力, 说明
能力	说明
识别论文核心认知更新	这篇论文让审稿人觉得"原来是这样"还是"原来还能这样"？
提取问题-动机-洞察	论文解决什么问题、为什么重要、核心 idea 是什么？
分析 Introduction 段落结构	每段的功能：铺背景→指问题→分析 gap→提洞察→预览方法→预览结果→总结贡献
检测承诺-兑现闭环	Introduction 里的每个 claim 在实验部分有没有被证据支撑？
推理方法设计的必然性	为什么模型要这样设计？去掉某个模块会怎样？
诊断叙事弱点	Introduction 写作的强项和弱项分别在哪？

快速体验 / Quick Demo

下面用一篇真实的 ICLR 2024 论文做演示。

输入：一篇完整论文

text

论文标题：The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context LearningAbstract:Alignment tuning has become the de facto standard practice for enabling baselarge language models (LLMs) to serve as open-domain AI assistants. The alignmenttuning process typically involves instruction learning through supervisedfine-tuning (SFT) and preference tuning via reinforcement learning from humanfeedback (RLHF). A recent study, LIMA, shows that using merely 1K examples forSFT can achieve significant alignment performance, suggesting that the effect ofalignment tuning might be "superficial." This raises questions about how exactlythe alignment tuning transforms a base LLM.We analyze the effect of alignment tuning by examining the token distributionshift between base LLMs and their aligned counterparts. Our findings reveal thatbase and aligned models perform nearly identically on 77.7% of token positions,with the shifts occurring primarily on stylistic tokens. Based on these findings,we introduce URIAL (Untuned LLMs with Restyled In-context ALignment), whichachieves effective alignment purely through in-context learning with as few asthree constant stylistic examples. Results demonstrate that base LLMs with URIALcan match or even surpass the performance of SFT/RLHF models.Introduction:Base large language models (LLMs) that are only pre-trained on unsupervised textcorpora typically cannot directly serve as open-domain AI assistants. In order toalign these base LLMs as helpful and harmless assistants, recent research hasfocused on alignment tuning through SFT and RLHF...On the other hand, LIMA proposes the "Superficial Alignment Hypothesis," whichargues that alignment tuning might simply teach base LLMs to select asubdistribution of data formats for interacting with users...To this end, we investigate the effects of alignment tuning by directly comparingtoken distributions between base LLMs and their aligned versions. Surprisingly,we find that 77.7% tokens are unshifted, and shifted tokens are predominantlystylistic (discourse markers, safety disclaimers)...Based on our findings, we pose the research question: how effectively can wealign base LLMs without any tuning? We propose URIAL, which achieves alignmentthrough in-context learning with only three constant stylistic examples...[后续正文包含 method、experiments 等]

论文标题：The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning Abstract:Alignment tuning has become the de facto standard practice for enabling baselarge language models (LLMs) to serve as open-domain AI assistants. The alignmenttuning process typically involves instruction learning through supervisedfine-tuning (SFT) and preference tuning via reinforcement learning from humanfeedback (RLHF). A recent study, LIMA, shows that using merely 1K examples forSFT can achieve significant alignment performance, suggesting that the effect ofalignment tuning might be "superficial." This raises questions about how exactlythe alignment tuning transforms a base LLM. We analyze the effect of alignment tuning by examining the token distributionshift between base LLMs and their aligned counterparts. Our findings reveal thatbase and aligned models perform nearly identically on 77.7% of token positions,with the shifts occurring primarily on stylistic tokens. Based on these findings,we introduce URIAL (Untuned LLMs with Restyled In-context ALignment), whichachieves effective alignment purely through in-context learning with as few asthree constant stylistic examples. Results demonstrate that base LLMs with URIALcan match or even surpass the performance of SFT/RLHF models. Introduction:Base large language models (LLMs) that are only pre-trained on unsupervised textcorpora typically cannot directly serve as open-domain AI assistants. In order toalign these base LLMs as helpful and harmless assistants, recent research hasfocused on alignment tuning through SFT and RLHF... On the other hand, LIMA proposes the "Superficial Alignment Hypothesis," whichargues that alignment tuning might simply teach base LLMs to select asubdistribution of data formats for interacting with users... To this end, we investigate the effects of alignment tuning by directly comparingtoken distributions between base LLMs and their aligned versions. Surprisingly,we find that 77.7% tokens are unshifted, and shifted tokens are predominantlystylistic (discourse markers, safety disclaimers)... Based on our findings, we pose the research question: how effectively can wealign base LLMs without any tuning? We propose URIAL, which achieves alignmentthrough in-context learning with only three constant stylistic examples... [后续正文包含 method、experiments 等]

任务示例 1：识别论文核心认知更新 / Central Realization

问模型：这篇论文让审稿人产生了什么核心认知更新？是"原来是这样"还是"原来还能这样"？

模型回答：

json

{  "realization_text": "Alignment tuning 并没有真正改变 base LLM 的知识，它只是在极少数 token 位置上做了风格调整（77.7% 的 token 完全不变）。因此不需要微调，仅靠 in-context learning 就能实现对齐。",  "realization_type": "原来是这样",  "why_this_type": "这篇论文揭示了 alignment tuning 的本质机制——表面看很复杂的 SFT+RLHF 流程，实际只改变了风格 token。这是对已知现象（alignment 有效）的深层解释，属于'原来是这样'类型。"}

{  "realization_text": "Alignment tuning 并没有真正改变 base LLM 的知识，它只是在极少数 token 位置上做了风格调整（77.7% 的 token 完全不变）。因此不需要微调，仅靠 in-context learning 就能实现对齐。",  "realization_type": "原来是这样",  "why_this_type": "这篇论文揭示了 alignment tuning 的本质机制——表面看很复杂的 SFT+RLHF 流程，实际只改变了风格 token。这是对已知现象（alignment 有效）的深层解释，属于'原来是这样'类型。"}

任务示例 2：分析 Introduction 段落结构 / Paragraph Role Analysis

问模型：这篇论文的 Introduction 每段分别承担什么功能？

模型回答：

json

[  {    "paragraph_id": "intro_p1",    "role": "scene_setting",    "function": "建立研究背景：base LLM 需要通过 alignment tuning 才能用作 AI 助手，这是当前标准做法。"  },  {    "paragraph_id": "intro_p2",    "role": "gap_analysis",    "function": "引入 LIMA 的'浅层对齐假说'，指出缺乏直接的定量证据，提出需要深入分析的动机。"  },  {    "paragraph_id": "intro_p3",    "role": "insight_delivery",    "function": "报告 token distribution shift 分析的核心发现：77.7% token 不变，shifted 的主要是风格 token。"  },  {    "paragraph_id": "intro_p4",    "role": "method_preview",    "function": "基于发现提出研究问题和 URIAL 方法：不微调，仅用 3 个 ICL 示例实现对齐。"  },  {    "paragraph_id": "intro_p5",    "role": "evidence_preview",    "function": "预览评估方案和主要结果：URIAL 在多维度上达到甚至超过 SFT/RLHF 模型的表现。"  },  {    "paragraph_id": "intro_p6",    "role": "contribution_summary",    "function": "总结三点贡献并强调对未来 alignment 研究方向的意义。"  }]

[  {    "paragraph_id": "intro_p1",    "role": "scene_setting",    "function": "建立研究背景：base LLM 需要通过 alignment tuning 才能用作 AI 助手，这是当前标准做法。"  },  {    "paragraph_id": "intro_p2",    "role": "gap_analysis",    "function": "引入 LIMA 的'浅层对齐假说'，指出缺乏直接的定量证据，提出需要深入分析的动机。"  },  {    "paragraph_id": "intro_p3",    "role": "insight_delivery",    "function": "报告 token distribution shift 分析的核心发现：77.7% token 不变，shifted 的主要是风格 token。"  },  {    "paragraph_id": "intro_p4",    "role": "method_preview",    "function": "基于发现提出研究问题和 URIAL 方法：不微调，仅用 3 个 ICL 示例实现对齐。"  },  {    "paragraph_id": "intro_p5",    "role": "evidence_preview",    "function": "预览评估方案和主要结果：URIAL 在多维度上达到甚至超过 SFT/RLHF 模型的表现。"  },  {    "paragraph_id": "intro_p6",    "role": "contribution_summary",    "function": "总结三点贡献并强调对未来 alignment 研究方向的意义。"  }]

任务示例 3：检测承诺-兑现闭环 / Commitment-Evidence Mapping

问模型：Introduction 里做了哪些 claim？这些 claim 在后文实验中被兑现了吗？

模型回答：

json

{  "explicit_commitments": [    {      "claim": "Alignment tuning 主要改变风格 token，知识来自 base LLM 本身",      "source_paragraph": "intro_p3",      "fulfilled_in": "Section 2.2, Figure 2（token distribution shift 分析）",      "fulfillment_strength": "strong"    },    {      "claim": "URIAL 仅用 3 个常量 ICL 示例即可实现有效对齐",      "source_paragraph": "intro_p4",      "fulfilled_in": "Section 3.3, Table 1（多维度评估对比）",      "fulfillment_strength": "strong"    },    {      "claim": "Base LLM + URIAL 可以匹配甚至超过 SFT/RLHF 模型",      "source_paragraph": "intro_p5",      "fulfilled_in": "Table 1, Figure 1（Mistral-7b-Instruct、Llama-2-70b-chat 对比）",      "fulfillment_strength": "strong"    }  ],  "implicit_promises": [    "方法具有跨模型泛化性（实际在 Llama-2、Mistral 上都测了）"  ]}

{  "explicit_commitments": [    {      "claim": "Alignment tuning 主要改变风格 token，知识来自 base LLM 本身",      "source_paragraph": "intro_p3",      "fulfilled_in": "Section 2.2, Figure 2（token distribution shift 分析）",      "fulfillment_strength": "strong"    },    {      "claim": "URIAL 仅用 3 个常量 ICL 示例即可实现有效对齐",      "source_paragraph": "intro_p4",      "fulfilled_in": "Section 3.3, Table 1（多维度评估对比）",      "fulfillment_strength": "strong"    },    {      "claim": "Base LLM + URIAL 可以匹配甚至超过 SFT/RLHF 模型",      "source_paragraph": "intro_p5",      "fulfilled_in": "Table 1, Figure 1（Mistral-7b-Instruct、Llama-2-70b-chat 对比）",      "fulfillment_strength": "strong"    }  ],  "implicit_promises": [    "方法具有跨模型泛化性（实际在 Llama-2、Mistral 上都测了）"  ]}

使用方法 / Usage

Python (Transformers)

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_name = "bhxdianzhang/ParaWriter-sft-stage1-4B"tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained(    model_name, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True) # 把论文的 abstract + introduction + method + experiments 拼成输入paper_content = """<PAPER_BODY>... 你的论文正文 ...</PAPER_BODY>""" instruction = """You are an expert academic paper analyst. Given the body of a research paper,identify the paper's CENTRAL REALIZATION — the single most important cognitive update itdelivers to its reader. Determine whether it is '原来是这样' or '原来还能这样'.""" messages = [{"role": "user", "content": f"{instruction}\n{paper_content}"}]text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer([text], return_tensors="pt").to(model.device) with torch.no_grad():    outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.3, do_sample=True) response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)print(response)

vLLM 部署

bash

vllm serve bhxdianzhang/ParaWriter-sft-stage1-4B --trust-remote-code

python

from openai import OpenAIclient = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy") response = client.chat.completions.create(    model="bhxdianzhang/ParaWriter-sft-stage1-4B",    messages=[{"role": "user", "content": "...论文内容 + 任务指令..."}],    temperature=0.3,)print(response.choices[0].message.content)

训练详情 / Training Details

Table with columns: 参数, 值
参数	值
Base Model	Qwen/Qwen3.5-4B
Framework	LLaMA-Factory
Data	65,678 samples (8,020 NeurIPS/ICML/ICLR papers, 2023-2026)
Tasks	9 types: realization, motivation, decomposition, roles, commitments, necessity, closure, critique, revision
Epochs	3
LR	3e-6
Batch Size	32 (effective)
Max Length	16,384 tokens
Hardware

这是 Stage 1 / This is Stage 1

当前模型是第一阶段训练的结果，专注于理解和诊断——看懂论文的叙事结构。

第二阶段（Stage 2）将加入生成能力——根据论证结构直接生成 abstract/introduction 的段落和大纲。

ParadoxGPT 模型家族 / Model Family

Table with columns: 模型, 用途, 状态
模型	用途	状态
ParaWriter-4B (Stage 1)	论文叙事理解与诊断	✅ 已发布
ParaWriter-9B (Stage 1)	同上，更强的理解能力	🔄 训练中
ParaWriter (Stage 2)	叙事理解 + 控制生成	📋 计划中
ParaReviewer-4B/9B	审稿意见分析与 concern grounding	📋 计划中

Citation

bibtex

@misc{paradoxgpt2026,  title={ParadoxGPT: Learning the Reviewer's "Aha Moment" for Better Scientific Writing},  author={Heng Zhang},  year={2026},  url={https://huggingface.co/bhxdianzhang/ParaWriter-sft-stage1-4B}}

@misc{paradoxgpt2026,  title={ParadoxGPT: Learning the Reviewer's "Aha Moment" for Better Scientific Writing},  author={Heng Zhang},  year={2026},  url={https://huggingface.co/bhxdianzhang/ParaWriter-sft-stage1-4B}}

License

Apache 2.0

ParaWriter-sft-stage1-4B

README

它能做什么 / What It Can Do

快速体验 / Quick Demo

输入：一篇完整论文

任务示例 1：识别论文核心认知更新 / Central Realization

任务示例 2：分析 Introduction 段落结构 / Paragraph Role Analysis

任务示例 3：检测承诺-兑现闭环 / Commitment-Evidence Mapping

使用方法 / Usage

Python (Transformers)

vLLM 部署

训练详情 / Training Details

这是 Stage 1 / This is Stage 1

ParadoxGPT 模型家族 / Model Family

Citation

License

Explore FriendliAI today

README

它能做什么 / What It Can Do

快速体验 / Quick Demo

输入：一篇完整论文

任务示例 1：识别论文核心认知更新 / Central Realization

任务示例 2：分析 Introduction 段落结构 / Paragraph Role Analysis

任务示例 3：检测承诺-兑现闭环 / Commitment-Evidence Mapping

使用方法 / Usage

Python (Transformers)

vLLM 部署

训练详情 / Training Details

这是 Stage 1 / This is Stage 1

ParadoxGPT 模型家族 / Model Family

Citation

License