Ruocheng Guo

Staff Research Scientist @Intuit
Ph.D. from DMML @SCAI_ASU
Tool-using Agents, Post-training
Causal ML
Ex-TikTok, CityUHK, MSR, Google X

[Ruocheng Guo @rguo_asu]

Easy to see the flaw of WMs in terms of causality, hard to figure out a solution that can scale up to the frontier WMs.

[Junfan Zhu 朱俊帆 ✈️ SIGGRAPH @junfanzhu98]

https://t.co/7BN6L59GeJ

[Ruocheng Guo @rguo_asu]

@jietang @elonmusk @teortaxesTex 期待使用glm作为主力coding agent的那一天

[Ruocheng Guo @rguo_asu]

@HuaxiuYaoML Too productive

[Ruocheng Guo @rguo_asu]

@cheryyun_l @NeurIPSConf Not invited either 😂

[Ruocheng Guo @rguo_asu]

@cwolferesearch Congrats!

[Ruocheng Guo @rguo_asu]

@ahatamiz1 I will be surprised if sota in automl cannot outperform autoresearch.

[Ruocheng Guo @rguo_asu]

@JustinLin610 Legend 🫡

[Ruocheng Guo @rguo_asu]

@_zifan_wang that's hard! Enjoy London, especially between March and Sep, at least no need to worry about visa too much anymore and 3yrs to PR via GTV, still need French/Spanish Schegen for fun :)

[Ruocheng Guo @rguo_asu]

@AlecStapp Anyone work in AI research gets administrative processing ~100% for any kind of US visa.

[Ruocheng Guo @rguo_asu]

Check out our work led by
@Dawei_Li_ASU

This is from the university collaboration program between Intuit AI Research and ASU.

Motivated by the fact that benchmarking tool PRMs relies on downstream policy learning or test time scaling, which adds another layer of complexity, we collected a comprehensive set of positive and negative actions of Tool-using Agents based on various Tool-using datasets to provide straightforward comparison between different tool PRMs.

We also show the effectiveness of training ToolPRMs with synthesized negative actions via SFT or RLVR!

[Dawei Li✈️ICLR2026 @Dawei_Li_ASU]

🚀 New Paper: ToolPRMBench — Evaluating Process Reward Models for Tool-Using Agents

🔗 Paper: https://t.co/PB0YU34947

Large language models are becoming powerful tool-using agents. But evaluating how they act, step by step, is still difficult. A single wrong tool call can break the entire process.

📊 What is ToolPRMBench?
ToolPRMBench is a new benchmark designed to evaluate process reward models (PRMs) in tool-using scenarios.
Instead of only checking final answers, it focuses on decision-level correctness—whether an agent chooses the right action at each step.

🧩 How is it built?
The benchmark transforms existing tool-use tasks into step-wise comparisons.
Each case includes the interaction history, the correct action, and a strong but incorrect alternative.
Both offline and online sampling are used, with multi-LLM verification to ensure data quality.

🔍 Key findings
Evaluations across 17 models show clear trends:
Tool-specific PRMs outperform general models
Larger models help, but specialization matters more
RL-based PRMs are more robust under distribution shift

🌟 Why it matters
ToolPRMBench provides a realistic and fine-grained way to evaluate tool-using agents. It also offers practical guidance for training better reward models.

[Dawei Li✈️ICLR2026 @Dawei_Li_ASU]

🚀 New Paper: ToolPRMBench — Evaluating Process Reward Models for Tool-Using Agents

🔗 Paper: https://t.co/PB0YU34947

Large language models are becoming powerful tool-using agents. But evaluating how they act, step by step, is still difficult. A single wrong tool call can break the entire process.

📊 What is ToolPRMBench?
ToolPRMBench is a new benchmark designed to evaluate process reward models (PRMs) in tool-using scenarios.
Instead of only checking final answers, it focuses on decision-level correctness—whether an agent chooses the right action at each step.

🧩 How is it built?
The benchmark transforms existing tool-use tasks into step-wise comparisons.
Each case includes the interaction history, the correct action, and a strong but incorrect alternative.
Both offline and online sampling are used, with multi-LLM verification to ensure data quality.

🔍 Key findings
Evaluations across 17 models show clear trends:
Tool-specific PRMs outperform general models
Larger models help, but specialization matters more
RL-based PRMs are more robust under distribution shift

🌟 Why it matters
ToolPRMBench provides a realistic and fine-grained way to evaluate tool-using agents. It also offers practical guidance for training better reward models.

[Ruocheng Guo @rguo_asu]

@huang_chao4969 It's actually pretty good and faster than gpt deep research in terms of literature review.

[Ruocheng Guo @rguo_asu]

On Dec 6, 10:30-11:30am, my colleague and I will present ParetoMIL: Early Risk Detection in Dialogue under Weak Supervision in Workshop on Multi-Turn Interactions in Large Language Models.

[Ruocheng Guo @rguo_asu]

I will be at neurips Dec 2-7, DM if you like to discuss tool using agents, RL for LLMs and causal ML.

On Dec 4, 11am-2pm, I will present our work Counterfactual Inference with Rank Preservation in the poster session.

https://t.co/m6xNiiZeTe

[Ruocheng Guo @rguo_asu]

That's very true, once one ran some evaluation and check, it is easy to catch either ground truth errors or metrics that are suboptimal.

[Lucas Beyer (bl16) @giffmana]

A lot of datasets are actually really bad! Even big conference ones, even ones that got awards!

It made me blanket lose trust.

It's simple to find out: Just spend 30min looking at it randomly. For vision, finetune a blind and a non-blind model and compare.

That's all it takes.

[Ruocheng Guo @rguo_asu]

@WenhuChen Feel Gemini3+cursor is better

[Ruocheng Guo @rguo_asu]

We are still calling for submissions! If your work is related to benchmarking/evaluation/reproducible causal ML methods (your estimator can be linear regression or LLMs), please consider to submit.

[Ruocheng Guo @rguo_asu]

@peter_richtarik Nowadays many reviewers are worse than gpt deep research.

[Ruocheng Guo @rguo_asu]

We are still calling for submissions! If your work is related to benchmarking/evaluation/reproducible causal ML methods (your estimator can be linear regression or LLMs), please consider to submit.

[Ruocheng Guo @rguo_asu]

CausalBench Workshop @ WSDM'26 is calling for papers.

If you have an early-stage work related to benchmarking/reproducing causal inference/causal ML methods and want to present and get feedback please consider to submit.

https://t.co/HDFHR79Fbx