Instead of retraining / adapting reasoning models for every domain, we can plug in reward modules to steer reasoning toward higher reliability, which is needed especially in high-stakes settings.

14/14

Bigger picture:
🧠 reasoning models = general-purpose
📚 process reward agents = domain-specific grounding modules
→ PRA enables a decoupling of how we reason from what we know

13/

Results / Take-aways (more details in paper):
✅ Strong improvements on medical reasoning benchmarks
✅ Works across multiple frozen policy models
✅ Generalizes beyond the model it was paired with

12/

Why this matters:
In knowledge-intensive domains, correctness is not merely logical consistency
→ it requires alignment with external knowledge.

PRA proposes a way to bring this into the reasoning & reward loop
11/

This unlocks something important:

👉 Tree-search reasoning paired with grounded knowledge

Instead of committing to one chain of thought, PRA can:
→ explore multiple paths
→ retrieve different evidence per step
→ evaluate them in real time

10/

Key idea:
PRA turns PRMs into active agents that can:
• Check each reasoning step dynamically
• Query external knowledge (guidelines, textbooks, etc.)
• Provide immediate rewards during generation

9/

Here, we introduce Process Reward Agents (PRA).

A framework that uses PRMs online to search, reward, and guide reasoning as it unfolds.

8/

Grounded (i.e. retrieval-augmented) PRMs typically operate post-hoc:
They score reasoning only after the full trace is generated.

This means:
❌ no real-time feedback
❌ no flexible search (e.g. tree exploration)
❌ limited ability to steer reasoning as it happens
7/

Recent work combines PRMs with retrieval (e.g. Med-PRM):
- pull in external knowledge
- critique reasoning traces step-by-step using external sources

But there’s a catch 👇
6/

One promising direction to go beyond final answers: Process Reward Models (PRMs)

Instead of only judging final answers, PRMs evaluate intermediate reasoning steps.

5/

This matters a lot.
In domains like medicine, LLM reasoning is not only about the final answer/decision
-
we urgently need sound and defensible justifications along the way!

4/

But in knowledge-intensive domains:

- step correctness often depends on external knowledge (consensus, guidelines, textbooks, local constraints etc.) spread across various sources

- individual steps may not be easily verifiable in isolation

3/

In math/code, intermediate steps are often locally verifiable
→ you can +- easily verify if a step is correct (e.g. formal rules, symbolic solvers, code compilation & execution etc.)

2/

Process Reward Agents for Steering Knowledge-Intensive Reasoning Reasoning in knowledge-intensive domains remains challenging as intermediate steps are often not locally verifiable: unlike math or code, evaluating step correctness may require synthesizing clues acr...

Preprint: arxiv.org/abs/2604.09482
Page: process-reward-agents.github.io
Code: github.com/eth-medical-...

Big thanks to a stellar team of co-authors: Jiwoong Sohn,
Tomasz Sternal, Kenneth Styppa, and Torsten Hoefler!
@ethz.ch

1/

[Preprint Alert] 𝐏𝐑𝐎𝐂𝐄𝐒𝐒 𝐑𝐄𝐖𝐀𝐑𝐃 𝐀𝐆𝐄𝐍𝐓𝐒 (𝐏𝐑𝐀)

Why is it relatively easy to get LLMs to produce strong reasoning traces in math/code…
but much harder in application domains like health? And what can we do against it?

Check out our new paper & 🧵below:

Check out Med-PRM, an approach for LLMs to verify each reasoning step against guidelines:

🔗 Page: med-prm.github.io
📄 Paper: arxiv.org/abs/2506.11474
🧠 Model: huggingface.co/dmis-lab/lla...
📚 Dataset: huggingface.co/datasets/dmi...
💻 Code: github.com/eth-medical-...
🧵 Thread: tinyurl.com/yu933dx6

Welcome to our new lab page 🚀
bsse.ethz.ch/mail

Great to see this out!

#AI agent labs are becoming better at producing autonomous research. Still, they operate in isolation w/o improving & interacting.

Here, we introduce 𝐀𝐠𝐞𝐧𝐭𝐑𝐱𝐢𝐯, where agent laboratories can upload & download latest research - which accelerates their progress:

Great effort led by Samuel Schmidgall!

Today, our two new faculty members held their inaugural lectures @ethzurich.bsky.social. Basile Wicky | Biomedical Design Lab, presented on designing proteins that interface with life; Michael Moor @michaelmoor.bsky.social | Medical AI Lab, spoke about AI in medicine. Recordings > u.ethz.ch/hQdQl

Today @michaelmoor.bsky.social took a train from Zurich to Innsbruck to kick-off our new Faculty of "AI and Scientific Computing"! 🎉 He talked about LLMs and Medical AI Agents. Exciting science and great discussions! Thanks! More info about our faculty: aiscm.i-med.ac.at #ai #scientificcomputing

Let's get paper sharing started here! I'll start:

Interesting new preprint on Multimodal medical preference optimization:
arxiv.org/pdf/2412.06141
@huaxiuyaoml.bsky.social (and others)

How can we build an Al virtual cell that simulates all functions and interactions of a cell? How will it transform research and drive breakthroughs in programmable biology, drug discovery and personalized medicine?

Take a look at our paper in @cellpress.bsky.social!
www.cell.com/cell/fulltex...

UI / UX suggestion for #bluesky:

I would remove the ".bsky.social" string that clutters the app. Like when looking at a list of n accounts, one has to visually ignore this suffix n times.

I suspect that small UI things like this could make a big impact in getting more momentum.

OpenAI coming to Switzerland! Congrats on the new roles!

Hello, world! 🤩

There is a new blue animal in town, it can fly but is not a bird #Xodus

Any #NewPI out there who just joined? Happy to connect! 🚀 🦋

Corrected link: bsky.app/profile/mich...

Finally figured out how to create a starter pack yay 😅

go.bsky.app/SNnu3ev

Just added a bunch of folks I could quickly find, far from exhaustive..