How do you design an LLM agent that decides for itself what to store in long term memory, what to keep in short term context and what to discard, without hand tuned heuristics or extra controllers? Can a single policy learn to manage both memory types through the same action space as text generation? Researchers from Alibaba Group and Wuhan University introduce Agentic Memory, or AgeMem, a framework that lets large language model agents learn how to manage both long term and short term memory as part of a single policy. Instead of relying on hand written rules or external…

What does an end to end stack for terminal agents look like when you combine structured toolkits, synthetic RL environments, and benchmark aligned evaluation? A team of researchers from CAMEL AI, Eigent AI and other collaborators have released SETA, a toolkit and environment stack that focuses on reinforcement learning for terminal agents. The project targets agents that operate inside a Unix style shell and must complete verifiable tasks under a benchmark harness such as Terminal Bench. Three main contributions: A state of the art terminal agent on Terminal Bench: They achieve state of the art performance with a Claude Sonnet…

In this tutorial, we demonstrate a realistic data poisoning attack by manipulating labels in the CIFAR-10 dataset and observing its impact on model behavior. We construct a clean and a poisoned training pipeline side by side, using a ResNet-style convolutional network to ensure stable, comparable learning dynamics. By selectively flipping a fraction of samples from a target class to a malicious class during training, we show how subtle corruption in the data pipeline can propagate into systematic misclassification at inference time. Check out the FULL CODES here. import torch import torch.nn as nn import torch.optim as optim import torchvision import torchvision.transforms…

How far can a mid sized language model go if the real innovation moves from the backbone into the agent scaffold and tool stack? Meta and Harvard researchers have released the Confucius Code Agent, an open sourced AI software engineer built on the Confucius SDK that is designed for industrial scale software repositories and long running sessions. The system targets real GitHub projects, complex test toolchains at evaluation time, and reproducible results on benchmarks such as SWE Bench Pro and SWE Bench Verified, while exposing the full scaffold for developers. https://arxiv.org/pdf/2512.10398 Confucius SDK, scaffolding around the model The Confucius SDK…

In this tutorial, we demonstrate how we use Ibis to build a portable, in-database feature engineering pipeline that looks and feels like Pandas but executes entirely inside the database. We show how we connect to DuckDB, register data safely inside the backend, and define complex transformations using window functions and aggregations without ever pulling raw data into local memory. By keeping all transformations lazy and backend-agnostic, we demonstrate how to write analytics code once in Python and rely on Ibis to translate it into efficient SQL. Check out the FULL CODES here. !pip -q install "ibis-framework[duckdb,examples]" duckdb pyarrow pandas import ibis…

A team of Stanford Medicine researchers have introduced SleepFM Clinical, a multimodal sleep foundation model that learns from clinical polysomnography and predicts long term disease risk from a single night of sleep. The research work is published in Nature Medicine and the team has released the clinical code as the open source sleepfm-clinical repository on GitHub under the MIT license. From overnight polysomnography to a general representation Polysomnography records brain activity, eye movements, heart signals, muscle tone, breathing effort and oxygen saturation during a full night in a sleep lab. It is the gold standard test in sleep medicine, but…

In this tutorial, we demonstrate how to build a unified Apache Beam pipeline that works seamlessly in both batch and stream-like modes using the DirectRunner. We generate synthetic, event-time–aware data and apply fixed windowing with triggers and allowed lateness to demonstrate how Apache Beam consistently handles both on-time and late events. By switching only the input source, we keep the core aggregation logic identical, which helps us clearly understand how Beam’s event-time model, windows, and panes behave without relying on external streaming infrastructure. Check out the FULL CODES here. !pip -q install -U "grpcio>=1.71.2" "grpcio-status>=1.71.2" !pip -q install -U apache-beam crcmod…

Technology Innovation Institute (TII), Abu Dhabi, has released Falcon-H1R-7B, a 7B parameter reasoning specialized model that matches or exceeds many 14B to 47B reasoning models in math, code and general benchmarks, while staying compact and efficient. It builds on Falcon H1 7B Base and is available on Hugging Face under the Falcon-H1R collection. Falcon-H1R-7B is interesting because it combines 3 design choices in 1 system, a hybrid Transformer along with Mamba2 backbone, a very long context that reaches 256k tokens in standard vLLM deployments, and a training recipe that mixes supervised long form reasoning with reinforcement learning using GRPO. Hybrid…

NVIDIA has just released its new streaming English transcription model (Nemotron Speech ASR) built specifically for low latency voice agents and live captioning. The checkpoint nvidia/nemotron-speech-streaming-en-0.6b on Hugging Face combines a cache aware FastConformer encoder with an RNNT decoder, and is tuned for both streaming and batch workloads on modern NVIDIA GPUs. Model design, architecture and input assumptions Nemotron Speech ASR (Automatic Speech Recognition) is a 600M parameter model based on a cache aware FastConformer encoder with 24 layers and an RNNT decoder. The encoder uses aggressive 8x convolutional downsampling to reduce the number of time steps, which directly lowers…

In deep learning, classification models don’t just need to make predictions—they need to express confidence. That’s where the Softmax activation function comes in. Softmax takes the raw, unbounded scores produced by a neural network and transforms them into a well-defined probability distribution, making it possible to interpret each output as the likelihood of a specific class.  This property makes Softmax a cornerstone of multi-class classification tasks, from image recognition to language modeling. In this article, we’ll build an intuitive understanding of how Softmax works and why its implementation details matter more than they first appear. Check out the FULL CODES here.…