Rafael Pardinas

RL & self-evolving agents at ServiceNow AI Research

[Rafael Pardinas @muchomuchacho]

So exciting to have our work featured among top class players!

[SemiAnalysis @SemiAnalysis_]

RL Systems Mind the Gap:
Matching Trainer and Generator Throughput
RL Training Infrastructure, GRPO,
PipelineRL, Async RL, Policy Staleness,
RL Sandbox Infra, CPU Requirements,
TCO Analysis, Thinking Machines Tinker

https://t.co/yr5oH99h4B

[Alex Gurung @AlexAag1234]

Been thinking recently about how to improve credit assignment in long horizon RL?

Our new MosaicLeaks blog post describes our method to accurately value actions via situational rewards, improving our privacy-aware research agent over outcome-only rewards!
https://t.co/xCbOKEcdwa

[Rafael Pardinas @muchomuchacho]

@eisokant @eliebakouch Base models are amazing for RL tuning and reasoning. If a company is looking to shape some behaviour into their models I would expect them to start with a base stage.

[Rafael Pardinas @muchomuchacho]

@johnschulman2 PipelineRL: rejected from NeurIPS 2025

[Rafael Pardinas @muchomuchacho]

@cwolferesearch Fantastic compilation. Thanks!

[Alex Gurung @AlexAag1234]

Excited to share my recent work @ServiceNowRSRCH ! We introduce a new privacy-centric deep research dataset and show models frequently leak enterprise information.
However, training with dense _situational_ rewards efficiently learns to jointly optimize performance and privacy

[Rafael Pardinas @muchomuchacho]

Led by @AlexAag1234 at ServiceNow AI Research, with @gspandana , @alexandredrouin , @ILaradji , @PerouzT and me.

Paper: https://t.co/2lGKyiJzW2

[Rafael Pardinas @muchomuchacho]

MosaicLeaks is now on arXiv. The Mosaic Effect captures a simple idea: small fragments can look harmless alone, but become revealing in aggregate. Deep research agents can leak enterprise information in exactly this way.

1/9

[Rafael Pardinas @muchomuchacho]

The core idea: Enterprise agent privacy failures will not only come from copying private text. They can also come from the external actions agents take while trying to be useful. Privacy shouldn't come at the cost of utility, we can optimise for both.

8/9

[Rafael Pardinas @muchomuchacho]

this is too good

[テコまる @tecomalupepepe]

長椅子振動主への反射システム

[Rafael Pardinas @muchomuchacho]

This is becoming really powerful. More to come for high latency agentic pipelines

[Rafael Pardinas @muchomuchacho]

Better reasoning does not have to mean longer reasoning.

Apriel OpenReasoner: fully reproducible multi-domain RL post-training using public datasets. 30-50% shorter traces, no quality trade-off.

@ServiceNowRSRCH @ehsk0 @dvazquezcv @alexandredrouin

[Rafael Pardinas @muchomuchacho]

@martinmbauer That may sit at a different point in time at different places

[Rafael Pardinas @muchomuchacho]

@nicogontier @ServiceNowRSRCH It has been a pleasure Nico!

[Rafael Pardinas @muchomuchacho]

@NandoDF Those two sentences are equivalent at some level, you may not understand it but that's probably because you didn't create it.