FAST addresses sampling inefficiency in autonomous driving reinforcement learning by introducing Dynamic Parallel Sampling Alignment to decouple sampling loops from individual episode terminations. It achieves up to 1.78 times wall-clock speedup over single-clip baselines while maintaining statistical unbiasedness through Scaled Mask-Padding Optimization.
A new framework models energy consumption in Transformer training on multiple GPUs. It uses BERT architectural sweeps to link measured energy to compute, memory traffic, and hardware efficiency proxies. The model, inspired by roofline analysis, includes a speedup-based hardware-efficiency factor and predicts training energy across diverse GPU configurations.
Randomized YaRN enhances long-context reasoning by combining YaRN positional extrapolation with randomized positional encoding and a length curriculum. It outperforms standard fine-tuning on benchmarks like BABILong and MRCR, showing significant gains at far out-of-distribution context lengths.
Adaptive Data Scheduling (ADS) introduces a dual-level data scheduling framework that replaces uniform sampling with adaptive distribution over semantic clusters and policy-boundary sample selection. Experimental results show ADS improves average accuracy by 5.2% over GRPO across three LLMs and seven reasoning benchmarks, demonstrating its effectiveness as a general strategy for LLM RL post-training.
A theoretical analysis shows that off-policy degree, determined by gradient steps per rollout, significantly impacts importance sampling ratios and token update dominance. The study introduces Adaptive Clip Policy Optimization (ACPO), which adjusts clipping boundaries by token group variance, outperforming DAPO and CISPO on 3B and 7B models across mathematical, QA, and logic reasoning tasks.
A 200-step training comparison between a base OLMo3 600M model and a version with a DeepSeek-style Engram graft shows lower training and evaluation loss, faster grad-norm stabilization, and improved early learning behavior. The Engram graft, injected into layers 1 and 5, increases trainable parameters to ~1.7B but maintains only a 40k increase in active parameters per token, indicating efficient memory usage.
I wrote a free 15-part series detailing LLM internals, using Gemma 4 12B as the core example. Each part covers technical aspects from tokenization to serving, with real math, tensor shapes, and hardware constraints. The series includes a companion vLLM Deep Dive and is fully accessible without paywalls or email.
The RTX 5090 MSI consumes 475-500W during inference or diffusion training. The user reports no issues with the power cable, emphasizing that it should not be bent to ensure safe and stable operation.
A long-context decode performance cliff on AMD Radeon AI PRO R9700 (RDNA4) was resolved by enabling AITER Unified Attention in vLLM 0.22.1. The fix involves relaxing a CDNA gate to include RDNA4, disabling other attention backends, and using bf16 KV cache, resulting in significant speedups across all context lengths. FP8 KV is ineffective on this hardware, and the model's native 262K context is fully achievable with bf16, offering ~2.9× concurrency without needing FP8.
EvoTensile uses evolutionary algorithms to tune GEMM kernels for AMD GPUs, improving NT layout performance from 20 to 40 TFLOPS on Strix Halo. This speedup represents a significant advance over unoptimized kernels, though it remains below the theoretical roofline of 59.4 TFLOPS.
A study identifies shrinkage bias in E2M1-based FP4 formats due to geometric asymmetry, causing multiplicative error accumulation and training instability. The proposed UFP4 recipe uses uniform E1M2/INT4 grids and applies Random Hadamard Transform to all GEMMs, achieving lower loss degradation than E2M1 baselines in large-scale LLM pretraining. The authors recommend E1M2/INT4 as a first-class training primitive for future accelerators.
A pretrained speech classifier is repurposed as a backbone for guided diffusion-based speech generation. By attaching a lightweight subnetwork and training it under denoising score matching, the approach achieves high speech quality with reduced memory and computational cost, using a single model instead of two separately trained components.
UltraQuant enables 4-bit KV caching for context-heavy agents, reducing P50 time-to-first-token by 3.47x in late rounds and boosting output throughput by 1.63x over FP8 KV baseline. It achieves this using FP8 queries, FP4 KV tensors, UE8M0 group scales, and native scaled-MFMA on AMD CDNA4 GPUs, with optimizations for decode-attention kernels and robust design choices like asymmetric K/V treatment and Walsh-Hadamard rotation.
This paper examines how mixture-of-experts models maintain calibration under distribution shift. It finds that expert-level calibration ensures overall model calibration in hard-routed models but is insufficient for soft-routed models. The authors propose adversarial reweighting to penalize calibration errors in routed aggregates, improving accuracy-calibration tradeoff across tasks and shifts.
Direct Advantage Estimation (DAE) is extended to partially observable domains with minimal modifications. A discrete latent dynamics model reduces computational overhead by efficiently approximating transition probabilities, enabling scalable and sample-efficient deep reinforcement learning in high-dimensional observation spaces.
A study shows diffusion models can achieve global minimizers without explicit timestep embeddings. Ablation studies on CelebA and CIFAR-10 reveal time-agnostic models maintain high fidelity and outperform conditioned ones in FID, precision, and recall.
DeepGaLA is a neural-network surrogate that provides uncertainty-aware predictions for inverse problems in partial differential equations. It achieves accuracy comparable to Gaussian-process surrogates while maintaining efficiency in high-dimensional parameter spaces and incorporating differential-equation constraints.
A synthetic framework reveals that superposition increases over time with transient dips at task boundaries, indicating boundary-specific interference. Higher feature sparsity promotes superposition without inevitable forgetting, provided representation strength is maintained. Task-level effective rank grows with sparsity, showing broader capacity usage under sparse conditions.
A two-stage evolutionary strategy improves Physics-Informed Neural Network performance by first screening hyperparameter candidates via low-fidelity training, then refining top candidates with gradient-based optimization. The approach reduces mean error significantly across Advection, Klein-Gordon, and Helmholtz equation problems under fixed computational budgets.
A pretrained speech classifier is repurposed as a backbone for guided diffusion-based speech generation. By attaching a lightweight subnetwork and training it under denoising score matching, the approach achieves high speech quality with reduced memory and computational cost, using a single model instead of two separately trained components.