Vals AI's MysteryMechanism Benchmark Shows GPT-6 Astra Tops Science Test at 53%
A new Vals AI benchmark forces frontier models to rediscover hidden math laws through experiments, exposing where extra reasoning stops paying off.
// Intelligence
Reasoning
A focused feed of models, agents, research and open-source releases for people building with AI.
Updated
Google's Tiny 10K-Parameter Neural Cellular Automata Beats Bigger AI on Maze Reasoning
A Google team shows tiny grids of locally-connected cells can solve Sudoku, mazes, and ARC-AGI through emergent, visible reasoning dynamics.
GPT, Gemini, and Claude All Pick the Same Right Answers
A new study of 270 puzzles with multiple correct answers finds AI models cluster tightly together and diverge sharply from how humans pick solutions.
Sakana AI's SAIL Triples Robot Success Rates by Searching Before Moving
Sakana AI and the University of Tokyo show that a vision language model can plan robot motions more reliably by testing and refining candidate trajectories in simulation before touching the real robot.
Anthropic's Claude Breaks Physics Record With a Nine-Loop Particle Calculation
Anthropic's Claude autonomously pushed a frontier theoretical physics calculation from eight to nine loops on Claude Science, costing only a few thousand dollars.
Recursive Self-Distillation Doubles Qwen3-8B Math Accuracy to 65.97%
A new arXiv paper proposes DCE+SRCL, letting the teacher model co-evolve with its student to lift Qwen3-8B math accuracy by 35 points.
Vals AI Finds Claude Opus 5.5 Beats GPT-6 Astra on Math Proofs
Vals AI ran GPT-6 Astra, Claude Opus 5, and Opus 5.5 across five reasoning levels on Lean 4 proof tasks and found diminishing returns kick in fast.
UCLA's PTTS Gives AI Parallel Reasoning Branches Separate Plans, Gaining 13.4 Points
A planner-executor framework that coordinates parallel reasoning branches instead of independent sampling, lifting pass@64 on math benchmarks by up to 13.4 points.
Vals AI's Claude Opus 5.5 Agents Proved a Faster Shortest-Path Algorithm
Ten Claude Opus 5.5 agents collaborated on a message board for 15 hours to produce C-HD, a Lean-verified shortest-path algorithm that beats published bounds in a specific density regime.
Microsoft's Agensh Runs 1,024 Coding Agents Without a Central Controller
Microsoft Research shows removing the central orchestrator lets thousands of coding agents self-organize, lifting test-pass rates by up to 21 points.
Microsoft Shows 5 Communicating Agents Matching 33 Independent AI Attempts
A Microsoft Research and UC Berkeley team shows agents sharing a scratchpad beat parallel independent runs, setting new records on ARC-AGI-3, polyomino packing, and MNIST compression.
Kyutai's Voice of Reason Thinks While It Speaks, Hitting 77% Math Accuracy
Kyutai's Voice of Reason turns GLM-4-Voice into a speech-native math solver, jumping GSM8K accuracy from 27.3% to 77.1% with no added latency.
GPT-6 Astra Cracks a Decade-Old Voting Theory Problem Nobody Could Solve
Epoch AI marks the first Major Advance on its unsolved-math benchmark, with GPT-6 Astra driving the proof in an interactive session with three mathematicians.
Sakana AI Bets Against Transformers With Its Frontier Intelligence Group
Sakana AI launches the Frontier Intelligence Group, a research collective betting that Transformers and scaling laws aren't the final answer to AGI.
Vals AI's MysteryMechanism Benchmark Tests If AI Can Rediscover 222 Scientific Laws
A new benchmark seals 222 scientific laws and asks agents to rediscover each one from scratch using a tight experiment budget, with GPT-6 Astra leading at 53.2%.
Yifan Zhang's RLT Grows Transformer Depth With Every Token Generated
A new architecture proposes latent reasoning that grows with sequence length, sharing one recurrent state across prompt, response, training, and RL replay.
OpenAI's GPT-6 Astra Cracks Epoch's Hardest Math Benchmark in 14 Months
A math benchmark built to resist AI just fell. GPT-6 Astra cracked the final Tier 4 problem, closing out a 98 percent run in 14 months.
Looped Flows Beats Rival Reasoners on ARC-AGI With Just 7M Parameters
A new recurrent reasoning method mixes denoising with looped hidden states, hitting 58.8% on ARC-AGI-1 and 12.2% on ARC-AGI-2 with just 7M parameters.
Together AI Says Open-Source Kimi K3 Beats Claude Fable 5.1 on Legal Tasks
Together AI highlights that Moonshot's open-weight Kimi K3 outscores Anthropic's newer Claude Fable 5.1 by 60% on the hard slice of Harvey's autonomous legal agent benchmark.
Autopoietic Game Theory Shows AI Agents Stop Cheating When Thinking Costs Energy
A new evolutionary model shows that when computation, replication, and social behavior all draw from one energy budget, cooperation emerges naturally in populations of random Z80 programs.
UT Austin Finds Fractal Chaos Explains Why AI Reasoning Costs Spike 10x
A new paper shows recurrent-depth reasoning models behave like chaotic dynamical systems, where hard problems create fractal basins that trap thinking near wrong answers.
OpenAI's 10,000-Agent Swarm Solves a 100-Year-Old Math Mystery
An internal OpenAI model coordinated roughly 10,000 agents for 88 hours to produce a Lean-verified finite-time blowup proof for 3D Navier-Stokes.
Kyoto University Finds a Hard Speed Limit on What SGD Can Learn
A new physics inspired bound explains which patterns stochastic gradient descent learns first, tying acquisition speed to Fisher information flow.
Meta's AIRA₃ Beats 4,000 Human Teams to Win Kaggle Gold
Meta's next-generation autonomous research agent AIRA₃ placed 8th of roughly 4,000 human teams in an NVIDIA-run Kaggle contest to fine-tune Nemotron.
No stories match these filters.