Reasoning Models
Reasoning models are a new class of AI that "thinks" before answering: spending extra compute at inference time, building chains of thought and solving problems ordinary LLMs cannot — math, code, science. They kicked off the test-time-compute race. Here is all our coverage of reasoning models: releases, benchmarks (GPQA, ARC-AGI), limits and cost.
Video-Language Models Confuse Car Maneuvers: LoRA and Probes Fix the Error
An arXiv study finds video LLMs unreliable at identifying car motion from video, while lightweight LoRA and probes on intermediate features restore strong…
Anthropic cuts distillers' extraction of Claude's reasoning by 10–100x
Anthropic added a two-stage classifier and preserved thinking: the data distillers can collect fell by 90–99%, with false blocks under 0.1%.
Deep Cogito trained its Cogito v1 family on Runpod infrastructure on the road to ASI
San Francisco startup Deep Cogito trained its open Cogito v1 models on Runpod's cloud. The company's stated goal is artificial superintelligence; the lineup spans…
LoopOPD Teaches Looped Language Models to Improve Through Their Own Loops
Lapras replaces CoT with latent reasoning: up to +10.79% F1 with 23.9x fewer tokens
STI-OPD: A Probabilistic Teacher Intervention for Multi-Step Agentic Tasks
StoreBench: LLM Agents Run an Online Store and Lose to a Script — 49% vs. 97%
Midjourney Adds Thinking Mode for V8.2 and Shared Folders in Alpha
Midjourney's alpha gets a thinking mode that reruns V8.2 jobs after the model checks its own output against the prompt. Shared folders add Collaborators and Viewers…
Midjourney Tests Thinking Mode: A Rerun (Thinking) Button for Better Text Accuracy
Midjourney has launched Thinking Mode on its Alpha site. Rerunning an image with added "thinking" should improve prompt accuracy, typography and coherence.
How to Build a Stable Pipeline on the Fable 5 Traces Dataset in Google Colab
Thinking Machines Lab Releases Inkling-Small: Open 276B MoE Model Outperforms Its Teacher
Claude Opus 5 builds 3D games — a shooter, kart racing and Minecraft — from a single prompt
Vercel's DeepsecBench: GPT-5.6 Leads at Finding Vulnerabilities in Application Code
MUX: Continuous LLM Reasoning Through Multiplexed Tokens
Researchers present MUX, a method that compresses a language model's chain of thought into continuous latent tokens with no information loss and speeds up…
Thinking Machines Releases Inkling, an Open 975B-Parameter MoE Model
Mira Murati's Thinking Machines Lab has opened the weights of Inkling, a 975B-parameter multimodal MoE model trained on 45T tokens, under Apache 2.0.