Researchers Catch LLMs Looping Into the Same Sentences After 50 Rewrites
A new paper models repeated LLM rephrasing and translation as Markov chains, revealing when outputs collapse into loops versus keep drifting.
20 stories
A new paper models repeated LLM rephrasing and translation as Markov chains, revealing when outputs collapse into loops versus keep drifting.
An open-source agent skill turns one photo into a finished AI talking-head video, with lip-sync, voice unification, and 48 packaging components.
An abliterated 35B mixture-of-experts coder with a grafted multi-token prediction head delivers frontier agentic coding at Q4 on 24GB consumer hardware.
A community photorealism checkpoint built on SDXL crosses 859k downloads, delivering lifelike portraits and scenes without a custom pipeline.
Nous Research consolidates eighteen first-party Hermes Agent plugins into a single repo, each one install away from the plugin catalog.
Valeo.ai's IGLOSS swaps CLIP text embeddings for text-to-image prototypes, setting new open-vocabulary segmentation records on nuScenes and SemanticKITTI.
CopilotKit's OpenMuse ships an MIT-licensed personal agent template with a persistent browser, Linux terminal, Gmail integration, and durable task plans.
A Czech-Moroccan robotics team shows how two drones can learn each other's antenna radiation patterns mid-flight, hitting 4.56 dB RMS prediction error.
A new open-source tool pushes every MLP neuron in Qwen3-4B up and down, saves every rollout, then lets you browse the model as a 3D map.
DecisionRAG replaces vector search with an LLM that judges candidates in parallel, beating dense retrieval plus an 8B reranker on multiple benchmarks.
A new preprint shows that steering vectors extracted from text can control image generation and vice versa inside unified multimodal models, zero-shot.
Pi Pocket turns the Pi coding agent into a phone-friendly, multiplayer web app with durable sessions that survive server restarts mid-run.
GitHub swapped its LLM-based secret scanner for a fine-tuned ModernBERT classifier, boosting acceptance 11.5%, cutting scan time 86%, and serving costs 98%.
DeepBeepMeep's repack of MiniMax's open-weight H3 video+audio model lands with GGUF quants and new 3-step accelerator LoRAs that make it run on modest GPUs.
Datalab's new 689M-parameter OCR model hits 85.9 on OmniParseBench, runs 1.7x faster than its predecessor, and ships with character-level tokenization.
NVIDIA's new 30B hybrid Mamba-Transformer MoE activates just 3B parameters, runs on a single H100, and handles up to 1M tokens of context.
Moonshot flipped the default so Kimi Code tackles work solo, with multi-agent swarms now strictly opt-in across CLI and Desktop.
StreamEval scores interactive streaming video models across five tasks, with incremental inputs, 15 to 300 second sessions, and an open harness.
Tencent's WeChat Vision team released an open 2B/4B vision-language model that handles detection, OCR, segmentation and alpha matting without a separate SAM backbone.
Comfy-Org bundles FLUX.1-dev and its control variants into single-file safetensors with built-in text encoders, cutting setup friction for ComfyUI users under 24GB VRAM.