🔥Telegram İfşa

Hope all conferences follow CVPR’s lead. 👏🏼

This paper is great. It shows why normal benchmarks cannot effectively convey the "goodness" of a certain model. I wondered if encouraging certain "words" may boost the performance of these reasoning chains. Language expressions are a proxy for the behaviors they investigate.

My hunch is that although the performance difference between base GPT-4 and 4.5 is (seemingly) small, it may actually lead to a quite significant difference with post-training.

first impressions on @apple intelligence:

if you see someone running a process on the GPUs for like 7+ days, they either know exactly what they're doing, or have no clue at all... (speaking from experience 😂)

multi-turn training got these models going wild.

Tree search is a natural choice for efficiently iterating and exploring diverse solutions for agentic tasks.

the "reject/accept" options on cursor are such an OP training signal... i'm jealous 🥹

📜🚨 Check out our latest work on "Self-Resource Allocation in Multi-Agent LLM Systems" where we explore how LLMs can be used to optimize task allocation in multi-agent systems 🤖 🧵(1/3)

If you're at ICLR you may want to grab the opportunity to talk with my incredible co-authors Kexun Zhang and Yuxi XIE on everything Search + Agent related at our SWE-Search poster, on Thursday at 3:00pm, Hall 3 + Hall 2B #156. 😁 iclr.cc/virtual/2025/p…

I'm presenting our mem v.s. gen paper at ICLR on Saturday morning at #244. Come and check it out if you're interested!

New paper 🚨📜🚀 Introducing “Agents of Change: Self-Evolving LLM Agents for Strategic Planning”! In this work, we show how LLM-powered agents can rewrite their own prompts & code to climb the learning curve in the board game Settlers of Catan 🎲 🧵👇

The thing people who dismiss LLMs don’t get is that even if they’re not the end game, they’ll be key to getting us there.

great paper thread.

😂😂...it's been so fun working with Kexun Zhang on SWE-Agents. thanks JetBrains. watch this space!

Great work on Searching in complex environments. The authors identified many of the problems we faced in SWE-Search: agent scaffolding, environment reliability, and selecting the correct final solution (in SWE-Search we addressed the latter using multi-agent debate verifier).

Elon's "search for the ultimate truth" paradigm for AI is kind of genius tbh. It's also very connected to RLVR training, where only the outcome matters, getting the "true" answer. :)

you know it's bad when it's thanking you for your patience.

interacting with AI is impacting humans in ways as fundamental as social media and the internet did.