Microsoft Research has released Echoverse, a collection of twelve deep synthetic environments designed to train computer-use agents. The initiative includes ten domain-specific worlds and two capability-focused worlds, all built to maintain coherent state and behavioral fidelity.
OpenAI has developed GPT-Red, an automated safety red-teaming model trained via self-play reinforcement learning to uncover vulnerabilities and generate adversarial data for improving future models.
Microsoft Research has introduced EvoLib, a framework that enables large language models to learn from their own experience during inference without requiring ground-truth labels or external feedback. Rather than storing raw experiences as static memories, EvoLib transforms them into reusable skills and reflective insights that are continually refined, consolidated, and reweighted.
DharmaOCR achieves higher extraction quality and stability than Mistral OCR4 and Unlimited-OCR on Brazilian Portuguese through domain-specific training.
Researchers propose Unsupervised On-Policy Self-Distillation (U-OPSD), a method that achieves post-training improvement for large language models using only the model's own generations without external supervision. The approach samples multiple rollouts to construct a pseudo-solution via majority vote, then distills a teacher distribution into prefixes of the model's longest incorrect completion.
Apple has introduced a new architecture for its third-generation foundation models (AFM3) that utilizes "Instruction-Following Pruning" to significantly reduce active parameter count. The model is designed to activate approximately 20% of its MLP layers by making routing decisions once per prompt rather than per token.
NVIDIA's NeMo team has released Molt, an open-source agentic reinforcement learning framework designed to reduce the complexity of algorithm iteration for researchers. The codebase is approximately 8.6K lines of RL code, significantly smaller than comparable frameworks like verl (62K lines) and slime (25K lines).
The author proposes a hypothesis that failed tool-using agent traces, such as wrong tool selections or malformed arguments, can serve as valuable fine-tuning data to train models away from these errors. The post seeks community feedback on whether training on corrected failure traces is effective or detrimental.
PIN Architecture v2 introduces a method to pre-select a neural network's size and inference speed before training by using weight tying. This technique forces groups of cells to hold identical values, allowing the network to maintain its width while drastically reducing the number of stored weights.
Huawei has released the open-source weights for openPangu-2.0-Pro, a Mixture of Experts (MoE) large language model trained on Ascend hardware. The model features 505 billion total parameters with 18 billion activated parameters and supports a context length of 512k tokens.
Researchers present Memory Decoder at Scale, a system that scales parametric long-term memory models up to 6.9B parameters and pretrains them on 300B tokens. To address the infeasibility of standard Faiss pipelines at this data scale, the authors implement a distributed indexing pipeline with sparse, batch-wise loading of kNN distributions.
Researchers introduce Relay On-Policy Distillation (Relay-OPD) to address prefix failure in standard on-policy distillation, where student errors lead to unreliable supervision. The method uses a teacher-student continuation asymmetry as a trigger for the teacher to briefly take over and redirect the trajectory before the student resumes.
Google is now offering a distillation service for its Gemini models. This allows users to distill the capabilities of larger Gemini models into smaller, more efficient versions via the Google Cloud platform.
OpenForgeRL is an open-source framework that allows researchers to train harness-based AI agents end-to-end using standard reinforcement learning stacks. It decouples training from inference by employing a lightweight proxy to record model calls as data and a Kubernetes orchestrator to manage rollouts in remote containers.
CrowdTensor is an Apache-2.0 open-source protocol for checkpointed volunteer model-training campaigns that pins immutable model and data revisions to give admitted CPU, GPU, or TPU cells bounded work. The project has released a Beta enrollment process alongside a Draft RFC for a Qwen2.5-7B GSM8K campaign.
A research preview of ThetaScan v0.1, a fixed-state nonlinear token mixer, has been released as an open-source implementation. The architecture ensures that every token computes its memory write from the current input and shared parameters rather than an evolving fast state, allowing writes to compose through an associative prefix scan.
Researchers present OpenForgeRL, an open-source framework that allows for the end-to-end training of AI agents using complex inference harnesses like Claude Code and OpenClaw. The system decouples training from inference by employing a lightweight proxy to record model calls as data and a Kubernetes orchestrator to manage remote container rollouts.
OpenForgeRL is an open-source framework that allows researchers to train harness-based AI agents end-to-end using standard reinforcement learning stacks. It achieves this by decoupling training from inference through a lightweight proxy that records model calls as data and a Kubernetes orchestrator that manages remote container rollouts.
A Hugging Face forum user investigates whether small, domain-specific large language models (LLMs) in the 7B–14B parameter range can effectively replace larger models in real-world production environments.
A new implementation of the ELIZA chatbot utilizes a sparse autoencoder architecture to store conversation entries in a latent memory layer, aiming to improve upon traditional keyword-matching methods. The system employs a T5 encoder-decoder structure with a 32,768-neuron memory layer, where similar entries are activated via contrastive learning.