I will be presenting two posters at #EMNLP2025 on Nov 5 (Hall C):
โข 1โ2 PM โ FaviComp: Familiarity-Aware Evidence Compression for RAG
โข 2:30โ4 PM โ Code Execution as Grounded Supervision for LLM Reasoning
Come chat about RAG and grounded reasoning supervision for LLM!
#emnlp2025 Iโm presenting several papers I authored at Tencent AI Lab about agent self evolving and LLM memory. There will also be an oral presentation on the WebEvolver paper at Thu. Nov 6 at 16:30-18:00 Location: A110. Come and chat with me ๐
Sad I canโt make it to EMNLP in person this year, but Iโll be presenting our paper virtually! Catch me on Nov 6, 8 AM CST / Nov 5, 9 PM EST at GatherTown booth 1715 (https://t.co/P3N4a7WIEh)! Drop by and say hi!
Last year, I led an unlearning effort at Amazon with the Nova Responsible AI and Pretraining teams, focusing on controlling model knowledge and behavior. We hosted an LLM unlearning challenge at Semeval and developed LUME unlearning bechmark a multitask optimization method, and a selective unlearning approach. I'm excited to see our insights contribute to the recently released Amazon Nova Customization feature. We will be presenting two of these papers at EMNLP and I'd be happy to connect with anyone interested in this topic. See more details in the following post from Rahul. https://t.co/QPWobTLItR
๐ง ๐ Excited to introduce Supervised Reinforcement Learningโa framework that leverages expert trajectories to teach small LMs how to reason through hard problems without losing their minds. ๐คฏ
Better than SFT && RLVR.
Read more: https://t.co/taEL8Vk4X5
#llms#RL#reasoning
๐Drop from Gemini VibeCoding Team:
Why does code functionality correctness NOT necessarily translate to ๐ฏ๐ฒ๐๐๐ฒ๐ฟ ๐๐๐ฒ๐ฟ ๐ฝ๐ฟ๐ฒ๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ฐ๐ฒ ๐ณ๐ผ๐ฟ ๐ฐ๐ผ๐ฑ๐ถ๐ป๐ด tasks?
In ๐ฉ๐ถ๐ฏ๐ฒ๐๐ต๐ฒ๐ฐ๐ธ๐ฒ๐ฟ: https://t.co/SINlvRIumJ, we find that the human preference correlates the best with a composite score of functional correctness and instruction following capability.
To evaluate the instruction following capability of LLM for coding, we propose a taxonomy of ๐ฏ๐ฌ ๐๐ฒ๐ฟ๐ถ๐ณ๐ถ๐ฎ๐ฏ๐น๐ฒ ๐ฐ๐ผ๐ฑ๐ถ๐ป๐ด ๐ถ๐ป๐๐๐ฟ๐๐ฐ๐๐ถ๐ผ๐ป๐ together with their corresponding deterministic checkers.
We evaluated 31 leading LLMs and
show that even the strongest models struggle to comply with multiple instructions and exhibit clear functional regression!
Please take a look at our paper and let us know how you like it!
We also look forward to seeing the community extend the work on real user coding prompts to understand how we can improve ๐ฟ๐ฒ๐ฎ๐น-๐๐ผ๐ฟ๐น๐ฑ ๐๐๐ฒ๐ฟ ๐ฐ๐ผ๐ฑ๐ถ๐ป๐ด ๐ฒ๐ ๐ฝ๐ฒ๐ฟ๐ถ๐ฒ๐ป๐ฐ๐ฒ, let us know if you want to collaborate on these topics! (0/1)
[Belated update]
Excited to share Iโll be starting my PhD with @johnhewtt at Columbia!
Iโm deeply grateful to my advisors Prof. @robinomial and Prof. @muhao_chen, my mentor @fwang_nlp, Prof. @_jessethomason_, Prof. Hao Zhou at TsinghuaAIR, and all my peers @USC.
๐งต
In addition, we show a rubric-based reward can help steering modelsโ thinking/reflection process, potentially key to future directions on infusing constraints in reasoning.
๐ We are thrilled to release a new open-source Deep Research Agent, Cognitive Kernel-Pro, from Tencent AI Lab! We focus on building a fully open-source agent with (to the maximum extent) free tools, showcasing impressive performance on GAIA with Claude-3.7-sonnet and surpass the counterpart, SmolAgents by a large margin.
In addition, we study the training recipe for an open-source Deep Research Agent Foundation Model. We curate high-quality training data (queries, trajectories, and verifiable answers across web, file, code, and reasoning domains). Our finetuned Qwen3-8B (CK-Pro-8B) surpasses WebDancer and WebSailor with the similar model size on the text-only subset of GAIA.
๐ Paper: https://t.co/CHLs2U0JLy
๐ง Code: https://t.co/8dxtnjdsxn
๐ค Data & Model:
https://t.co/jzaV1LywnK
https://t.co/KD8k9WqGM9
This work builds on the previous efforts of Tencent AI Lab (Fig. 2). Be sure to check them out if you're interested!
1+1=3
2+2=5
3+3=?
Many language models (e.g., Llama 3 8B, Mistral v0.1 7B) will answer 7. But why?
We dig into the model internals, uncover a function induction mechanism, and find that itโs broadly reused when models encounter surprises during in-context learning. ๐งต
A little late but nevertheless happy to share that our paper โOutlier Gradient Analysis: Efficiently Identifying Detrimental Training Samples for Deep Learning Modelsโ was accepted as an oral presentation at #ICML2025 (top 1%). More details here: https://t.co/0E5wkLUA8X!
๐ Check out our paper: WebEvolver: Enhancing Web Agent Self-Improvement with Coevolving World Model, from Tencent AI Lab!.
We present a world model-driven framework for self-improving web agents, addressing critical challenges in self-trainingโsuch as limited exploration and performance plateaus.
๐ โโKey Innovations:โโ
- โโCo-Evolving World Modelโโ:
The world model is implemented as an LLM that predicts the next webpage state (observation) given the current state and a planned action.
In addition to fine-tuning the agent's policy model using self-generated trajectories, the same data is repurposed to train the world model.
- World Model as Web Server
Training Phase:โโ sample pseudo trajectories by replacing the real web server with the world model.
โโInference Phase:โโ simulate the outcome of candidate actions with 1-2 step look-ahead planning, to help better select actions.
๐ โโResults on Real-World Tasks (Mind2Web-Live & WebVoyager):โโ
โ โโ~10% higher success rateโโ vs. pure self-training.
โ significantlyโ fewer environment interactionsโโโefficient yet powerful!
arxiv: https://t.co/7tMVy4iI0J
code: https://t.co/wzutAx8zNR
๐จ New paper accepted to #ACL2025!
We propose SudoLM, a framework that lets LLMs learn access control over parametric knowledge.
Rather than blocking everyone from sensitive knowledge, SudoLM grants access to authorized users only.
Paper: https://t.co/gzzs9L6S1z
๐งต[1/6]๐
๐ฆ Speaker Spotlight: Sercan ร. Arฤฑk (@sercanarik)
We're thrilled to welcome Sercan, Senior Staff Research Scientist at Google Cloud AI, as an invited speaker at the ICML 2025 Workshop on Computer Use Agents!
Sercan's focuses on democratizing AI and applying it to impactful use cases across various industries. We're excited to hear his insights on making AI agents more interpretable, robust, and reliable.
#ICML2025 #ComputerUseAgents #LLMAgents #AI