Lai Dang Quoc Vinh @LDQuocVinh

Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference” We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to…

Cool research from Microsoft! They release rStar2-Agent, a 14B math reasoning models trained with agentic RL. It reaches frontier-level math reasoning in just 510 RL training steps. Here are my notes:

Towards a Unified View of LLM Post-Training This work proposes Hybrid Post-Training, which switches between RL and SFT using simple performance feedback to balance exploration and exploitation. More below:

RL’s Razor: On-policy RL forgets less than SFT. Even at matched accuracy, RL shows less catastrophic forgetting Key factor: RL’s on-policy updates bias toward KL-minimal solutions Theory + LLM & toy experiments confirm RL stays closer to base model

@jasonth0 I love doing this actually :). I think it's a pretty powerful eval too. Have all models generate something, then put it all together and give it back to all of them and ask them to rank all outputs. I thought models might have a bias to prefer their own outputs, but this doesn't…

DuPO: Enabling Reliable LLM Self-Verification via Dual Preference Optimization "We present DuPO, a dual learning-based preference optimization framework that generates annotation-free feedback via a generalized duality" "DuPO decomposes a primal task’s input into known and…

Nice paper that is basically an empirical survey of all the recent RL algorithm improvements.

Couldn't resist. Here's a pure PyTorch from-scratch re-implementation of Gemma 3 270M in a Jupyter Notebook (uses about 1.49 GB RAM): github.com/rasbt/LLMs-fro…

Sebastian Raschka @rasbt

2 months ago

Couldn't resist. Here's a pure PyTorch from-scratch re-implementation of Gemma 3 270M in a Jupyter Notebook (uses about 1.49 GB RAM): github.com/rasbt/LLMs-fro… https://t.co/9vF9U29pWh

26 202 1K 469K 786

Download Image

Introducing Jan-v1: 4B model for web search, an open-source alternative to Perplexity Pro. In our evals, Jan v1 delivers 91% SimpleQA accuracy, slightly outperforming Perplexity Pro while running fully locally. Use cases: - Web search - Deep Research Built on the new version…

Presenting the GLM-4.5 technical report!👇 arxiv.org/abs/2508.06471 This work demonstrates how we developed models that excel at reasoning, coding, and agentic tasks through a unique, multi-stage training paradigm. Key innovations include expert model iteration with…

Top AI Papers of The Week (August 4-10): - CoAct-1 - ReaGAN - Agentic Web - Seed Diffusion - Efficient Agents - A Taxonomy of Hallucinations - Unified Retrieval Agent for AI Search Read on for more:

I'm thrilled to announce the definitive course on Claude Code, created with @AnthropicAI and taught by Elie Schoppik @eschoppik. If you want to use highly agentic coding - where AI works autonomously for many minutes or longer, not just completing code snippets - this is it.…

China just dropped an absolute bombshell AI for Math paper: not just Gold in IMO 2025, but >50% of all Putnam and 78% of all past IMO problems. It BEATS Google's AlphaGeometry2, achieves 100% on OpenAI's miniF2F. Uses Lean for proofs and novel approaches for geometry. And it's…

9 new policy optimization techniques ▪️ GSPO ▪️ LAPO ▪️ HBPO ▪️ SOPHIA ▪️ RePO ▪️ CISPO ▪️ PAPO ▪️ OPO ▪️ EXPO Save the list, and check this out for the links and more info: huggingface.co/posts/Kseniase…

Gemini 2.5 Pro Capable of Winning Gold at IMO 2025 This work shows that you can really push LLMs to achieve remarkable performance on hard tasks. Self-verification and careful orchestration did the trick here! Good insights for devs on this one. Here are my notes:

Introducing GLM-4.5 and GLM-4.5 Air: new flagship models designed to unify frontier reasoning, coding, and agentic capabilities. GLM-4.5: 355B total / 32B active parameters GLM-4.5-Air: 106B total / 12B active parameters API Pricing (per 1M tokens): GLM-4.5: $0.6 Input / $2.2…

Finally, a good modern book on causality for ML: causalai-book.net by @eliasbareinboim. This looks like a worthy successor to the ground breaking book by @yudapearl which I read in grad school. (h/t @JoshuaSafyan for the ref).

Beautiful @GoogleResearch paper. LLMs can learn in context from examples in the prompt, can pick up new patterns while answering, yet their stored weights never change. That behavior looks impossible if learning always means gradient descent. The mechanisms through which this…

New Anthropic research: Building and evaluating alignment auditing agents. We developed three AI agents to autonomously complete alignment auditing tasks. In testing, our agents successfully uncovered hidden goals, built safety evaluations, and surfaced concerning behaviors.

>>> Qwen3-Coder is here! ✅ We’re releasing Qwen3-Coder-480B-A35B-Instruct, our most powerful open agentic code model to date. This 480B-parameter Mixture-of-Experts model (35B active) natively supports 256K context and scales to 1M context with extrapolation. It achieves…