Dr Philip Osborne

PhD in Artificial Intelligence from the University of Manchester, UK.

Founder of - Language Agents for Reinforcement Learning

[naklecha @naklecha]

today, i'm excited to release a reinforcement learning guide that carefully explains the intuition and implementation details behind every single fundamental algorithm in the field. enjoy :)

https://t.co/cWJTM2Lr1P

[Dr Philip Osborne @DataOsborne]

Really interesting work by Stephen Chung to improve RL safety by predicting future agent actions.

Title: Predicting Future Actions of Reinforcement Learning Agents

Paper: https://t.co/t08V01vnso

[Ling Yang @LingYang_PU]

🚀 Excited to share "Demystifying Reinforcement Learning in Agentic Reasoning" (https://t.co/y8ELYkdXwm)

Despite rapid progress in agentic RL, key questions remain: What data works best? Which algorithms are most effective? How should agents balance thinking vs. tool use?

We conducted a systematic investigation across 3 dimensions:
📊 Data curation - Real trajectories vs. synthetic data
⚙️ Algorithm design - Exploration-exploitation dynamics
🧩 Reasoning modes - Internal reasoning vs. tool calls

Our findings reveal simple yet powerful practices that dramatically improve training efficiency and final performance.

🏆 Result: DemyAgent-4B achieves SOTA on challenging benchmarks, even outperforming 32B models on AIME2024/2025, GPQA-Diamond & LiveCodeBench!

🎁 Open-sourced: ✅ 3K high-quality end-to-end SFT dataset ✅ 30K diverse RL dataset ✅ DemyAgent-4B model ✅ Complete training recipes & actionable insights

📄 Paper: https://t.co/y8ELYkdXwm
💻 Code: https://t.co/TMb8v28HjR
Detailed breakdown in thread below 🧵👇

#AgentRL #ReinforcementLearning #LLM #AgenticAI

[abranti @joaoabrantis]

I am writing a blog series exploring Kinship-Aligned Multi-Agent Reinforcement Learning.

The first post introduce Territories: a new environment where agents with divergent interests either learn to cooperate or see their lineage go extinct.

https://t.co/lIZGygSBtK

[@levelsio @levelsio]

Another bug submitted by a user, that I just copy pasted into Cursor and then immediately fixed

I think I need to create an automated workflow that pulls bugs reported from my @featurebase bug board, then sends them into Cursor/AI etc, to immediately fix it, then do a pull request that I can then check

Fully automated AI bug fixing workflow will be the standard I think

And then my daily work will just be checking the AI's pull requests and approve or reject

[Sakana AI @SakanaAILabs]

The AI Scientist Generates its First Peer-Reviewed Scientific Publication

We’re proud to announce that a paper produced by The AI Scientist-v2 passed the peer-review process at a workshop in ICLR, a top AI conference.

Read more about this experiment → https://t.co/50p2t9tgHC

To our knowledge, this is the first fully AI-generated paper that has passed the same peer-review process that human researchers go through. The paper was produced by an improved version of the original AI Scientist, called The AI Scientist-v2. We’ll be sharing the full details of v2 in an upcoming release.

We conducted this experiment with the full cooperation of both the ICLR leadership and the organizers of the ICLR workshop, @ICBINBWorkshop. We (@_yutaroyamada @cong_ml @shengranhu @RobertTLange) proudly collaborated with UBC (@jeffclune) and Oxford (@FLAIR_Ox) on this exciting project.

[Seth Karten @sethkarten]

Can a Large Language Model (LLM) with zero Pokémon-specific training achieve expert-level performance in competitive Pokémon battles?
Introducing PokéChamp, our minimax LLM agent that reaches top 30%-10% human-level Elo on Pokémon Showdown!
New paper on arXiv and code on github!

[Dr Philip Osborne @DataOsborne]

@sethkarten Amazing work! I saw the feedback on OpenReview and disagree with some of the feedback on lack of generalisability, its an issue with all RL work but at least using LLMs is a positive step on improving that, keep up the good work and hopefully with adjustments it can be accepted!

[Peter Jansen (@ACL2026) @peterjansen_ai]

Can language models perform end-to-end scientific discovery? In our NeurIPS Spotlight paper, we show: very rarely.

Our best model found <20% of discoveries, our best PhDs found nearly all.

Paper: https://t.co/WyidZ9CTcN
Code/Web: https://t.co/sJWrlRENy8
@allen_ai @MSFTResearch

[Peter Jansen (@ACL2026) @peterjansen_ai]

Great to see that tables from our DiscoveryWorld paper are now the go-to examples of hard-to-automatically-parse tables 😅

(But the colors and slanted headings really do make them nice for humans to read)

https://t.co/WyidZ9CTcN

[Dr Philip Osborne @DataOsborne]

@FrancescoD_Ales Excel in most workplaces I've seen

[Dr Philip Osborne @DataOsborne]

A pleasure to be presenting at #eacl2023 this week

[Dr Philip Osborne @DataOsborne]

https://t.co/fitvmiys2R

[Francesco D'Alessio @FrancescoD_Ales]

🗳 Time to vote for you #ProductivityAwards2020

Now with...
📺 Top Productivity YouTubers 2020
🧠 Best Feature of 2020
🆕 Best New YouTubers in Productivity 2020

🙌🏼 VOTE!
WINNERS ANNOUNCED 10/12
🎺 https://t.co/1s0CBuj9CL

[Dr Philip Osborne @DataOsborne]

@ClickUp_app Hi, is this only for full-time and on-site availablity? Am a doctoral student in AI, always interested in new opportunities and a big fan of ClickUp. I would love to apply my work towards helping improve the service if this can be discussed

[Dr Philip Osborne @DataOsborne]

For anyone looking at these new Email apps but cant justify the cost, @missiveapp's free version limits you to a 15 day window which you can turn to an inbox for only current items with all the other features. Productivity by accidental decluttering! @KeepProductive @AliAbdaal

[Dr Philip Osborne @DataOsborne]

Following up on your "My Workflow of Collecting Information", for those on a tight budget, you can export highlights via APIs for completely free. Took me about an hour with no API exp to setup and in less than 50 lines of code. Example output in image. @ShuOmi3

[Dominik Haitz @d_haitz]

I just published in @TDataScience Cyberpunk Style with Matplotlib https://t.co/ACIyJSmmBe

[Dr Philip Osborne @DataOsborne]

@ClickUp_app @NotionHQ @iAyush_Sharma I use both myself as a researcher but I've found Clickup to be vastly more powerful for time tracking/management and use Notion for notes and everything else. Notion can in theory do everything but requires a lot more manual input to do it.

[Microsoft Research @MSFTResearch]

Multi-armed bandits are a simple but very powerful framework for algorithms that make decisions over time under uncertainty. “Introduction to Multi-Armed Bandits” by Alex Slivkins provides an accessible, textbook-like treatment of the subject: https://t.co/RzkaVZTLMS