Machine Learning

Open-Reasoner-Zero pioneers scalable, accessible RL training for reasoning in LLMs, achieving superior performance with a minimalist approach.

RL with Verifiable Rewards is now expanding to diverse domains like medicine!

ReaRec: Unleashing latent reasoning power for sequential recommendation through inference-time multi-step reasoning.

This paper enhances Reinforcement Learning from Human Feedback (RLHF) by tackling reward hacking and response diversity issues through improved data construction methods.

LogQuant: 2-bit quantization for KV cache, superior accuracy!

CURA: Verbal process supervision improves coding agents.

DeLoRA: Decoupling angles and strength in low-rank adaptation for robust & efficient finetuning of large models!

Gumbel-Softmax Flow Matching enables controllable biological sequence generation with straight-through guidance, scaling efficiently to high-dimensional simplices.

RL fine-tuning enhances reasoning in small LLMs, achieving competitive performance with limited resources, despite optimization & length challenges.

UAE-3D: A unified latent space approach for efficient & high-quality 3D molecular generation, outperforming existing methods in accuracy and speed.

Frac-Connections: An efficient alternative to Hyper-Connections that divides hidden states into fractions.

DAPO: Open-sources a LLM reinforcement learning system that achieves SOTA AIME scores, fostering reproducible research at scale.

Transformers can achieve state-of-the-art performance without normalization layers via Dynamic Tanh (DyT), offering a simpler and more efficient alternative.

KArAt: Can Learnable Attention Beat Standard Attention in Vision Transformers?

Navigating millions of models is hard. This paper charts Hugging Face, revealing model relationships and attribute predictions.

LLMs can now reason more efficiently!

Rimer: RWKV-7 empowers superior time series modeling, offering a simple yet effective alternative to Transformers with fewer parameters.

LoRACode enhances code embeddings using LoRA, achieving SOTA in code retrieval with minimal computational cost.

This paper reviews AI4SE benchmarks, introduces BenchScout for benchmark discovery, and proposes BenchFrame for benchmark enhancement, demonstrated via HumanEvalNext.

Multi-attempt RL refines LLMs, significantly boosting accuracy on math tasks by enabling them to learn from failures through user feedback.