@testerpce.bsky.social
How distillation is used today and what performance uplift it gives to open models (a rant) natolambert.substack.com/p/how-distil...
How distillation is used today and what performance uplift it gives to open models
This is in response to Ben Thompson’s recent piece where he said distillation is happening during RL and becoming more important to model performance.
natolambert.substack.com
This looks awesome! We need benchmarks that consider coordination between agents. Also super cool to see cross comparison between LLM/ VLM and MARL agents! Congrats to @kale-ab.bsky.social & co. 👏
Kale-ab Tessera@kale-ab.bsky.social · 3w agoCan LLM agents coordinate in long-horizon, open-ended worlds? We test 13 LLMs in Alem, a new benchmark. Most struggle, averaging ~6% normalised return. Yet on Hard, zero-shot Gemini 3.1 Pro matches the best MARL agent after 1B training steps. Our ablations show communication matters most. 🧵
Super happy to reveal our new paper! 🎉🙌♟️ We trained a model to play four games, and the performance in each increases by "external search" (MCTS using a learned world model) and "internal search" where the model outputs the whole plan on its own!
Nenad Tomasev@nenadtomasev.bsky.social · 2y agoI'm excited to share a new paper: "Mastering Board Games by External and Internal Planning with Language Models" storage.googleapis.com/deepmind-med... (also soon to be up on Arxiv, once it's been processed there)
Puneet Kant, Monika Tanwar A Synthetic Reliability-Aware PINN Benchmark for Offshore Wind Turbine Support-Structure Monitoring with Bayesian Inverse Identification https://arxiv.org/abs/2606.24176
Tianyu Ding, Juan Pablo De la Cruz Weinstein When Retrieval Metrics Mislead: Measuring Policy Signal in Long-Horizon Tool-Use Agents https://arxiv.org/abs/2606.23937
It's still really hard to tell which robotics models are "best," and there seems to be a good amount of competition at the top -- but at least I guess the willingness of some people to cheat on leaderboards is a good sign? open.substack.com/pub/itcanthi...
What Do Robotics Leaderboards Tell Us About The State of Robot Learning?
There remains no chatbot arena for robotics
open.substack.com
Hello all! 👋 I’m delighted to share a 🚨 new preprint 🚨: “Active Evaluation of General Agents: Problem Definition and Comparison of Baseline Algorithms”. A paper thread! 🤩📄🧵 1/N
Unlike board games, real-world strategic interactions are messy. Traditional game theory thus needs a boost for the age of agentic AI. Our #AAMAS2026 workshop "Strategic Engineering"(sites.google.com/view/se-aama...) in Cyprus aims to bridge the gap. Come join us to unlock truly strategic AI!
Here's an excerpt from the most recent video on how Shannon studied the entropy of English, animated by Mitchell Zemil
This week's #PaperILike is "Human-Guided Complexity-Controlled Abstractions" (Peng et al., NeurIPS 2023). Selecting the right levels and kinds of abstractions remains important and open for many forms of human-AI / human-robot interaction. PDF: arxiv.org/abs/2310.17550
Human-Guided Complexity-Controlled Abstractions
Neural networks often learn task-specific latent representations that fail to generalize to novel settings or tasks. Conversely, humans learn discrete representations (i.e., concepts or words) at a va...
arxiv.org
#MemoryModay #NLProc 'Visualizing Regional Language Variation Across Europe on Twitter' by Dirk Hovy et al. uncovers language differences across Europe with stunning visuals. Language is art! #Linguistics link.springer.com/referencewor...
Visualizing Regional Language Variation Across Europe on Twitter
Geotagged Twitter data allows us to investigate correlations of geographic language variation, both at an interlingual and intralingual level. Based on data-driven studies of such relationships, this…
link.springer.com
WE DON'T NEED PETROLEUM-BASED PLASTICS ANYMORE! "The [bamboo plastic] outperforms most commercial plastics and bioplastics in mechanical and thermo-mechanical metrics while maintaining full biodegradability in soil within 50 days and closed-loop recyclability with 90% retained strength."
High-strength, multi-mode processable bamboo molecular bioplastic enabled by solvent-shaping regulation - Nature Communications
Bioplastics derived from biomass show promise as sustainable alternatives to petrochemical plastics, but their adoption is hindered by their inferior mechanical properties and processability. Here, th...
nature.com
This points to an important direction: layering symbolic systems on top of LLMs. These can overcome the main shortcomings of LLM architectures: probabilistic execution, continual learning, attribution, and (maybe) uncertainty quantification. 1/
Antonin Raffin@araffin.bsky.social · 3mo ago"Codex grew programmatic policies with no neural nets: max score on Breakout, and SOTA-level scores on MuJoCo. Maybe heuristics were not too weak. Maybe they were just too expensive to maintain. Maybe it's the next paradigm." trinkle23897.github.io/learning-bey...
This paper is getting a lot of (deserved) attention; I think this paper serves as a nice complement arxiv.org/abs/2602.18710
Many AI Analysts, One Dataset: Navigating the Agentic Data Science Multiverse
Empirical conclusions depend not only on data but on analytic decisions made throughout the research process. Many-analyst studies have quantified this dependence: independent teams testing the same h...
arxiv.org
Per Engzell@pengzell.bsky.social · 3mo agoWhat can we learn from automating an entire quantitative social science paper, from prompt to finished product? Thread about ongoing work with @natewilmers.bsky.social 1/12 Paper: osf.io/preprints/so...
Super excited to get this at my door today! Ursula Laa and my book on exploring high-d data and models is in print! Book website is www.routledge.com/Interactivel... if interested I think my 30% discount code might work for you - msg me. Time to update dicook.github.io/mulgar_book/ too! #rstats
A Human-Centric Framework for Data Attribution in LLMs (FaCCT'26) TLDR: LLMs screwed up data economy, and NLP should help to re-design incentives with data attribution. Here's a moonshot for an alternative LLM use paradigm, for the users, creators & intermediaries. arxiv.org/abs/2602.10995 /1
Should all our resources go towards building chatbots? What if we built systems that actually give people meaningful agency? Finally out as an accepted @facct.bsky.social paper! Joint work with my past collaborators Sourojit, Pranav and Sanjana. huggingface.co/papers/2605....
Paper page - What if AI systems weren't chatbots?
Join the discussion on this paper page
huggingface.co
🚨 New preprint w/ Valerio Rubino and Peter Dayan: how do people discover and use compositional structure under constraints? osf.io/preprints/ps... A key factor is a simple heuristic that favors reuse of repeated and symmetric fragments across scales, is robust to time pressure, and sped up RTs 🧵👇
K-Dense BYOK Your own AI research assistant, running on your computer, powered by your API keys. K-Dense BYOK (Bring Your Own Keys) is an open-source app that lets you chat with an AI assistant called Kady. You ask Kady a question or give it a task, and it figures out the best way to handle it.
Meituan's LongCat-Next: A Discrete Native Autoregressive Multimodal Model (open-weight) 68.5B total params, 3B active, LongCat-Flash-Lite MoE backbone, excels at seeing, painting, and speaking in a unified discrete autoregressive framework.
What do people mean when they talk about world models for robots? Why are world models so exciting? New blog post talking about the ~3 kinds of world models, what they do and why they matter: itcanthink.substack.com/p/will-world...
Will World Models Allow Robots to Think?
A brief review of the technology powering a new generation of intelligent robots
itcanthink.substack.com
Can't wait for when I can vibe code a production recommender system. Until then, here's some system designs: • Retrieval vs. Ranking: eugeneyan.com/writing/syst... • Real-time retrieval: eugeneyan.com/writing/real... • Personalization: eugeneyan.com/writing/patt...
Dataset Distillation (2018/2020) They show that it is possible to compress 60,000 MNIST training images into just 10 synthetic distilled images (one per class) and achieve close to original performance with only a few gradient descent steps, given a fixed network initialization.
"Wild posteriors in the wild" A cool-looking paper if you're interested in funky posterior geometry Link: arxiv.org/abs/2503.00239 Code: github.com/YunyiShen/we... #stats
Anyone interested in interactive story generation? I genuinely love this stuff, and would love to talk about it with anyone who's interested This is a little tool I made to experiment with generating like murder mysteries automatically~
Xiaoyu Deng, Ye Zhang, Tianmin Guo, Yongzhe Zhang, Zhengjian Kang, Hang Yang ChallengeMe: An Adversarial Learning-enabled Text Summarization Framework https://arxiv.org/abs/2502.05084
Roman Vashurin (Mohamed bin Zayed University of Artificial Intelligence), ... CoCoA: A Generalized Approach to Uncertainty Quantification by Integrating Confidence and Consistency of LLM Outputs https://arxiv.org/abs/2502.04964
Haohao Zhu, Junyu Lu, Zeyuan Zeng, Zewen Bai, Xiaokun Zhang, Liang Yang, Hongfei Lin Commonality and Individuality! Integrating Humor Commonality with Speaker Individuality for Humor Recognition https://arxiv.org/abs/2502.04960