Knowledge catalogue
I am trying to learn concepts like On Policy Distillation (OPD), On Policy Self Distillation (OPSD) and how do they compare to RL algorithms like GRPO. There are a lot of papers on this, but because o
It's an encoder-only transformer that consumes past(blood glucose + carbs + insulin) and future(carbs + insulin) and predicts future blood glucose for the next 2 hours. Announced meals and boluses/bas
Kimi K3 by Moonshot reached the frontier as an open-weight model. Artificial Analysis ranks it fourth of 580 models, behind only Claude Opus 5, Fable 5, and GPT-5.6 Sol. Moonshot released more than th
Hey everyone, I recently open-sourced a new Python package called ganfs (Generative Adversarial Network Feature Selection), and I wanted to share it with the community. The Problem: Selecting the best
I've always found it a little strange that AI is everywhere, but the codecs we use in practice are the traditional hand-engineered systems like h.264, h.265, av1. Alexnet started the wave of neural ne
We developed a leaderboard ranking frontier model security. There's no shortage of model capability rankings, but we didn't find anything comparable for model security. Yet security is becoming increa
While building this workflow a thing that surprised me was that, initially I thought the pipeline was complete: From Goal to → Decompose → Research → Specification → Implementation It successfully bro
I work on PostSlate, a video editing tool, and this comes out of our own work. We run ML models on-device, face detection and embedding among other things, which means we can't assume anything about t
ML research is progressing at breakneck speed where frontier labs in both academia and industry have access to considerably large computes (GPUs). Where do small labs or independent researchers go in
My data set is a list of images, each equipped with a a couple sentences of text. A user would search primarily with text only. My default approach is using BM25, but how would I facilitate searching
Hello guys. I found some problems in current frontier models. And want to share. # math_code_hallucination > Record of a failure caused by combining mathematics and code in a single prompt. --- ## Cas
I'm really confused about what the point of the prompt injection was (speaking as an author). Is it just a study? I would really prefer that they took action against the AI-generated reviews. Obviousl
One of the papers I reviewed has what seems to be entirely LLM-generated rebuttals, and the original paper is also clearly LLM-generated, with Claude-speak everywhere. While the authors acknowledge LL
TL;DR: Most RL post-training algorithms optimize the current batch and move on. But after an update, did the new policy actually become better? We introduce Policy Improvement Reinforcement Learning (
Largest empirical study of AI penetration in academic publishing ever conducted—51%-by-2025 is the most authoritative quantitative marker yet of how thoroughly LLMs have reshaped scientific writing, a
Hi everyone! 👋 I built and trained the complete Transformer architecture from scratch using pure PyTorch (`torch.nn` primitives) based on the original 'Attention Is All You Need' paper. I trained the
This was my Bachelor's Final Project: implementing YOLO26n inference completely from scratch using ARM64 Assembly Language and C, without relying on existing inference frameworks. The goal was to unde
I'm a software engineer who mainly builds softwaes/applications, and I'm starting to work on machine learning projects. Since ML workloads often require GPUs, I know services like Google Colab and Kag
NOTE -> I expect answer from people who actually have experience and strong understanding of these. please give something beneficial. I'm building a SaaS platform in Sri Lanka that handles documents a
I have been running some experiments with smaller open-weight LLMs on multiple-choice questions of Swedish medical licensing exams. On a dataset called MedQA-SWE, GPT-4 scored 84% accuracy in 2024 and
Hey everyone, I have been looking into how people source compute for their Inference workloads (and in general). I wanted to understand some specific pain points here. If you've used online services l
There are a few reasons why problems from International Mathematical Olympiad function as a good benchmark for LLMs: - The problems are new, not included in the training data of any model - Hard math
I've been chasing the question of what algorithms a transformer can actually express -- separate from what it can learn. So I built a compiler: define a computation graph in ordinary Python, and it pr
Built an open-source AI coding agent that was 7%–75% cheaper than a cold 'claude -p' run on 6/6 well-localized tasks across repositories up to ~82k LOC. The biggest difference: Cold agent: 6.83, 207 t
Hi everyone, I recently received an interview offer for ML position at Adyen, and the first round will be a live coding round on HackerRank. I scheduled it for the end of August because that was the l
The interesting finding from a new [arXiv paper](https://arxiv.org/abs/2607.16165) isn't that a frontier vision model failed a new benchmark, that happens weekly, but the specific shape of the failure
We spent the last months consolidating seven separate sequence classifiers into one multi-head model, our apex model, so to speak, and since the weights are now public, I wanted to share what worked a
Paper:https://arxiv.org/abs/2607.19058 Code (GitHub):https://github.com/nuemaan/skewadam Hi everyone, I just published a preprint on a new optimizer designed to tackle the massive VRAM bottleneck in M
I've been building an AI that learns to play the classic Snake game through reinforcement learning. The goal is to reach high scores while keeping training time as low as possible. The current version
Hi everyone, I'm working on extracting the hierarchical structure of long PDF documents (legal/regulatory text, lots of numbered sections) and would like to gather some feedback on my approach before
TL;DR: I’m reproducing the trait-persistence result from arXiv:2606.24014 on one RTX 3090. Before I can test persistence I need to install a trait via RL — and my GRPO run moves the trait only +2.4 po
Hi everyone, We've open-sourced Tri-Net v2, the official implementation accompanying our recently published Scientific Reports (Nature Portfolio) paper: 'Tri-Net: Unified Deep Learning for Skin Lesion
I worked on this project (https://github.com/workofart/harness-training) for the past few months to reframe 'Agent-driven Self-improving Harness' to 'Harness Training'. The idea is simple, the harness
This post likely presents an enhanced demonstration of DVD-JEPA, a video variant of the Joint Embedding Predictive Architecture model. The JEPA framework has been extended to video tasks (V-JEPA) , an
A researcher released an open-source softmax-free attention model at GPT-2 Medium scale (354M parameters trained on 11.5B tokens) that uses structural sparsity and tile-skipping kernels to reduce VRAM
This handbook covers the technical aspects of running large language models efficiently at scale, focusing on GPU optimization techniques including GPU internals, key-value (KV) cache management, batc
TSAuditor is a contamination auditing framework for time series foundation models that uses probe adaptation dynamics to detect when evaluation datasets may have been exposed during pretraining, valid
This post likely covers the process of creating a custom drone simulation environment using MuJoCo, a physics engine commonly used in machine learning research. The project involves leveraging MuJoCo'
Gemini 3.1 Pro and Gemini 3-Pro lead visual reasoning benchmarks , with GPT-5.2, Kimi-K2.5, and GPT-5.2-Pro following . A 2026 evaluation benchmarked 15 leading multimodal models on visual reasoning a
NeurIPS 2026 used an AI detector to identify policy violations, resulting in 178 desk-rejected submissions (18.4% of all submissions) and 123 flagged for further review (12.7%) . The detection approac
Research demonstrates that LLM-based agents can generate functionally correct patches that pass all tests while still containing security vulnerabilities, challenging the assumption that test-passing
I need to check the actual content of this Reddit discussion to provide an accurate summary. Graph Neural Networks can learn environmental effects on galaxy properties, incorporating spatial relations
This research discusses a technique for enabling real-time multilingual automatic speech recognition (ASR) on edge devices by dynamically switching between compact monolingual models rather than using
Reviews and discussions for all accepted papers at ICML are made public on OpenReview after the reviewing period concludes. Authors of rejected papers may also opt-in to have their reviews and discuss
The N.E.A.T (NeuroEvolution of Augmenting Topologies) algorithm is an evolutionary machine learning approach that evolves neural networks to solve control problems. This post describes applying N.E.A.
Physics-informed neural networks (PINNs) are machine learning models that incorporate physical laws and equations as constraints during training to solve differential equations. This post likely discu
This call for papers announces a workshop focused on the growing need for efficient and effective techniques for editing trained models, especially large generative models. The workshop solicits paper
This post describes implementing DCGAN (Deep Convolutional Generative Adversarial Network) inference on resource-constrained microcontroller hardware, achieving image generation with a 12.6 million pa
MergeNB is a VSCode extension to resolve Jupyter Notebook git merge conflicts . It applies three-way merge logic on matched notebook entities (source, metadata, outputs, execution_count) to automatica
WordDetectorNet predicts an axis-aligned bounding box for each inner-word pixel using a ResNet18-based neural network, then applies DBSCAN clustering to refine the predictions . The system demonstrate
NuExtract3 is a unified 4B vision-language reasoning model for document understanding that combines structured information extraction with image-to-Markdown conversion, suitable for OCR and RAG prepro
Liveness detection and synthetic media detection models often fail to generalize across unseen data and struggle with content from different models. Understanding how factors like data source diversit
LxMLS 2026 is a 6-day in-person event scheduled for July 20-25 at Instituto Superior Técnico that covers machine learning topics from theory to practice for solving natural language processing problem
CANTANTE addresses the challenge of optimizing LLM-based multi-agent systems where system-level performance scores are available but individual agent parameters cannot be directly optimized. The frame
An OpenAI general-purpose reasoning model autonomously disproved a conjecture posed by Paul Erdos in 1946, overturning 80 years of mathematical belief. The breakthrough concerns the planar unit distan
A researcher from a resource-limited region discusses challenges with attending an accepted ICML workshop (GlobalSouthML) in South Korea and seeks advice on affordable attendance options, likely explo
Physics-informed neural networks (PINNs) use specialized loss functions that combine data-fitting terms with physics constraints derived from differential equations, allowing the network to learn solu
A developer created a visualization tool that displays GPT-2's internal concept activations in real-time during text generation, representing the model's 'thinking' process as a 3D graph that updates
Witchcraft is a Rust reimplementation of Stanford's XTR-Warp semantic search engine that uses a single-file SQLite database for storage, enabling client-side deployment. The system operates completely
A vulnerability in DeepSeek's website exposed a significant amount of data, including user chats. A publicly accessible ClickHouse database belonging to DeepSeek allowed full control over database ope