Reading an AI's "chain of thought" to predict its behavior? Turns out that's not very reliable 🔮 The fresh idea: make behavior prediction itself a learning task. Title: Forecasting Future Behavior as a Learning Task URL: 🔮 Overview A method to predict how large reasoning models (LRMs) will behave on new inputs. Instead of relying on explicit explanations, it introduces trainable "Behavior Forecasters" that analyze a single reasoning trajectory to predict outputs. ❓ Challenges Solved We want to understand and predict LRM behavior, but prior approaches have limits. ・Existing explanation methods don't scale to long reasoning trajectories ・Read as natural language, those trajectories are often unreliable A model's written "thoughts" don't necessarily reflect its actual behavior. 💡 Methodology & Proposed Approach ・It treats behavior prediction itself as a learnable task ・Training data comes directly from querying LRMs — no human annotation needed ・It runs in a single forward pass at inference ・Instantiated on two tasks: estimating answer consistency across reruns, and predicting how input modifications affect outputs ・End-to-end fine-tuning of the backbone, and initializing from the target LRM's weights, proved essential 📊 Experimental Results ・Behavior Forecasters outperform GPT-5.4 and Claude Opus-4.6 as "naive readers" ・And they achieve higher accuracy at a small fraction of the inference cost #LLMInterpretability# #ReasoningModels#

there will soon come a time, perhaps next year, when voice interaction + multimodal reasoning models are good enough and fast enough that you can actually just program like this

🗺️ Even frontier GPT-5 succeeds on just 14.4% of real-world spatial tasks. A new benchmark goes beyond staring at a static image and exposes how weak AI agents still are at active spatial reasoning. Title: SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks URL: 📝 Overview SpatialWorld measures whether multimodal LLMs can solve tasks by actively exploring 3D environments from a vision-only, egocentric viewpoint. It unifies eight different simulators across indoor, outdoor, and digital-game settings under a shared protocol, and evaluates 15 frontier models on 760 human-annotated tasks. The agent gets no prior map and no reference solution; it has to look, move, and decide on its own. ❓ Challenges Solved Prior spatial-reasoning benchmarks relied on passive evaluation via static VQA or pre-recorded video. That can't capture the interactive spatial understanding the real world demands, where an agent must move its own viewpoint to gather visual evidence and replan on the fly under partial observability. There was a large gap between recognizing a static scene and actually moving through an unfamiliar space to get a task done. 💡 Methodology & Proposed Approach ・The task is framed as a vision-only POMDP (Partially Observable Markov Decision Process) ・The agent receives only a natural-language goal and a single native-resolution egocentric RGB image, with no depth, maps, or semantic metadata ・Actions are issued through a high-level text interface covering navigation, viewpoint control, object interaction, and task completion ・It integrates eight backends: indoor (AI2-THOR, ProcTHOR, VirtualHome), outdoor (CARLA, EmbodiedCity), and digital games (Block3D, Snake3D, Rubik's Cube) ・Success is judged by whether the final terminal state satisfies the goal, not by matching the trajectory, and is validated by human annotators ・Beyond success rate, it measures step efficiency against human reference trajectories to surface inefficient behavior 🎯 Use Cases It offers a unified, fair way to evaluate the spatial abilities of home robots and autonomous agents before real-world deployment. It can systematically diagnose where long-horizon tasks that combine navigation and manipulation break down, serving as a rigorous testbed for improving spatial-reasoning models. 📊 Experimental Results ・Across 15 frontier models, physical-task success was 14.4% for GPT-5, 12.2% for Qwen-3.5-397B, 9.2% for Gemini-3.1-Pro, and 9.2% for Kimi-K2.5 ・On digital games, Gemini-3.1-Pro led at 39.0%, followed by GPT-5 at 36.4% ・By complexity, interaction-only tasks averaged 50.2%, navigation-only dropped to 8.6%, and combined navigation-and-interaction collapsed to just 4.2% ・Models with similar success rates showed very different efficiency scores, revealing heavy reliance on trial-and-error exploration ・Model rankings shifted dramatically across environments, with no single model dominating every category #AIAgents# #SpatialReasoning#

Meet the Grok 3 family, now on our API! Grok 3 Mini outperforms reasoning models at 5x lower cost, redefining cost-efficient intelligence. Grok 3, the world's strongest non-reasoning model, excels in tasks that need real world knowledge like law, finance, and healthcare.

In 2022, OpenAI researchers found something that broke every rule of machine learning. Their tiny model trained for 10,000 epochs. It learned absolutely nothing. Validation accuracy was dead stuck at 50%. Then at epoch 12,000, without warning, it jumped to 99%. This phenomenon is called "Grokking". And in 2026, it might be the most important discovery in AI nobody talks about. Neural networks can train for thousands of cycles without seeming to learn anything useful. Then, in a single epoch, they suddenly achieve near-perfect generalization. What started as a weird training glitch has become a foundational insight into how models truly learn. We’ve always been told: “If validation loss stops improving for a few hundred epochs, stop training.” Early stopping was the golden rule. Grokking says the exact opposite: Keep going. The model might look completely stuck, but real understanding is quietly forming under the hood. During that long, dead plateau, the machine isn't idle. It's doing deep internal work: - Circuits form, dissolve, and reform. - Spurious correlations get pruned away. - Weight patterns crystallize around true underlying rules. - The model shifts from brute-force memorization to genuine comprehension. It’s the machine version of a human “aha!” moment—a long, agonizing buildup followed by sudden clarity. Take modular addition as a real-world example. Researchers fed a small model just 30% of all possible examples. At epoch 500, it hit 100% training accuracy but stayed at 50% validation. It had memorized the test answers, but couldn't solve a new problem. At epoch 10,000, it still sat at 50% validation. It looked utterly hopeless. Then at epoch 12,000, it instantly shot to 99%. It didn't just guess right; it had grokked the actual mathematical rule. This explains the hidden mechanics behind the massive reasoning models we use today. When you see modern reinforcement learning or long-context reasoning models suddenly "click" after looking stuck, you are witnessing grokking at scale. Massive training runs aren’t wasteful, they are deliberately forcing the AI to stop memorizing and start thinking. And we are learning to induce this at inference time. Extended Chain-of-Thought prompts that force a model to think for thousands of tokens, self-consistency loops, and verification passes are all designed to do one thing: teach the model to grok your problem on the fly. The big philosophical takeaway is brutal for our short attention spans. Learning isn’t smooth. It isn’t gradual. It is discontinuous. Models, and humans, can stay “dumb” for ages, right up until they suddenly understand everything.