Differences

This shows you the differences between two versions of the page.

--- nlp:human-in-the-loop [2025/05/01 09:44] – [RLHF] jmflanig
+++ nlp:human-in-the-loop [2025/05/31 07:43] (current) – [RLHF] jmflanig
@@ Line 35: / Line 35: @@
   * [[https://arxiv.org/pdf/2312.14925.pdf|Kaufmann et al 2023 - A Survey of Reinforcement Learning from Human Feedback]]
   * [[https://arxiv.org/pdf/2307.04964|2023 - Secrets of RLHF in Large Language Models Part I: PPO]]
-  * [[https://arxiv.org/pdf/2401.06080|Wang 2024 - Secrets of RLHF in Large Language Models Part II: Reward Modeling]]
+  * [[https://arxiv.org/pdf/2401.06080|Wang et al 2024 - Secrets of RLHF in Large Language Models Part II: Reward Modeling]]
-  * [[https://arxiv.org/pdf/2404.08555|Chaudhari et al 2024 - RLHF Deciphered: A Critical Analysis of Reinforcement Learning from Human Feedback for LLMs]]
+  * [[https://arxiv.org/pdf/2404.08555|Chaudhari et al 2024 - RLHF Deciphered: A Critical Analysis of Reinforcement Learning from Human Feedback for LLMs]] Really good explanation of PPO in section 6 (why each piece is necessary)
+  * [[https://arxiv.org/pdf/2406.11191|Jiang et al 2024 - A Survey on Human Preference Learning for Large Language Models]]
   * [[https://rlhfbook.com/book.pdf|Lambert - Reinforcement Learning from Human Feedback]] Nathan's RLHF book. Very good [[https://rlhfbook.com/|website]]
@@ Line 44: / Line 45: @@
   * [[https://arxiv.org/pdf/2009.01325.pdf|Stiennon et al 2020 - Learning to summarize from human feedback]]
   * [[https://arxiv.org/pdf/2112.00861.pdf|Askell et al 2021 - A General Language Assistant as a Laboratory for Alignment]] Introduced the acronym RLHF
-  * InstructGPT: [[https://arxiv.org/pdf/2203.02155.pdf|Ouyang et al 2022 - Training language models to follow instructions with human feedback]] This is essentially inverse-reinforcement learning (such as [[https://www.ri.cmu.edu/pub_files/2009/7/learch.pdf|this]]) applied to LMs.
+  * InstructGPT: [[https://arxiv.org/pdf/2203.02155.pdf|Ouyang et al 2022 - Training language models to follow instructions with human feedback]] This is essentially inverse-reinforcement learning (such as [[https://www.ri.cmu.edu/pub_files/2009/7/learch.pdf|this]]) applied to LMs. Background papers:
     * Used PPO: [[https://arxiv.org/pdf/1707.06347.pdf|Schulman et al 2017 - Proximal Policy Optimization Algorithms]]
+    * [[https://arxiv.org/pdf/1909.08593|Ziegler 2019 - Fine-Tuning Language Models from Human Preferences]]
+    * [[https://arxiv.org/pdf/2009.01325|Stiennon et al 2020 - Learning to Summarize from Human Feedback]]
   * [[https://arxiv.org/pdf/2204.05862.pdf|Bai et al 2022 - Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback]]
   * [[https://arxiv.org/pdf/2305.18290.pdf|Rafailov et al 2023 - Direct Preference Optimization: Your Language Model is Secretly a Reward Model]]
   * [[https://arxiv.org/pdf/2306.01693.pdf|Wu et al 2023 - Fine-Grained Human Feedback Gives Better Rewards for Language Model Training]]
   * [[https://arxiv.org/pdf/2410.04612|Gao et al 2024 - Regressing the Relative Future: Efficient Policy Optimization for Multi-turn RLHF]]
+  * [[https://arxiv.org/pdf/2505.22338|Wang et al 2025 - Text2Grad: Reinforcement Learning from Natural Language Feedback]]
 === Crowdsourcing & Data Collection ===