@FrankYueWu1 @stillpedant @MengdiWang10 @yisongyue @NECLabsAmerica @Caltech @Princeton @rpi 🧵 Thanks for reading — feel free to reach out or share! 🙌 #neurips2025 #LLMs #AIResearch #MachineLearning #Reasoning #InferenceScaling #DISC #DeepSeek #LLMReasoning

🧵10/n Huge thanks to my amazing collaborators — Wei Cheng, Benjamin Riviere, @FrankYueWu1 , @stillpedant, @MengdiWang10, @yisongyue , Santiago Paternain, and Haifeng Chen Truly a pleasure collaborating across @NECLabsAmerica , @Caltech , @Princeton, and @rpi . 🙏

🧵9/n Inference scaling doesn’t need to be brute force. It can be intelligent. By letting models adapt their reasoning granularity, DISC opens a new path toward efficient, self-reflective inference. Paper: https://t.co/rR8NhdRgJl Website:

🧵8/n DISC is plug-and-play: ✅ Works with greedy, beam, or MCTS search ✅ Boosts open-source models like Llama, Mistral, and Qwen ✅ Keeps runtime overhead negligible One algorithm, many models — no retraining needed. 🧠

🧵7/n Reasoning models like DeepSeek-R1 love DISC. ❤️‍🔥 With only 10 samples, it boosts performance by 85% — and even when capped to the same token budget as one vanilla generation, DISC still wins by +33%. Smarter reasoning, not longer decoding. ⚡️

🧵6/n Up to 4× higher accuracy over base model with open source models, using just 10 samples. 🔥

🧵5/n So, how well does it work? 📈 Across benchmarks: - APPS: -5.0% error - MATH: -6.7% error - LiveCodeBench: -10.5% error Better scaling with compute budget across the board Beats baselines at any budget

🧵4/n At its core, DISC teaches models where to think harder. It dynamically allocates compute — 🧩 subdividing challenging steps ⚙️ focusing search on high-reward prefixes 🚀 avoiding wasted effort on trivial tokens All with no extra training, heuristics, or supervision.

🧵3/n Instead of predefining step sizes, DISC learns to adaptively break down reasoning in real time — taking bigger leaps on easy parts and finer steps where the model struggles. It’s like giving your LLM a dynamic zoom lens for reasoning. 🔍

🧵2/n Static inference methods split reasoning steps into fixed steps (token, line, or sentence)… but that’s like using the same stride length for every terrain. 🥾 How can we make inference both scalable and adaptive? That’s where DISC comes in. 💡

🎉 Thrilled to share that our paper “DISC: Dynamic Decomposition Improves LLM Inference Scaling” has been accepted to #neurips2025 ! 🚀 Here’s how we push reasoning and inference scaling to new heights 🧵 1/n

DISC: Dynamic Decomposition Improves LLM Inference Scaling DISC adaptively partitions reasoning traces during inference so that language models devote more compute to the hardest reasoning steps — leading to faster, more accurate inference across coding and math benchmarks.

And thanks for the great talk on hallucination and HLE from @ofirnachum, Edwin Zhang, and Long Pham!

Love seeing research and industry collide in real time at @agihouse_org paper reading roundtables

📅 TOMORROW is LM4Sci #COLM2025! ⚡🔬 Are you excited? Today's spotlight: Jennifer Sun (Cornell, Google DeepMind) @JenJSun on Accelerating Knowledge & Discovery in Scientific Workflows 🧵

The Interlink at YC AI SUS was super fun — it matched you with others who shared your interests automatically. Would love to see academic AI/ML conferences like #NeurIPS, #ICLR, and #ICML adopt something similar so that it is easier to find people now that conferences have

Andrej Karpathy also showed how easy it’s becoming to “vibe code” complex apps — like this AI-powered menu generator he built in a night: https://t.co/DpFMnQkp4K Excited for a future where building powerful tools is more accessible than ever for everyone!

Y Combinator AI SUS last week was a blast. Got to hear Andrej Karpathy talk about the future of AI-driven software engineering — he's phenomenal at making complex ideas feel intuitive.

We don’t hardcode pain into games—so why does losing feel so real? And what does that teach us about AI and reward modeling? https://t.co/0NXzcpFJ8B #AI #ReinforcementLearning #MachineLearning