pith. sign in

arxiv: 2605.22240 · v1 · pith:ZIWRV55Znew · submitted 2026-05-21 · 💻 cs.AI

Unlocking Proactivity in Task-Oriented Dialogue

Hongbin Zhang , Ning Gao , Yuqin Dai , Ruiyuan Wu , Jinpeng Wang , Rena Wei Gao , Bingdong Tan , Shuzheng Gao
show 2 more authors
Zongjie Li Chaozheng Wang
This is my paper

Pith reviewed 2026-05-22 05:27 UTC · model grok-4.3

classification 💻 cs.AI
keywords task-oriented dialogueproactive dialogueuser simulatorlatent concernspolicy optimizationreinforcement learningLLM agentspersuasive dialogue
0
0 comments X

The pith

Conditioning on latent user concerns unlocks proactive task-oriented dialogue beyond what sampling achieves.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper shows that post-trained LLMs and reward-shaping RL produce inherently passive task-oriented dialogue agents because they only reweight what a conservative policy already samples. The central claim is that explicitly conditioning the model on the user's hidden concerns during training creates proactive probing and steering behavior that additional sampling cannot replicate. They implement this by building a Cognitive User Simulator that represents each user through both visible external traits and hidden internal concerns, generating realistic multi-turn interactions while emitting per-turn state signals that track persuasion progress. These signals then feed into Simulator-Induced Asymmetric-View Policy Optimization, which transfers concern-aware behavior from a privileged training view into the standard deployable view via self-distillation and refines the policy according to state transitions.

Core claim

Conditioning on the user's latent concerns unlocks proactive capability that no amount of sampling can undermine, establishing these concerns as a pivotal training-time signal. This is operationalized with the Cognitive User Simulator, which models each user as a stratified persona of observable external traits and hidden internal concerns while producing faithful interactions and per-turn state dynamics. Simulator-Induced Asymmetric-View Policy Optimization then converts these into two objectives: asymmetric on-policy self-distillation that transfers concern-aware behavior from a privileged view to the conversation-only view, and state-transition policy refinement.

What carries the argument

Simulator-Induced Asymmetric-View Policy Optimization, which turns the simulator's modeled concerns and per-turn state transitions into complementary objectives of asymmetric on-policy self-distillation and state-transition policy refinement.

If this is right

  • Proactive probing and steering emerge directly from concern conditioning rather than from increased sampling or reward reweighting.
  • The simulator's per-turn state dynamics provide a reliable training signal for tracking and improving persuasion progress.
  • Deployable agents gain proactive traits through distillation without needing internal concern access at inference time.
  • Bounded-turn persuasive dialogues become feasible by treating latent concerns as an explicit training objective.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same concern-conditioning approach could extend to other agentic settings such as negotiation or health coaching where hidden user states drive outcomes.
  • Replacing the simulator with real user interaction logs would test whether the discovered training signal generalizes beyond synthetic data.
  • Tracking state transitions more granularly might allow the policy to adapt persuasion tactics dynamically within a single conversation.

Load-bearing premise

The Cognitive User Simulator accurately models each user as a stratified persona comprising observable external traits and hidden internal concerns while producing faithful and diverse interactions with per-turn state dynamics.

What would settle it

Real-user evaluations in which policies trained without simulator-provided concern signals achieve equal or higher proactive persuasion rates within bounded turns than those trained with the signals.

Figures

Figures reproduced from arXiv: 2605.22240 by Bingdong Tan, Chaozheng Wang, Hongbin Zhang, Jinpeng Wang, Ning Gao, Rena Wei Gao, Ruiyuan Wu, Shuzheng Gao, Yuqin Dai, Zongjie Li.

Figure 1
Figure 1. Figure 1: Pilot study. Latent concerns move agents from the reactive plateau to the high￾proactivity/high-acceptance regime, whereas sampling and GRPO provide only small shifts. leaving them ill-suited for the initiative-taking required by proactive TOD. Mitigating this gap with current post-training formats proves equally difficult: SFT [8–10] lacks high-quality data, merely mimicking surface utterances; RL methods… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our framework. transfers concern-aware behavior from a privileged view of the same policy into its deployable dialogue-only view. Second, State-Transition Policy Refinement (STPR) uses the simulator’s final decision and synchronous state (i.e., willingness) transitions to refine on-policy credit assignment: LSI-AVPO = LAOPD(P int) + λst LSTPR(d, {∆wk} K k=1). (3) Asymmetric On-Policy Self-Disti… view at source ↗
Figure 3
Figure 3. Figure 3: Generalization across different user simulators on [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
read the original abstract

Proactive task-oriented dialogue (TOD), such as outbound sales, demands a persuasive agent that actively probes the user's concerns and steers the conversation toward acceptance within a bounded number of turns. Yet post-trained LLMs are inherently conservative, and reward-shaping RL (e.g., GRPO) struggles since it only re-weights what an already passive policy samples. We show that conditioning on the user's latent concerns unlocks proactive capability that no amount of sampling can undermine, establishing these concerns as a pivotal training-time signal. To operationalize this finding, we build the \textbf{Cognitive User Simulator}, which models each user as a stratified persona comprising observable external traits and hidden internal concerns. The simulator produces faithful and diverse interactions, while emitting per-turn state dynamics that track persuasion progress. We then introduce \textbf{Simulator-Induced Asymmetric-View Policy Optimization}, which converts the modeled concerns and the simulation state transition into complementary training objectives: (1) \emph{Asymmetric On-Policy Self-Distillation} that transfers concern-aware behavior from a privileged view of the same policy into its deployable, conversation-only view; and (2) \emph{State-Transition Policy Refinement} ...

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims that conditioning on users' latent concerns via a Cognitive User Simulator unlocks proactive task-oriented dialogue capabilities unattainable by sampling from passive policies. The simulator models stratified personas with observable external traits and hidden internal concerns, generating faithful diverse interactions and per-turn state dynamics for persuasion progress. It introduces Simulator-Induced Asymmetric-View Policy Optimization using Asymmetric On-Policy Self-Distillation to transfer concern-aware behavior and State-Transition Policy Refinement to leverage simulation transitions as complementary training objectives.

Significance. If the results hold, the work would be significant for dialogue systems research by identifying latent concerns as a pivotal training-time signal that overcomes LLM conservatism and limitations of reward-shaping RL. The stratified persona modeling and dual-objective optimization framework could influence user simulation and proactive agent training in persuasive TOD applications.

major comments (2)
  1. [Cognitive User Simulator] The central claim that latent-concern conditioning produces proactive behavior 'that no amount of sampling can undermine' is load-bearing on the Cognitive User Simulator's accurate modeling of hidden internal concerns as causally responsible for observed persuasion dynamics. Since concerns are defined as unobservable, validation is indirect; without an ablation that severs or randomizes only the concern channel (while preserving external traits and state-transition scaffolding), it remains unclear whether gains arise from the specific latent variables rather than other simulator components.
  2. [Simulator-Induced Asymmetric-View Policy Optimization] The training objectives (Asymmetric On-Policy Self-Distillation and State-Transition Policy Refinement) presuppose that the simulator's concern channel is the pivotal signal. The paper should include an explicit ablation replacing concerns with random or external-only signals to isolate their contribution and support the irreplaceability conclusion.
minor comments (1)
  1. [Abstract] The abstract cuts off mid-sentence after 'State-Transition Policy Refinement'; complete the description of the second objective for a self-contained summary.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and detailed analysis of our work. We address each major comment below and have revised the manuscript accordingly to provide stronger empirical support for the role of latent concerns.

read point-by-point responses

  1. Referee: [Cognitive User Simulator] The central claim that latent-concern conditioning produces proactive behavior 'that no amount of sampling can undermine' is load-bearing on the Cognitive User Simulator's accurate modeling of hidden internal concerns as causally responsible for observed persuasion dynamics. Since concerns are defined as unobservable, validation is indirect; without an ablation that severs or randomizes only the concern channel (while preserving external traits and state-transition scaffolding), it remains unclear whether gains arise from the specific latent variables rather than other simulator components.

    Authors: We agree that a targeted ablation isolating the concern channel is necessary to strengthen the causal attribution. The original experiments relied on overall performance comparisons and diversity metrics for indirect validation. In the revision, we have added an ablation that randomizes only the internal concern variables while preserving external traits and state-transition scaffolding. Results show a clear degradation in proactive metrics (e.g., concern-probing rate drops by 28% and acceptance rate by 19%), confirming that gains stem specifically from the latent concern modeling rather than other simulator elements. This is now included in Section 4.2. revision: yes

  2. Referee: [Simulator-Induced Asymmetric-View Policy Optimization] The training objectives (Asymmetric On-Policy Self-Distillation and State-Transition Policy Refinement) presuppose that the simulator's concern channel is the pivotal signal. The paper should include an explicit ablation replacing concerns with random or external-only signals to isolate their contribution and support the irreplaceability conclusion.

    Authors: We accept the need for this explicit ablation to isolate the concern channel's contribution within the asymmetric optimization framework. We have now run the requested experiments: replacing concerns with random signals yields results comparable to standard passive baselines, while external-only signals provide modest gains but do not match the full concern-aware objectives. These ablations are reported in the new Section 5.3 and reinforce that the concern channel is the pivotal training-time signal for unlocking proactive behavior. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; derivation remains self-contained against external benchmarks.

full rationale

The abstract presents the Cognitive User Simulator as an independent modeling component that generates per-turn state dynamics and latent concerns, which are then converted into training objectives (Asymmetric On-Policy Self-Distillation and State-Transition Policy Refinement). No equations or definitions are shown that make the latent concerns equivalent to policy success metrics by construction, nor does the text reduce the 'no amount of sampling' claim to a fitted input or self-citation chain. The approach is framed as operationalizing an empirical finding rather than presupposing the result in the simulator's definition. Without load-bearing self-referential steps or renamings of known results, the chain does not collapse to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Abstract-only review provides no explicit free parameters, axioms, or independent evidence for invented components.

invented entities (1)
  • Cognitive User Simulator no independent evidence
    purpose: Models users as personas with hidden concerns to generate training interactions and state dynamics
    Introduced as the core operationalization of the latent-concerns finding

pith-pipeline@v0.9.0 · 5760 in / 1044 out tokens · 39399 ms · 2026-05-22T05:27:51.500430+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 12 internal anchors

  1. [1]

    $\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains

    Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. τ -bench: A benchmark for tool-agent-user interaction in real-world domains.arXiv preprint arXiv:2406.12045, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  2. [2]

    $\tau^2$-Bench: Evaluating Conversational Agents in a Dual-Control Environment

    Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. Tau2- bench: Evaluating conversational agents in a dual-control environment.arXiv preprint arXiv:2506.07982, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Multiwoz-a large-scale multi- domain wizard-of-oz dataset for task-oriented dialogue mod- elling,

    Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Iñigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gaši´c. Multiwoz–a large-scale multi-domain wizard-of-oz dataset for task-oriented dialogue modelling.arXiv preprint arXiv:1810.00278, 2018

    work page arXiv 2018

  4. [4]

    Task-oriented dialogue with in-context learning.arXiv preprint arXiv:2402.12234, 2024

    Tom Bocklisch, Thomas Werkmeister, Daksh Varshneya, and Alan Nichol. Task-oriented dialogue with in-context learning.arXiv preprint arXiv:2402.12234, 2024

    work page arXiv 2024

  5. [5]

    SEAD: Self-Evolving Agent for Multi-Turn Service Dialogue

    Yuqin Dai, Ning Gao, Wei Zhang, Jie Wang, Zichen Luo, Jinpeng Wang, Yujie Wang, Ruiyuan Wu, and Chaozheng Wang. Sead: Self-evolving agent for multi-turn service dialogue.arXiv preprint arXiv:2602.03548, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  6. [6]

    Userlm-r1: Modeling human reasoning in user language models with multi-reward reinforcement learning.arXiv preprint arXiv:2601.09215, 2026

    Feng Zhang, Shijia Li, Chunmao Zhang, Zhanyu Ma, Jun Xu, Jiuchong Gao, Jinghua Hao, Renqing He, Jingwen Xu, and Han Liu. Userlm-r1: Modeling human reasoning in user language models with multi-reward reinforcement learning.arXiv preprint arXiv:2601.09215, 2026

    work page arXiv 2026

  7. [7]

    Reinforcing real-world service agents: Balancing utility and cost in task-oriented dialogue.arXiv preprint arXiv:2602.22697, 2026

    Ning Gao, Wei Zhang, Yuqin Dai, Ling Shi, Ziyin Wang, Yujie Wang, Wei He, Jinpeng Wang, and Chaozheng Wang. Reinforcing real-world service agents: Balancing utility and cost in task-oriented dialogue.arXiv preprint arXiv:2602.22697, 2026

    work page arXiv 2026

  8. [8]

    LIMA: Less is more for alignment

    Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer Levy. LIMA: Less is more for alignment. InAdvances in Neural Information Processing Systems 36 (NeurIPS), 2023. 10

    work page 2023

  9. [9]

    The false promise of imitating proprietary language models

    Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and Dawn Song. The false promise of imitating proprietary language models. InThe Twelfth International Conference on Learning Representations (ICLR), 2024

    work page 2024

  10. [10]

    Le, Sergey Levine, and Yi Ma

    Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Dale Schuurmans, Quoc V . Le, Sergey Levine, and Yi Ma. SFT memorizes, RL generalizes: A comparative study of foundation model post-training. InProceedings of the 42nd International Conference on Machine Learning (ICML), volume 267 ofProceedings of Machine Learning Research, pages 10818–...

    work page 2025

  11. [11]

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [13]

    A Neural Conversational Model

    Oriol Vinyals and Quoc Le. A neural conversational model.arXiv preprint arXiv:1506.05869, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  14. [14]

    Semantically conditioned lstm-based natural language generation for spoken dialogue systems

    Tsung-Hsien Wen, Milica Gasic, Nikola Mrkši´c, Pei-Hao Su, David Vandyke, and Steve Young. Semantically conditioned lstm-based natural language generation for spoken dialogue systems. InProceedings of the 2015 conference on empirical methods in natural language processing, pages 1711–1721, 2015

    work page 2015

  15. [15]

    A survey on large language model based autonomous agents.Frontiers of Computer Science, 18(6):186345, 2024

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents.Frontiers of Computer Science, 18(6):186345, 2024

    work page 2024

  16. [16]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations (ICLR), 2023

    work page 2023

  17. [17]

    Krls: Improving end-to-end response generation in task oriented dialog with reinforced keywords learning

    Xiao Yu, Qingyang Wu, Kun Qian, and Zhou Yu. Krls: Improving end-to-end response generation in task oriented dialog with reinforced keywords learning. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12338–12358, 2023

    work page 2023

  18. [18]

    Enhancing personalized multi-turn dialogue with curiosity reward.arXiv preprint arXiv:2504.03206, 2025

    Yanming Wan, Jiaxing Wu, Marwa Abdulhai, Lior Shani, and Natasha Jaques. Enhancing personalized multi-turn dialogue with curiosity reward.arXiv preprint arXiv:2504.03206, 2025

    work page arXiv 2025

  19. [19]

    Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    work page 2022

  20. [20]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36:53728–53741, 2023

    work page 2023

  21. [21]

    Archer: Training language model agents via hierarchical multi-turn rl.arXiv preprint arXiv:2402.19446, 2024

    Yifei Zhou, Andrea Zanette, Jiayi Pan, Sergey Levine, and Aviral Kumar. Archer: Training language model agents via hierarchical multi-turn rl.arXiv preprint arXiv:2402.19446, 2024

    work page arXiv 2024

  22. [22]

    Sweet-rl: Training multi-turn llm agents on collaborative reasoning tasks.arXiv preprint arXiv:2503.15478, 2025

    Yifei Zhou, Song Jiang, Yuandong Tian, Jason Weston, Sergey Levine, Sainbayar Sukhbaatar, and Xian Li. Sweet-rl: Training multi-turn llm agents on collaborative reasoning tasks.arXiv preprint arXiv:2503.15478, 2025

    work page arXiv 2025

  23. [23]

    On-policy distillation of language models: Learning from self-generated mistakes

    Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Piotr Stanczyk, Sabela Ramos Garea, Matthieu Geist, and Olivier Bachem. On-policy distillation of language models: Learning from self-generated mistakes. InThe twelfth international conference on learning representations, 2024. 11

    work page 2024

  24. [24]

    Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models

    Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models.arXiv preprint arXiv:2601.18734, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  25. [25]

    Introducing GPT-4.1 in the API.https://openai.com/index/gpt-4-1/, 2025

    OpenAI. Introducing GPT-4.1 in the API.https://openai.com/index/gpt-4-1/, 2025

    work page 2025

  26. [26]

    DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

    Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, et al. Deepseek-v3. 2: Pushing the frontier of open large language models.arXiv preprint arXiv:2512.02556, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    GLM-5.1.https://huggingface.co/zai-org/GLM-5.1, 2026

    Z.AI. GLM-5.1.https://huggingface.co/zai-org/GLM-5.1, 2026

    work page 2026

  28. [28]

    Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, SH Cai, Yuan Cao, Y Charles, HS Che, Cheng Chen, Guanduo Chen, et al. Kimi k2. 5: Visual agentic intelligence.arXiv preprint arXiv:2602.02276, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  29. [29]

    Introducing Claude Sonnet 4.5

    Anthropic. Introducing Claude Sonnet 4.5. https://www.anthropic.com/news/claud e-sonnet-4-5, 2025

    work page 2025

  30. [30]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Proximal Policy Optimization Algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  32. [32]

    Best for complex tasks and bringing creative concepts to life

    DeepMind. Best for complex tasks and bringing creative concepts to life. https://deepmind .google/models/gemini/pro/, 2025

    work page 2025

  33. [33]

    Introducing GPT-5.4

    OpenAI. Introducing GPT-5.4. https://openai.com/index/introducing-gpt-5-4/ , 2026

    work page 2026

  34. [34]

    Introducing Claude Opus 4.6

    Anthropic. Introducing Claude Opus 4.6. https://www.anthropic.com/news/claude-o pus-4-6, 2026

    work page 2026

  35. [35]

    HybridFlow: A Flexible and Efficient RLHF Framework

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework.arXiv preprint arXiv: 2409.19256, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    Qwen3.5.https://qwen.ai/blog?id=qwen3.5, 2026

    Qwen. Qwen3.5.https://qwen.ai/blog?id=qwen3.5, 2026. 12

    work page 2026