H

Harsh Sharma

@harshsharma_ai, 26K on Instagram

yooooo

Curated stories

Posterior Matching Distillation improves few-step image generation

Diffusion distillation compresses iterative sampling processes to speed up image generation, but traditional methods often lose sample diversity. Posterior Matching Distillation addresses this by combining mode-covering posterior sampling with flexible generation steps.

RL post-training in diffusion transformers is activation steering

Reinforcement learning post-training significantly improves image and video generation models. A new study shows that much of this improvement comes from a simple constant shift rather than complex prompt-specific changes.

Nvidia researchers argue huge AI context windows are a dead end

Recent advances allow artificial intelligence models to process massive amounts of text at once. However, a new study suggests that simpler methods often match the performance of these expensive memory expansions.

Anthropic and Stanford researchers independently release proofs of Riemann hypothesis

An Anthropic mathematician and the Claude model have presented a computer-verified proof of the Generalized Riemann Hypothesis. Just hours later, Stanford researchers released an independent human-authored draft attempting the same feat.

Inkling 2 Mini uses looped layers to reason locally

Thinking Machines has released Inkling 2 Mini, an open mixture-of-experts model that uses looped layers to think internally before generating text. The model runs locally on standard hardware while matching larger systems on coding and intelligence benchmarks.

Calibrated decision rewards improve language model training without answer keys

A new reinforcement learning method uses a probability-based judge model to score language model outputs without needing a reference answer. The approach matches the performance of traditional answer keys while avoiding common evaluation flaws.

Resources