Graphcore Research

Would you rather use 1 million × 16-bit weights, 4 million × 4-bit weights, or even 16 million × 1-bit weights? In joint work between Aleph Alpha Research and Graphcore, we asked this question of LLMs — the answer encouraged us to embrace the wonder ✨ of 1-bit weights, which can outperform 4-bit and 16-bit weights on a fixed weight memory budget. In our work - ⚖️ A scaling laws evaluation prompts us to consider very low-bit formats - 📈 Scaled-up tests show the power of memory-matched models with 1-bit weights - ⚡ Kernel benchmarking demonstrates their feasibility for autoregressive inference Read all about it in our blog and paper (link below! ⬇️) Massive thanks to our collaborators at Aleph Alpha Research! Authors: @SohirMaskey, Constantin Eichenberg, @atomicflndr and @douglasahorr

Link to blog ➡️ graphcore-research.github.io/2026-03-11-1-b… Link to paper ➡️ arxiv.org/abs/2602.15563

Would you rather use 1 million × 16-bit weights, 4 million × 4-bit weights, or even 16 million × 1-bit weights? In joint work between Aleph Alpha Research and Graphcore, we asked this question of LLMs — the answer encouraged us to embrace the wonder ✨ of 1-bit weights, which can outperform 4-bit and 16-bit weights on a fixed weight memory budget. In our work - ⚖️ A scaling laws evaluation prompts us to consider very low-bit formats - 📈 Scaled-up tests show the power of memory-matched models with 1-bit weights - ⚡ Kernel benchmarking demonstrates their feasibility for autoregressive inference Read all about it in our blog and paper (link below! ⬇️) Massive thanks to our collaborators at Aleph Alpha Research! Authors: @SohirMaskey, Constantin Eichenberg, @atomicflndr and @douglasahorr

✨ Check out our summaries and follow our blog ✨graphcore-research.github.io/papers-of-the-…

Our picks for November’s Papers of the Month are here. Out of our shortlisted papers, we spotlight three looking at LLM efficiency from different angles! 📊 First up, How to Scale Second-Order Optimization is studying optimal tuning of second order optimizers such as Muon. 🌫️ In Intelligence per Watt, the authors discuss our favorite metric on large language models: energy efficiency. And how to take advantage of edge AI inference. 🕸️ Finally, Int vs FP: A comprehensive study of fine-grained formats is contributing to an old-timer topic in quantization: integer vs floating (block) point formats. Check out our summaries 👇

Come to the Frontiers in Probabilistic Inference today at #NeurIPS2025 to see Graphcore researcher Michael Pearce presenting "Variational Entropy Search is Just 1D Regression"

Come to the ML for Systems workshop at #NeurIPS2025 to see Graphcore Researcher Callum McLean presenting "MXNorm: Reusing Block Scales for Efficient Tensor Normalisation"

At #NeurIPS in San Diego? Come swing by poster #1016 this afternoon to see Graphcore Researcher Johanna Vielhaben presenting "Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers"

✨Paper✨ arxiv.org/abs/2409.04103 ✨Blog post✨ graphcore-research.github.io/posts/the-role…

How does the structure of a Knowledge Graph influence model accuracy in #DrugDiscovery? Our comprehensive study with @AstraZeneca on the effects of graph topology on Knowledge Graph Completion models has just been published in Bioinformatics! Learn more in the paper and the blog post below! 👇

Apply here: job-boards.greenhouse.io/graphcore/jobs…

🚨 Graphcore is hiring AI Research Interns! 🚨 Join us to work at the intersection of hardware and AI and help shape the future of AI systems. Whether you're excited about efficient inference, large-scale training, or advancing frontier-model capabilities, we’ve got cutting-edge projects for you to dive into. Interested? Apply below 👇

✨ Check out our summaries and follow our blog ✨ graphcore-research.github.io/papers-of-the-…

Our picks for October’s Papers of the Month are here. Out of 49 shortlisted papers, we spotlight 4 that stand out for their clever ideas on making #LLMs faster, smarter, and more efficient! 📊 First up, Grouped Lattice Vector Quantisation introduces a novel technique for a fine-grained post-training quantisation of LLMs, retaining good performance even at low bit widths. 🌫️ In Planned Diffusion, @danielmisrael and colleagues combine autoregressive and diffusion models. While the autoregressive model creates a scaffold and plan, the diffusion model fills the gaps, achieving extremely low-latency text generation. 🤔 Is your LLM overthinking it? Rethinking Thinking addresses the problem of lengthy reasoning chains by bounding their thinking space and gradually distilling their thoughts, speeding up reasoning without losing depth. 🕸️ Finally, When Structure Doesn’t Help compares techniques for how LLMs read text attributed graphs. The results are rather surprising: sometimes, too much structure can hurt. Check out our summaries 👇

Learn more by signing up here! luma.com/0j3nptp7

LLM using too many reasoning tokens? 😕 Generation slow? 🐌 Or simply too many steps before EOS? 🪜🪜🪜 Douglas Orr (@douglasahorr), our beloved research scientist, has got you covered! He will tell you the remedies to all of the above in the shortest time possible. Registration link in the 🧵 below! (Special thanks to @CodeWordsAI and @join_ef)

✨ graphcore-research.github.io/papers-of-the-… ✨

September's Papers of the Month is here, and this month is all about LLMs! 🧠 Out of all papers released this month, our editor @robhu92 has curated: 📊 "FlowRL: Matching Reward Distributions for LLM Reasoning“ (review by @samot_gc): A clever usage of #GFlowNets to align an #RL policy model with the *full* reward distribution, encouraging mode coverage ☁️ "Soft Tokens, Hard Truths" (review by @lukehudlass): A simple and scalable method to integrate #continuousthoughts with RL training schemes 🏎️ "Set Block Decoding is a Language Model Inference Accelerator" (review by @douglasahorr): Exactly what it says on the 🥫! 🔁 "Turning Recurring LLM Reasoning into Concise Behaviors" (review by @DobrikG): An improvement to #CoT reasoning, trying to reduce self-repetition via metacognitive reduce Check out our reviews in the 🧵 below!

Finally, Graph-R1 is another addition to the stack of agentic RAG approaches, but this time, using knowledge hypergraphs! Summary: #graph-r1-towards-agentic-graphrag-framework-via-end-to-end-reinforcement-learning" target="_blank" rel="nofollow noopener">graphcore-research.github.io/papers-of-the-…

Next, Guiding Diffusion Models with RL for Stable Molecule Generation introduces reinforcement learning with physical feedback to accomplish exactly as its name suggests! Summary: #guiding-diffusion-models-with-reinforcement-learning-for-stable-molecule-generation" target="_blank" rel="nofollow noopener">graphcore-research.github.io/papers-of-the-…

Summer may be over, but Papers of the Month certainly isn’t! For August’s edition, we covered the following papers: ➡️ ADMIRE-BayesOpt ➡️ Guiding Diffusion Models with RL for Stable Molecule Generation ➡️ Graph-R1 🧵