A focused feed of models, agents, research and open-source releases for people building with AI.
Ai2 released StateBench , a benchmark splitting state tracking into four algebraic complexity tiers. Frontier models including GPT-6 Sol, Claude Opus 5.5, and Gemini 3.8 Flash drop to 13% or lower on hardest problems.
AI agent lab at Institute of Science Tokyo ran a full quantum physics study end-to-end. Agents derived an exact L^2 + 1 calibration law for NV-center ensembles in diamond. 75 proposals were generated; only one was accepted after PI-agent review.
New arXiv paper studies how transformers internally represent sequences from multiple interacting agents. Proves any joint stochastic process factorizes into per-agent Bayesian beliefs via a feedback-loop HMM construction.
LongChang tests whether LLMs can notice and explain altered physics with no task or hints given. Only GPT-6-Astra discovered hidden laws, scoring 16.7% while 11 other frontier models scored 0%.
Kyutai shows Neural Thickets gains come from logit diversity, not weight-space experts. Logit PSV adds one random bias vector over the vocabulary, model stays frozen. Beats full-weight PSV on 13 of 15 reasoning benchmarks across four tested LLMs.
Vals AI explains how ten Claude agents produced a Lean proof for seven electrons on a sphere, and names which cases should fall next.
A new paper treats atoms as agents chatting through Bayesian inference, letting a neural model discover order parameters from raw spin snapshots alone.
A new study shows cranking reasoning effort to maximum can cost 42x more, lose easy tasks, and nudge agents into shortcut behaviors like reading gold fixes.
An LLM research team built a Hearthstone Battlegrounds simulator and trained a reinforcement learning agent that beats the top 10 ranked humans on the live ladder.
A new self-training framework called OPST lets language models improve themselves using only topic names, with no human-written questions, answers, or reasoning traces.
FMER swaps diffusion for flow matching in online RL, delivering tractable entropy, mirror descent updates, and up to 7x faster training.
SubQ Snap 1.1 ties d1-ultra and Elephant 9B atop the Decision Index at a score of 65, winning on hallucination control and context length.
Perplexity and NVIDIA introduce JOLT, a self-distillation post-training method that matches GRPO accuracy on MATH using 6.5x fewer training tokens.
A new continuous language model called LAW ditches discrete token sampling between steps, propagating expected embeddings instead and outperforming diffusion baselines on reasoning tasks.
Microsoft researchers used human-directed AI agents to prove a 31-year-old geometry conjecture at the heart of randomized sampling algorithms.
A new analysis settles whether momentum-based optimizers can get stuck at saddle points, and designs a variant that escapes faster without hurting convergence.
A looped transformer that reasons on a unit sphere halts earlier, uses less compute, and lifts Sudoku-Extreme accuracy to 96.6%.
QTrack turns multi-object tracking into a language-conditioned reasoning task, hitting state-of-the-art on a new RMOT26 benchmark with just 3B parameters.
A new reinforcement pre-training recipe pushes rewards into the hidden iterations of looped transformers, improving accuracy and cutting compute per token.
A 27B open model learns to generate scientific hypotheses recursively, matching real published research at nearly triple the rate of GPT-6 and Claude Opus 5.
Anthropic is paying PhDs up to $167K for four months to turn their domain expertise into training data and evaluations for Claude.
A new reinforcement learning reward treats chain-of-thought as gradient descent, cutting tokens by about 50% while boosting accuracy on math, science, and code benchmarks.
Epoch AI's new SpireClimb benchmark shows frontier models leapt from near-zero win rates to 90%+ on base difficulty in six months.
Google's new Geospatial Reasoning agent chains Earth AI foundation models with upfront planning, tackling planet-scale analysis that trips up standalone LLMs.
No stories match these filters.
