Yanming Wan

Personalization methods for LLMs often rely on extensive user history. We introduce Curiosity-driven User-modeling Reward as Intrinsic Objective (CURIO) to encourage actively learning about the user within multi-turn dialogs. 📜 arxiv.org/abs/2504.03206 🌎 sites.google.com/cs.washington.…

Excited to share that "Learning to summarize user information for personalized reinforcement learning from human feedback" is accepted to ICLR. TL;DR We can train a conversation summarizer with RL to capture diverse user preferences for pluralistic LLM alignment. w/ @natashajaques @mickel_liu @yanming_wan @PeterAhnnDD arxiv: arxiv.org/abs/2507.13579 website: sites.google.com/stanford.edu/p… code: github.com/nam630/plurali…

It’s a great honor to receive this award! Our related paper "Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward" has been accepted to #NeurIPS2025. Looking forward to sharing and discussing with everyone! 📜 arxiv.org/pdf/2504.03206 🌏 sites.google.com/cs.washington.…

Forget modeling every belief and goal! What if we represented people as following simple scripts instead (i.e "cross the crosswalk")? Our new paper shows AI which models others’ minds as Python code 💻 can quickly and accurately predict human behavior! shorturl.at/siUYI🧵

@SuJinyan6 Thank you for your interest in our work! Since this work was done during an internship, we're currently unable to release the internal code. However, we plan to reimplement the multi-turn training framework using an external open-source codebase and release that version publicly.

@yanming_wan Hi, Yanming, has the code been released?

Personalization methods for LLMs often rely on extensive user history. We introduce Curiosity-driven User-modeling Reward as Intrinsic Objective (CURIO) to encourage actively learning about the user within multi-turn dialogs. 📜 arxiv.org/abs/2504.03206 🌎 sites.google.com/cs.washington.…

@manuelsriosb18 Thank you for your interest in our work! Our experiments were conducted using internal TPU resources at Google, but in practice, any compute setup that can handle multiple Gemma 2B models should suffice. The implementation details can be found in the paper for reference.

@yanming_wan @yanming_wan What hardware do you need to run these experimentes ? Great work!

Overall, we propose CURIO for enhancing personalization in LLMs for multi-turn dialogs, which encourages LLMs to actively learn user traits and adapt its responses accordingly. This work was done with my awesome collaborators: @jiaxing_jxwu @marwaabdulhai @LiorShan @natashajaques

Baselines and entropy-based rewards lead to "controlling behavior", where the model gets high rewards by convincing the user to adopt a particular preference that is easier to cater to, rather than adhering to the ground-truth. "Grounded" rewards stop this reward hacking. (8/9)

We introduce MMMG: a Comprehensive and Reliable Evaluation Suite for Multitask Multimodal Generation ✅ Reliable: 94.3% agreement with human judgment ✅ Comprehensive: 4 modality combination × 49 tasks × 937 instructions 🔍Results and Takeaways: > GPT-Image-1 from @OpenAI leads image generation at 78.3% accuracy—13.7% ahead of the next-best model. The top open-source model, BAGEL from #ByteDance , achieves 45.5% accuracy. > Audio generation is still challenging: Top open-sourced models achieve only 48.7% accuracy in sound (Make-An-Audio 2 from #ByteDance) and 41.9% in music (MusicGen from @AIatMeta). 📜 Paper: arxiv.org/abs/2505.17613… 🛠️ Code and Evaluation Suite: github.com/yaojh18/MMMG 🥇Leaderboard: #leaderboard" target="_blank" rel="nofollow noopener">yaojh18.github.io/mmmg-leaderboa… 🧵1/N

Links Paper: arxiv.org/abs/2409.18073 Code: github.com/Simon-Wan/FISER Website: sites.google.com/view/fiser-hmt/ (10/10)

Overall, we present FISER for ambiguous instruction following by building a model that explicitly performs social reasoning to infer the human’s intentions from prior actions. This work was done with my awesome collaborators: @YueWu7677 @ypwang61 @maojiayuan @natashajaques (9/10)

We filter out a proportion of irrelevant objects to assess the impact of excessive item quantity on GPT-4. LLM relies on a very large proportion of objects being filtered out, showing that they can't effectively select relevant information and focus on relevant objects. (8/10)

We compare training models end-to-end with predicting goals as an auxiliary task, vs separating into two models that first predict goals, then the actions. Multi-staged is significantly better, implying that fully separating social from embodied reasoning performs better. (7/10)

In FISER, we explicitly model a human's intention by modeling the human’s overall plan as a set of predicates. We further assume that the human selects a subgoal that needs help, and specifies a robot’s task, which is the underlying intention of the instruction. (5/10)

We train FISER models from scratch using the following architecture. The first 2N layers form the social reasoning and the last N form the embodied reasoning. The embeddings at Layer 2N are used to recognize the robot’s task and the last layer is used to predict actions. (6/10)