Katherine Tieu @kthrn_uiuc

Today on Connectionism: establishing the conditions under which LoRA matches full fine-tuning performance, with new experimental results and a grounding in information theory

Thinking Machines @thinkymachines

2 days ago

Today on Connectionism: establishing the conditions under which LoRA matches full fine-tuning performance, with new experimental results and a grounding in information theory

77 524 3K 1.0M 2K

Download Image

Multi-Agent Debate (MAD) has been hyped as a collaborative reasoning paradigm — but let me drop the bomb: majority voting, without any debate, often performs on par with MAD. This is what we formally prove in our #NeurIPS2025 Spotlight paper: “Debate or Vote: Which Yields…

With the help from Bytedance @verl_project team, we have integrated collabllm as a recipe to veRL, a to-be-most-popular open RL library for LLMs. Now you are only one step away from making your LLM a great collaborator in multiturn conversations. verl.readthedocs.io/en/latest/algo…

Shirley Wu @ShirleyYXWu

3 months ago

With the help from Bytedance @verl_project team, we have integrated collabllm as a recipe to veRL, a to-be-most-popular open RL library for LLMs. Now you are only one step away from making your LLM a great collaborator in multiturn conversations. verl.readthedocs.io/en/latest/algo… https://t.co/W0TplOcynS

18 32 210 49K 44

Efficient training of neural networks is difficult. Our second Connectionism post introduces Modular Manifolds, a theoretical step toward more stable and performant training by co-designing neural net optimizers with manifold constraints on weight matrices.…

We keep pushing the limits of speculative decoding (SD) in LLM inference -- check out our latest NeurIPS’25 paper: Lookahead Reasoning (LR). The high-level rationale is pretty simple: SD alone isn’t enough now: as GPUs get stronger (H200 -> B200 -> Rubin CPX), we'll be able to…

Hao AI Lab @haoailab

a week ago

We keep pushing the limits of speculative decoding (SD) in LLM inference -- check out our latest NeurIPS’25 paper: Lookahead Reasoning (LR). The high-level rationale is pretty simple: SD alone isn’t enough now: as GPUs get stronger (H200 -> B200 -> Rubin CPX), we'll be able to…

3 24 137 41K 76

Download Video

Very cool work from Meta Superintelligence Lab. They are open-sourcing Meta Agents Research Environments (ARE), the platform they use to create and scale agent environments. Great resource to stress-test agents in environments closer to real apps. Read on for more:

✨ Accepted as a Spotlight at #NeurIPS2025 !! 🥁 TLDR: AB-MCTS resolves the “go wider” vs. “go deeper” trade-off in LLM inference-time scaling using a principled Bayesian method, achieving strong multi-turn performance across LiveCodeBench, ARC-AGI & MLE-Bench. See you in SD!

Takuya Akiba @iwiwi

3 months ago

✨ Accepted as a Spotlight at #NeurIPS2025 !! 🥁 TLDR: AB-MCTS resolves the “go wider” vs. “go deeper” trade-off in LLM inference-time scaling using a principled Bayesian method, achieving strong multi-turn performance across LiveCodeBench, ARC-AGI & MLE-Bench. See you in SD!

10 72 377 84K 253

Download Image

Accepted as NeurIPS2025 Spotlight! Existing large multimodal models (LMMs) have very poor visual understanding and reasoning over part-level attributes and affordances (only 5.9% gIoU). We developed novel part-centric LMMs to address these challenges arxiv.org/pdf/2505.20759

🧠 For many AI workflows, the Memory is becoming the main bottlenech rather than compute. And this new XQuant method claims - 10–12.5x memory savings vs. FP16 - Near-zero accuracy loss - Beats state-of-the-art KV quantization🔥 The slow, expensive part in modern AI is moving,…

The paper shows LLMs can play repeated prisoner’s dilemma, stay cooperative, and adapt fast. Each round is a choice, cooperate or defect. Both gain if both cooperate, defection pays only when the other cooperates. The study plays LLM agents against 240 classical strategies…

This is a new 100-page RL for LLM literature review. It appears fairly complete. It also covers static/dynamic data and frameworks. And it has some nice figures! 🔗arxiv.org/abs/2509.08827

keyboard_arrow_left Previous keyboard_arrow_right Next

I'm hiring 1-2 new postdocs to work on AI for Scientific Discovery (especially on drug discovery and material discovery) and Science-Inspired AI (especially on scientific foundation models). Please drop me an email if you are interested or know someone who might be a great fit!

Besides the fun fact that Connectionism is connected with the early days of the AI field and highlights similarities between neural networks and human brains, the flagship product of the (first) Thinking Machines is named Connection Machine. — 🧑‍🎓Enjoy reading and more is coming!

Thinking Machines @thinkymachines

3 weeks ago

Besides the fun fact that Connectionism is connected with the early days of the AI field and highlights similarities between neural networks and human brains, the flagship product of the (first) Thinking Machines is named Connection Machine. — 🧑‍🎓Enjoy reading and more is coming!

241 1K 8K 3.2M 5K

Download Image

I'm hiring a postdoc on #AI for #science starting soon. We are building AI-enabled 3D human liver printing technology with many experts across disciplines! If you are excited about this, drop me an email with your CV and a brief introduction! Repost is much appreciated!

Using small language models in a multi-agent system is not always good. Here is one scenario (multi-agent debate) where weaker LLM agents often disrupt the performance of the stronger agents. This focuses on LLM agents engaged in debate, but it might be a more pervasive issue.

OpenAI realesed new paper. "Why language models hallucinate" Simple ans - LLMs hallucinate because training and evaluation reward guessing instead of admitting uncertainty. The paper puts this on a statistical footing with simple, test-like incentives that reward confident…

keyboard_arrow_left Previous keyboard_arrow_right Next

A comprehensive survey on trustworthiness in reasoning with LLMs. Great read for AI devs.

Who is inducing failure in LLM Agentic Systems? This is a cool idea to diagnose errors in multi-agent interactions. AgenTracer-8B outperforms giant proprietary LLMs like Gemini-2.5-Pro and Claude-4-Sonnet by up to 18.18%.

🌠Flow Matching Meets Biology and Life Science: A Survey Flow matching is emerging as a powerful generative paradigm. We comprehensively review its foundations and applications across biology & life science🧬 📚Paper: arxiv.org/abs/2507.17731 💻Resource: github.com/Violet24K/Awes…

🚀 Excited to announce our new Agentic AI MOOC | Fall 2025! Following our successful MOOCs in the Fall’24 and Spring’25 w. 25k+ registered learners, we're diving deeper into #AgenticAI—from LLM foundations and agentic frameworks to real-world applications in code generation, web…