Ji Xue

We (w/ Shansan Gong, Ruixiang ZHANG, Huangjie Zheng, Jiatao Gu, Navdeep Jaitly, Lingpeng Kong) released a family of 7B diffusion language models, DiffuCoder, that specializes on code generation, with a focus on understanding and improving masked diffusion models. A core analysis of DiffuCoder is the autoregressiveness (AR-ness) score, a novel metric that quantifies the causal patterns in decoding, revealing how diffusion models break from strict left-to-right generation for more flexible, non-linear code planning. Recent advances in autoregressive (AR) models dominate code generation, but diffusion-based LLMs (dLLMs) like DiffuCoder offer a promising alternative, especially for complex programming tasks. DiffuCoder explores how these models decode differently—showing less global AR-ness in code tasks compared to math—and how temperature affects both token selection and generation order, unlike traditional AR models. We also introduce coupled-GRPO, a post-training RL method with a coupled-sampling scheme, to reduce performance drops during accelerated decoding, boosting parallelism and efficiency. We use a self-improvement pipeline that leverages AR-ness analysis, coupled-GRPO optimization, and evaluation on benchmarks like AceCode-89k to refine decoding strategies. This approach enables DiffuCoder to navigate diverse code generation pathways and enhance performance with modest computational overhead. Looking ahead, we aim to further leverage Reinforcement Learning to steer code generation through these decoding patterns, with the discrete nature of AR-ness scores providing a foundation for search-based strategies—ideal for the sparse rewards of optimizing complex code structures. Check out our full paper and code for a deeper dive! Paper: https://lnkd.in/gVWU3BDJ Code: https://lnkd.in/gmXTZ_6n Models: https://lnkd.in/gTcKCDr9 #MachineLearning #AI #CodeGeneration #DiffusionModels #NLP

220

5 Comments

🚀 Smarter, Faster RAG: A Policy-Optimized, Memory-Efficient Framework for Test-Time Reasoning 📄 New arXiv print 🔗 Scaling Test-Time Inference with Policy-Optimized, Dynamic Retrieval-Augmented Generation via KV Caching and Decoding https://lnkd.in/dErdTDf7 🔍 The Problem Retrieval-Augmented Generation (RAG) boosts LLMs with external knowledge but suffers from: ❌ Static retrieval schedules ❌ Distractor-induced hallucinations ❌ Linear memory growth in KV caching ❌ Inefficient test-time compute without reward alignment. 🧠 Our Framework: PORAG + ATLAS + CRITIC We propose a three-component architecture that dynamically optimizes retrieval, rewards factual and high-quality outputs, and reduces compute/memory costs — all without altering the pretrained base model at inference. 1️⃣ PORAG: Policy-Optimized Retrieval-Augmented Generation • Built on Group Relative Policy Optimization (GRPO) • Fine-tunes with QLoRA adapters, guided by two specialized reward heads: – Retrieval Fidelity (factual correctness) – Response Quality (fluency, coherence, utility) • Enables single-shot decoding with KL-regularized policy updates for stability 2️⃣ ATLAS: Adaptive Token-Layer Attention Scoring • Detects when to retrieve via Multi-Layer Attention Gradient (MLAG) • Builds what to retrieve via Layerwise Representation Pooling (LRP) • Leverages hybrid attention-embedding relevance, with entropy-aware scaling to manage retrieval load 3️⃣ CRITIC: Cache Reduction via Importance-based Token Inclusion Criteria • Selectively retains high-impact tokens in KV cache • Hybrid importance signal: attention strength, entropy spread, gradient sensitivity • Reduces memory and latency for long-context generation, with adaptive compression ⚙️ Exhaustive Test-Time Inference Scaling — No Fine-Tuning Required Our framework integrates advanced inference-time scaling techniques that allow LLMs to trade off compute, latency, and answer quality — dynamically and intelligently. These include: ✔️ Entropy-Guided Decoding — invoke deeper reasoning under uncertainty ✔️ Monte Carlo Tree Search (MCTS) — strategic decoding via multi-branch simulation ✔️ Mixture of Agents — task-routed decoding using a pool of specialized policies ✔️ Lookahead Decoding — speculative multi-token generation and verification These methods are plugged into our test-time control module, which uses token-level entropy, compute budget, and reward models to dynamically select the optimal strategy for each input. 🧠 Conclusion Our framework enables scalable, retrieval-aware LLM inference by combining reward-optimized generation, introspective retrieval, and memory-efficient decoding. It is: ✔️ Transformer-agnostic ✔️ Plug-and-play for SLMs ✔️ Ready for enterprise-scale, compute-sensitive deployments 📌 Full paper on arXiv: https://lnkd.in/dErdTDf7 🤝 With Akash Das, SHIVAM GUPTA, and Venkataramana Runkana #LLMInference #TestTimeScaling

17

Most existing #lean4 datasets contain only correct proofs. Models learn error correction with RL, that's expensive. With UW Math AI lab we release a dataset of 260k erroneous Lean proofs with - compiler feedback - reasoning trace - corrected proof Improvements in Error Correction: - Goedel 8B: 2x - Kimina 8B: 3x Paper: https://lnkd.in/gaybt4bd

117

2 Comments

I used to think we might need to preprocess and structure input text at inference time to make LLM classifiers more robust (output consistent to similar input). But this paper introduced an interesting idea: it's enough to teach the model abstract reasoning during training — no need to force structure at inference. Promising direction for making LLM outputs more consistent without adding complexity at runtime.

4

Google Research Director Denny Zhou, who founded the LLM Reasoning Team at DeepMind, recently gave a great talk at Stanford’s CS25 class: Large Language Model Reasoning. This was one of the most intuitive talks about reasoning model training and application. He outlined new directions for training models to handle questions where answers aren’t easily verifiable. My team is currently working on an enterprise LLM project focused on using reasoning to extract relevant context and generate responses for unverifiable cases. It’s a tough but promising area, and it’s an interesting time to be building. [Link to Denny Zhou's talk: https://lnkd.in/gdJ6Mzi8]

1,026

8 Comments

We open-sourced Overmind: cuts ML model loading from 15s to 0.2s using zero-copy shared memory. At Meshy, loading model weights took longer than inference itself. So Bin Wang on our infra team built a caching layer that reconstructs PyTorch tensors without copying memory. Monkey-patch it in, no code changes needed. Give it a spin and let us know what you think! https://lnkd.in/gMtg4C48 https://lnkd.in/gzMVGu87 #OpenSource #PyTorch #AIInfra #AI3D #GenAI

56

My colleague David Crespo and I were recently on the Oxide and Friends podcast, where we talked about some of the ways we've been able to use LLMs like Claude Opus 4.5 to achieve a higher level of rigor and robustness than possible with humans alone. If you're a systems engineer who is despondent at how the dominant narrative is slop and shipping as much code as possible, quality be damned, then please give the episode a listen! LLMs have a lot to offer to detail-oriented infrastructure engineers. YouTube: https://lnkd.in/gAcrpQNb Transistor: https://lnkd.in/gzhsEFuQ

258

4 Comments

We developed an open-source AI tool to design concrete mixes that are stronger, more sustainable, and ready to build with faster—speeding up construction while reducing environmental impact. https://lnkd.in/gPCk8tCM But the impact of this AI tool is not just hypothetical! Amrize used Meta’s AI-based technologies to design a new low-carbon mix, and successfully deployed it in an at-scale slab-on-grade application at Meta's new data center in Rosemont, MN. Compared to the legacy mix, this new AI-designed mix is: 🦁 Stronger ⏱ Faster 🍃 Lower carbon ⏱ ️The ideal set time All this was achieved without needing any new materials, nor special equipment. Best of all, the AI is open-sourced. https://lnkd.in/g2KA7KZW This work was featured in a Meta engineering blog article published today! https://lnkd.in/gBU9HY8H

98

9 Comments