Hunyuan (Tencent)

Hunyuan (Chinese: 腾讯混元大模型; Tencent Hunyuan, Tencent HY) is a family of foundation models and large language models (LLM) developed by the Tencent Hunyuan Foundation Model Team and available through the Tencent Cloud service, as well as in the form of open weights on the Hugging Face and GitHub platforms. The family covers models for text generation and understanding, logical and mathematical reasoning, coding, and long-context processing, along with multimodal extensions (images, video, 3D). Hunyuan is positioned as Tencent's flagship generative-AI line and is integrated into the company's product ecosystem (Yuanbao, WeChat, Tencent Meeting, Tencent Cloud).

History and development timeline

Hunyuan was developed in the context of the global LLM race and China's strategy of technological self-reliance in artificial intelligence. The lineup evolved from proprietary dense models to open Mixture-of-Experts (MoE) and hybrid Transformer–Mamba architectures.

First generation (2023)

The Hunyuan series was publicly announced on September 7, 2023, at the Tencent Global Digital Ecosystem Summit in Shenzhen. The first version was a proprietary dense model with more than 100 billion parameters, pre-trained on over 2 trillion tokens. The model focused on Chinese-language tasks, logical reasoning, and content generation; it was integrated into more than 50 Tencent products and made available through the Tencent Cloud API.

Hunyuan-Large and the shift to MoE (2024)

In February 2024 the flagship model was moved to a Mixture-of-Experts architecture, and in May 2024 Tencent launched the Yuanbao chatbot built on Hunyuan. In November 2024 Tencent released Hunyuan-Large — the largest open Transformer-MoE model at the time, with 389 billion total and 52 billion active parameters. The model was published with open weights on Hugging Face and GitHub and accompanied by an arXiv preprint. The release marked Tencent's strategic turn toward open development.

Hybrid and reasoning models (2025)

In 2025 the lineup was expanded with several key releases:

• Hunyuan-TurboS (February 2025) — the first industrial large-scale hybrid Transformer-Mamba-MoE model, with 560 billion total and 56 billion active parameters, pre-trained on 16 trillion tokens. It introduced an adaptive Chain-of-Thought (CoT) mechanism for dynamic switching between fast and deep thinking.
• Hunyuan-T1 (March 2025) — a specialized reasoning model built on top of TurboS with large-scale reinforcement learning (96.7% of post-training compute devoted to reinforcement learning).
• Hunyuan-A13B (June 2025) — a compact MoE model (80 billion total / 13 billion active parameters) balancing performance and cost, with support for fast and slow thinking.
• Small dense models (August 2025) — 0.5B, 1.8B, 4B, and 7B variants for edge devices and high-concurrency deployment scenarios, with 256K context support.

Hunyuan 2.0 and subsequent developments (2025–2026)

On December 5, 2025, Tencent officially released the second generation of its commercial model — Hunyuan 2.0 (HY 2.0), based on an MoE architecture (406 billion total / 32 billion active parameters) with a 256K-token context window; the Think and Instruct variants had become available on Tencent Cloud in November. The lineup is split into HY 2.0 Think (deep reasoning, code) and HY 2.0 Instruct (dialogue, creative generation). The models are accessible via the Tencent Cloud API and integrated into the Yuanbao and ima applications. Compared with the predecessor model (Hunyuan-T1-20250822), the developers highlight improved pre-training data and a two-stage reinforcement-learning pipeline combining verifiable rewards (RLVR) and human feedback (RLHF).

In December 2025 Tencent also reorganized its AI divisions (creating the AI Infra, AI Data, and Data Computing Platform departments) and appointed former OpenAI researcher Yao Shunyu as Chief AI Scientist, heading the AI Infra and Large Language Model departments. In the first half of 2026 the company released Hy-MT2, the second generation of its translation models (a 1.8B dense model and a 30B-A3B MoE), under the Apache 2.0 license.

Summary timeline

Date	Event
September 2023	Public announcement of Hunyuan as a proprietary LLM (>100B parameters); launch of the Tencent Cloud API
February 2024	Flagship model moved to an MoE architecture
April 2024	Rebranding: "advanced" → hunyuan-pro, "standard" → hunyuan-standard; hunyuan-lite added
May 2024	Launch of the Yuanbao chatbot; hunyuan-lite moved to MoE, extended to 256K context
September 2024	Launch of the new hunyuan-turbo (next-generation MoE)
November 2024	Open release of Hunyuan-Large (389B/52B MoE), preprint arXiv:2411.02265
February–March 2025	Hunyuan-TurboS (hybrid Mamba-MoE); Hunyuan-T1 (reasoning)
June 2025	Hunyuan-A13B (80B/13B, fine-grained MoE)
August 2025	Small dense models 0.5B–7B
September 2025	Hunyuan-MT (specialized translation model)
November 2025	HunyuanOCR (1B, open weights)
November–December 2025	Hunyuan 2.0 (HY 2.0 Think / Instruct), 406B/32B MoE — official release on December 5; reorganization of Tencent's AI divisions
First half of 2026	Hy-MT2 translation family (1.8B and 30B-A3B, Apache 2.0)

Source: Tencent Cloud product dynamics; official repositories.

Theoretical foundations and architecture

Mixture-of-Experts architecture

The key models of the series (Hunyuan-Large, A13B, HY 2.0) use the Mixture-of-Experts (MoE) architecture, in which some Transformer layers contain multiple "experts" (sub-networks), while a router (gating network) selects a subset of active experts for each token. The general formula of an MoE layer:

$\mathrm{O}\mathrm{u}\mathrm{t}\mathrm{p}\mathrm{u}\mathrm{t}=\sum _{i=1}^{N}G(x{)}_i\cdot E_i(x)$

where $G(x)$ is the gating (routing) function, $E_i$ is the $i$-th expert, and $N$ is the total number of experts. Under this approach, the total number of model parameters $P_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}$ substantially exceeds the number of parameters active in a single forward pass $P_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}$:

$P_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}\ll P_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}$

which allows model capacity to grow without a proportional increase in inference compute.

Hunyuan-Large implements a scheme with 1 shared expert and 16 specialized experts, activating 1 specialist per token (top-1 routing). Additional innovations include mixed routing with a recycle mechanism (reprocessing of rejected tokens) and expert-specific learning rates to improve the balance of expert contributions.

Hybrid Transformer–Mamba architecture

Hunyuan-TurboS and T1 introduce a hybrid architecture combining Transformer (attention) blocks and Mamba (state-space model) blocks. Mamba provides linear complexity in sequence length:

$h_t=A{h}_{t-1}+B{x}_t$

where $A$, $B$ are learnable matrices, $h_t$ is the hidden state at step $t$, and $x_t$ is the input token. This yields $O(1)$ memory scaling with length (versus $O(N)$ in a standard Transformer).

TurboS contains 128 layers organized into blocks: 57 Mamba2 layers, 7 Attention layers, and 64 FFN-MoE layers with 32 experts. AMF (Attention → Mamba2 → FFN) and MF (Mamba2 → FFN) blocks alternate to balance global and local context.

Attention mechanisms and KV cache

Hunyuan-Large uses Grouped Query Attention (GQA) — an attention variant in which several queries share common keys and values — and Cross-Layer Attention (CLA) to reduce KV-cache size. The standard KV-cache size for a self-attention layer with $H$ heads, sequence length $L$, and head size $d_h$ is:

$S_{\mathrm{K}\mathrm{V}}\propto 2\cdot H\cdot L\cdot d_h$

where $S_{\mathrm{K}\mathrm{V}}$ is the KV-cache size. With GQA and $H_g<H$ groups, the size decreases to:

$S_{\mathrm{K}\mathrm{V}}^{\mathrm{G}\mathrm{Q}\mathrm{A}}\propto 2\cdot H_g\cdot L\cdot d_h$

CLA additionally reuses the KV cache across layers. Together, these optimizations provide memory savings of up to roughly 95%.

Positional encoding and long context

The models use Rotary Position Embedding (RoPE) with scaling to support long sequences. Context training is performed in stages (curriculum learning): from 32K to 256K tokens. The activation function is SwiGLU. The tokenizer vocabulary is 128K (tiktoken with a Chinese-language extension).

Training and scaling

For MoE models, scaling laws are studied in the form of empirical power-law dependencies:

$Q\approx a-b\cdot P_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}^{-\alpha }-c\cdot N^{-\beta }$

where $Q$ is a quality metric, $P_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}}$ is the number of active parameters, $N$ is the data volume, and $a,b,c,\alpha ,\beta$ are experimentally fitted parameters. Hunyuan-Large emphasizes the use of synthetic data at a scale of 1.5 trillion tokens — orders of magnitude more than in previous MoE publications.

Training pipeline and alignment

The training process of Hunyuan models includes several stages typical of modern LLMs:

Pre-training

Large-scale training on a broad multilingual corpus (Chinese, English, and other languages). Hunyuan-Large was pre-trained on 7 trillion tokens (including 1.5 trillion synthetic). TurboS — on 16 trillion tokens. The exact dataset composition is not disclosed.

Supervised fine-tuning (SFT)

Fine-tuning on more than 1 million instructions (instruction–response pairs), teaching the model to follow user instructions.

Alignment with reinforcement learning

To align model behavior with human preferences, the following are used:

Direct Preference Optimization (DPO) — an alignment method without an explicit reward model, used in Hunyuan-Large.

Reinforcement Learning (RL) — Hunyuan-T1 uses large-scale reinforcement learning with curriculum learning; according to the developer, 96.7% of post-training compute is devoted to RL.

Adaptive Long-Short Chain-of-Thought — TurboS implements a mechanism for dynamic switching between fast and deep thinking, allowing the model to adapt reasoning depth to task complexity.

Model family

The family is divided into closed cloud API models and open-weight models.

Core models (overview)

Model	Release date	Architecture	Parameters (total / active)	Context	Availability
Hunyuan (2023)	September 2023	Dense Transformer	>100B / —	not disclosed	Proprietary API
Hunyuan-Large	November 2024	Transformer-MoE	389B / 52B	256K (pretrain), 128K (instruct)	Open weights (GitHub, HF)
Hunyuan-TurboS	February 2025	Hybrid Transformer-Mamba-MoE	560B / 56B	256K (arch.), 32K/16K (API)	Proprietary API (the GitHub repository hosts the technical report)
Hunyuan-T1	March 2025	TurboS base + large-scale RL	—	32K/64K (API)	Proprietary API
Hunyuan-A13B	June 2025	Fine-grained MoE	80B / 13B	256K (open), 224K/32K (API)	Open weights + API
Small models (0.5–7B)	August 2025	Dense	0.5–7B	256K	Open weights
HY 2.0 Think	December 2025	MoE	406B / 32B	256K (announced); API: 128K in / 64K out	Proprietary API
HY 2.0 Instruct	December 2025	MoE	406B / 32B	256K (announced); API: 128K in / 16K out	Proprietary API

Sources: arXiv:2411.02265; arXiv:2505.15431; Tencent Cloud docs; GitHub.

Note. For several models, the architectural context limits (from the technical reports) differ from the API service limits (from the product docs). This is not a contradiction; it reflects differences between research and production configurations.

Specialized models

Hunyuan-MT (September 2025) — a specialized machine-translation model covering 33 languages, which took 1st place at WMT25 in 30 of 31 categories.

Hy-MT2 (2026) — the second generation of the translation line: a 1.8B dense model and a 30B-A3B MoE, released under the Apache 2.0 license.

HunyuanOCR (November 2025) — a compact open OCR model with 1B parameters.

HunyuanImage, HunyuanVideo, Hunyuan3D, HY-World — multimodal models of the family for generating images, video, 3D objects, and navigable 3D worlds (HY-World 2.0, 2026), respectively.

Positioning and lineage

The Tencent Cloud documentation traces the following evolutionary links:

• hunyuan-a13b is listed as an upgrade of hunyuan-standard-256K.
• hunyuan-t1 is built on top of TurboS with intensified RL post-training.
• HY 2.0 Think/Instruct is the branch in which the base model was updated from TurboS to Hunyuan 2.0.
• The open branch (Hunyuan-Large, A13B, compact dense models) develops in parallel with the closed API branch, providing research access.

What's new in key generations

Hunyuan-Large (2024)

Compared with the first generation and previous MoE approaches, Hunyuan-Large introduces:

• A scale of 389 billion parameters (52 billion active) — the largest open Transformer-MoE model at the time of publication.
• Support for 256K context (pretrain) and 128K (instruct) — well above the typical values of mainstream 2024 LLMs.
• KV-cache compression based on GQA + CLA (about 95% memory savings).
• Large-scale synthetic data (1.5 trillion synthetic tokens).
• Mixed expert routing and expert-specific learning rates.

Hunyuan-TurboS (2025)

The key architectural shift:

• A Mamba2 + Attention + MoE hybrid: 560B total / 56B active, 128 layers.
• Pre-training on 16 trillion tokens.
• Adaptive Long-Short Chain-of-Thought for dynamic control of reasoning depth.
• A claimed 1.8–2× decoding speedup over the previous Turbo version.

Hunyuan-T1 (2025)

A reasoning model built on TurboS:

• 96.7% of post-training compute devoted to reinforcement learning.
• A twofold decoding speedup within a comparable deployment envelope.
• Curriculum learning to improve reasoning strategies.

HY 2.0 (2025)

An update of the commercial flagship:

• MoE architecture: 406B total / 32B active parameters.
• A 256K-token context window.
• A two-stage reinforcement-learning pipeline (RLVR + RLHF) and length-penalty strategies against verbose outputs.
• Substantial gains on specialized benchmarks: IMO-AnswerBench 73.4 (a claimed gain of nearly 20% over Hunyuan-T1-20250822), SWE-bench Verified 53.0 (up from 6.0), τ²-Bench 72.4 (up from 17.1).

Evaluation and benchmarks

Evaluation was performed on standard benchmarks using zero-shot and few-shot protocols. The results are based on the official technical reports and repositories; for several models, independent external verification is still limited.

Hunyuan-Large benchmarks (pretrain)

Benchmark	Hunyuan-Large	Llama 3.1-405B	Llama 3.1-70B	Mixtral-8x22B	DeepSeek-V2
MMLU	88.4	85.2	79.3	77.8	78.5
CMMLU	90.2	—	—	—	84.0
C-Eval	91.9	—	—	—	81.7
GSM8K	92.8	89.0	83.7	83.7	79.2
MATH	69.8	53.8	41.4	42.5	43.6
HumanEval	71.4	61.0	58.5	53.1	48.8

Source: arXiv:2411.02265.

According to the technical report, Hunyuan-Large (pretrain) leads in 15 of 19 benchmarks when compared with Llama 3.1-405B, Mixtral-8x22B, and DeepSeek-V2.

Hunyuan-Large-Instruct benchmarks

Benchmark	Hunyuan-Large-Instruct
MMLU	89.9
MATH	77.4
HumanEval	90.0
Arena-Hard	81.8

Source: arXiv:2411.02265; Hugging Face model card.

According to the developers, Hunyuan-Large-Instruct outperforms LLaMA 3.1-405B on MMLU by 2.6 percentage points.

TurboS and T1 benchmarks

Benchmark	TurboS	T1
LMSYS Chatbot Arena	1356 (top-7)	—
Average across 23 benchmarks	77.9%	—
MMLU-Pro	—	87.2
GPQA-Diamond	—	69.3
MATH-500	—	96.2
LiveCodeBench	—	64.9
Arena-Hard	—	91.9

Sources: arXiv:2505.15431; T1 official page.

Hunyuan-A13B benchmarks

Benchmark	Hunyuan-A13B	Hunyuan-Large
MMLU	88.17	88.4
MMLU-Pro	67.23	60.2
BBH	87.56	86.3
MATH	72.35	69.8
GPQA	49.12	—
MBPP	83.86	—

Source: GitHub Hunyuan-A13B README.

A13B outperforms Hunyuan-Large on several post-training-sensitive tasks (MMLU-Pro, MATH, MBPP), which is consistent with its positioning as a more recent, application-oriented variant.

Compact dense branch benchmarks

Model	MMLU	MMLU-Pro	BBH	GSM8K	MATH
Hunyuan-0.5B	54.02	31.15	45.92	55.64	42.95
Hunyuan-1.8B	64.62	38.65	74.32	77.26	62.85
Hunyuan-4B	74.01	51.91	75.17	87.49	72.25
Hunyuan-7B	79.82	57.79	82.95	88.25	74.85

Source: GitHub Hunyuan-7B README.

HY 2.0 benchmarks

The figures are reported by the developer (the only explicit sources are the official Hunyuan account and press coverage):

Benchmark	HY 2.0	Hunyuan-T1-20250822 (predecessor)
IMO-AnswerBench	73.4	— (a gain of ≈20% is claimed)
SWE-bench Verified	53.0	6.0
τ²-Bench	72.4	17.1

Note. These figures are classified as "developer-reported, pending broad external verification."

Technical usage details

Context windows by branch

Model	Input (tokens)	Output (tokens)
HY 2.0 Think	128K	64K
HY 2.0 Instruct	128K	16K
T1 (API)	32K	64K
A13B (API)	224K	32K
TurboS (API)	32K	16K
Lite (API)	250K	6K
Hunyuan-Large (open)	256K (pretrain) / 128K (instruct)	—
Compact 0.5–7B	256K	—

Source: Tencent Cloud product overview.

Supported tools

The Tencent Cloud API documents:

• Function Calling (tools, tool_choice).
• AI Search Enhancement — built-in retrieval over external sources (retrieval-augmented generation).
• SearchInfo and Citation — citation markers in responses.
• EnableMultimedia — multimedia insertions in the output.
• EnableThinking — a chain-of-thought toggle for A13B.
• EnableRecommendedQuestions — generation of recommended follow-up questions.

API compatibility

The Hunyuan API supports an OpenAI-compatible format:

• Base URL: https://api.hunyuan.cloud.tencent.com/v1.
• /v1/chat/completions and /v1/embeddings are supported.
• Embedding model: hunyuan-embedding, dimensionality 1024.
• Streaming via SSE.
• Default concurrency: 5 per account.

For the open Hunyuan-Large, the standard input/output format follows the PyTorch/Transformers frameworks. Quantization is supported: FP8, INT4 (GPTQ/AWQ).

Applications and integration

Hunyuan models are integrated into the Tencent ecosystem and available to external developers; by the end of 2025 Hunyuan was used in more than 900 internal Tencent applications. The main documented scenarios:

Tencent products

• Yuanbao — a chatbot supporting Hunyuan and DeepSeek.
• Tencent Meeting — meeting summarization.
• Tencent Docs — text generation and analysis.
• ima — an AI assistant for knowledge management (intelligent workspace).

Enterprise applications

• Text generation: articles, advertising copy, scripts, product descriptions.
• Intelligent customer support: chatbots, FAQs, response automation.
• Code generation and analysis (especially HY 2.0 Think and T1).
• Mathematical and logical reasoning, analytical tasks.
• Multilingual translation (33 languages; Hunyuan-MT / Hy-MT2).

Open models

The open variants are used for research, fine-tuning, and deployment on edge devices (the small dense models). The multimodal extensions (HunyuanImage, HunyuanVideo, Hunyuan3D) are applied to 3D modeling and video generation.

Limitations and open problems

• Closed commercial models. For HY 2.0 and T1 there are no open weights or detailed architectural specifications; access is limited to the API, hindering independent reproducibility.
• Opaque training data. While the use of large-scale synthetic data is emphasized, the exact dataset composition and the shares of languages and domains are not disclosed.
• Compute requirements. The open models require substantial resources: at least 32 GPUs for full fine-tuning of Hunyuan-Large; tens of GB of VRAM even with FP8.
• Hallucinations. Like other LLMs, the models are prone to generating plausible but factually incorrect statements. Systematic studies of this aspect for Hunyuan in peer-reviewed open literature remain limited.
• Differences between architectural and service limits. For several models (TurboS, A13B), the API service limits are substantially lower than the architectural capabilities.
• Regional availability. The primary focus is the Chinese market; the API for users outside China carries language and legal restrictions.
• Lack of detailed safety reports. The documentation outlines general safety and filtering principles, but there are no public technical safety reports comparable in scope to those of some international models.
• Open-source branch compatibility. After the releases, integration issues with transformers / vllm were recorded, where the libraries did not recognize the hunyuan_v1_dense architecture type and required trust_remote_code or special branches.

Incidents and known issues

As of the available materials (mid-2026), no major public incidents specifically associated with Hunyuan (large-scale data leaks, unique security threats) have been recorded.

Documented product fixes

The Tencent Cloud product dynamics record micro-fixes:

• 2024-11-20: hunyuan-turbo-latest updated to fix repetition caused by special characters, improve the stability of Markdown output, and reduce unjustified refusals.
• 2025-04-03: a T1 update fixing the mixing of Simplified/Traditional Chinese and of Chinese/English, plus improved project-level code generation.
• 2025-04-20: Search Enhancement switched from default-on to default-off — effectively an API behavior change for integrations.

MoE security research

At the level of research critique (preprints), Hunyuan-A13B appears in work on attacks against safety localization in MoE models. In particular, the papers Large Language Lobotomy and GateBreaker report high attack success rates when "silencing" safety experts. This does not constitute a confirmed production incident, but it shows that the security of MoE routing is regarded as a vulnerable research vector.

Ethical and regulatory aspects

Hunyuan operates within China's legal and regulatory context, including the national rules for generative AI governance (CAC) and content filtering, as well as Tencent's internal data-security policies. In September 2023 the model was registered under the Interim Measures for the Management of Generative Artificial Intelligence Services, and in December 2023 it passed the national Large Model Standard Compliance Evaluation. The Tencent Cloud documentation stresses that the use of the Hunyuan API must comply with applicable law and platform policy.

Specific measures include:

• Built-in content moderation in streaming output, with a possible FinishReason=sensitive.
• Alignment via RLHF/DPO to reduce hallucinations and undesirable behavior.
• Training-data filtering to minimize bias.
• Open licenses (Tencent Hunyuan Community License Agreement) for the open models, with the caveat that the license territory excludes the European Union, the United Kingdom, and South Korea for a number of variants.

Dedicated independent studies of bias and fairness for Hunyuan are still scarce; more research is expected as open weights and independent testing expand.

Community reception

The most substantive external signal for the closed branch is TurboS's result on LMSYS Chatbot Arena (1356, top-7 globally). For the open branch, an indirect indicator is GitHub activity: on the order of 1.6 thousand stars for Hunyuan-Large, about 800 for A13B, and about 700 for Hunyuan-MT (as of early 2026); the Hunyuan-MT-7B translation model also topped the Hugging Face trending chart in September 2025. This reflects a noticeable, though not dominant by the standards of the open-LLM ecosystem, level of engagement. Subjective community assessments (Hugging Face, Reddit) note strengths in Chinese-language and agent tasks.

Outlook and research directions

Directions for further development outlined in public materials:

• Further hybridization of architectures (Mamba + Transformer + MoE) and study of scaling laws for MoE.
• Expansion of multimodal capabilities: integration of Hunyuan3D, HunyuanVideo, and HunyuanImage with the language models; a full-modality model, Hunyuan-O, announced in May 2025.
• Improved tool use and agentic capabilities.
• Lower inference costs: quantization (2-bit for edge), optimizations via TRT-LLM/vLLM.
• Expansion of the open model portfolio.
• Development and publication of detailed safety and alignment reports.

External links

• Official Tencent Cloud Hunyuan documentation
• Hunyuan-Large on Hugging Face
• Hunyuan-Large GitHub repository
• Hunyuan-A13B GitHub repository
• Hunyuan-T1 official page
• Hunyuan launch announcement (September 2023)

See also

• Mixture-of-Experts (MoE)
• Transformer architecture
• DeepSeek
• LLaMA (Meta)
• Yi (01.AI)

Literature

• Sun, X. et al. Hunyuan-Large: An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent. arXiv:2411.02265, 2024. https://arxiv.org/abs/2411.02265
• Tencent Hunyuan Team. Hunyuan-TurboS: Advancing Large Language Models through Mamba-Transformer Synergy and Adaptive Chain-of-Thought. arXiv:2505.15431, 2025. https://arxiv.org/abs/2505.15431
• Tencent Hunyuan Team. Hunyuan-MT Technical Report. arXiv:2509.05209, 2025. https://arxiv.org/abs/2509.05209
• Vaswani, A. et al. Attention Is All You Need. NeurIPS, 2017. https://arxiv.org/abs/1706.03762
• Rafailov, R. et al. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. NeurIPS, 2023. https://arxiv.org/abs/2305.18290