GR2 (Generative Reasoning Re-ranker) demonstrates strong gains for LLM-based re-ranking, outperforming recent work such as Google DeepMind’s PLUM and Kuaishou’s OneRec-Think. GR2 extends reasoning traces by enabling the LLM to invoke an optimized tool that loads item features associated with a Semantic ID directly into its context. These reasoning traces are a key lever for performance improvements and are optimized using Reinforcement Learning with verifiable rewards. We also present a practical recipe to prevent reward hacking.

RL post-training as a practical lever for turning general-purpose LLMs into specialized, policy-aligned classifiers. We demonstrate that RL 1. follows Sigmoid-like scaling behavior, 2. provides Strong label efficiency compared to SFT 3. need careful recipe for reward shaping and Monte-Carlo score aggregation to stabilize training and improve reliability.

As the second paper in a series on understanding chain-of-thought and in-context learning, this work investigates the role of CoT meta-training in shaping reasoning behavior in LLMs. We show that while CoT can accelerate learning, excessive reliance on explicit reasoning traces can harm generalization, and we outline principled strategies for effective CoT usage.

Move fast and live in future

The video of the talk is available on Youtube and the slides can be access on here

I explained how 360Brew unifies core AI components into a single layer, streamlining ranking and recommendation workflows, eliminating technical debt, and accelerating developer productivity.

CoT-ICL Lab is a framework to study chain-of-thought (CoT) and in-context learning (ICL) by decoupling causal structure from token processing functions. Experiments show that 1) CoT accelerates the accuracy transition to higher values across model sizes 2) deeper models require fewer in-context examples to leverage CoT effectively while more examples help shallow models match deeper model performance. Along with detailed analyses we provide theoretical insights. The code is available here.

In this work we show how to leverage On-Policy forward Knowledge Distillation, Model Compression (Pruning and Quantization), and Serving Optimizations (detailed on RadixAttention, FlashInfer, and tensor parallelism) to deliver a 20x reduction in both cost and latency while maintaining the quality of our 360Brew XL model.

In this report, we demonstrate that 360Brew model, a 150B parameter foundation model trained on 1T tokens can solve over 30 personalization tasks on LinkedIn platform without task-specific fine-tuning and no complex feature engineering. It can generalize to out-of-domain tasks and surfaces, and achieves performance similar to or better than the production model.

In this work we demonstrate that LLM's performance is affected by the relative distance between pieces of information in the context. The further apart the information is within long context, the more the model’s performance deteriorates.

Liger Kernel is a collection of Triton kernels designed specifically for LLM training. It can effectively increase multi-GPU training throughput by 20% and reduces memory usage by 60%.

This system-model co-design work focuses on leveraging synchronization in data parallelism hierarchical partitioning to avoid race condition in gradient updates for LLM training.

We formulate CoT as a reasoning graph and propose a prompt strategy for multi-step reasoning that can capture complex processes in tasks such as mathematics and commonsense reasoning.

The race condition between AllGather and device-to-device copy for the 2nd partition causes instability in training large models such as Llama-7B and Falcon-40B on a moderately large number of GPUs. After discovering the algorithmic issue, we landed the fix in the DeepSpeed repository.

We propose a framework for understanding how Data Augmentation interacts with class-level learning dynamics. We show that simple class-conditional augmentation strategies informed by our framework improve performance on the negatively affected classes.

Leveraging multimodal foundation models — XLM-R for language and FLAVA (plus MViT) for vision/video — to parse semantics across text, images and clips, we pair a hierarchical retrieval-and-ranking stack that distills billions of posts into the few hundred most relevant to each person in milliseconds, fueling unconnected-content discovery across Feed, Explore and Reels.

We trained a deep learning model on millions of explicit SMSL feedback, fusing them with existing engagement features to predict how users preference on new content—and then blend that preference score into real-time ranking. The result: giving everyone finer-grained, self-tuning personalization.

FAIM consolidated dozens of individual models into a unified PyTorch-based framework that processes text, images, and video simultaneously. This shift to generalized AI improved detection of hate speech, bullying, harassment, and violence while reducing computational overhead. The framework now powers over 85% of Facebook's multimodal integrity models and includes advanced features like Whole Post Integrity Embeddings (WPIE) for contextual understanding across content types.