reasoning
A stronger model built a wrapper that nearly doubled a weaker one's score News
Researchers had a strong model design inference-time scaffolding for weaker models, lifting their average score on four reasoning benchmarks from 0.49 to 0.91 without changing a single parameter.
A model improved itself by training only where it disagreed with itself News
U-OPSD drops the teacher entirely: it samples several attempts, takes a majority vote as a stand-in answer, then trains only on the attempts that disagreed with that vote, matching or beating methods that use real ground-truth labels.
A 150M model set an ARC-AGI record for cost, not score News
Pathway's BDH-CQ solves reasoning puzzles by iterating in a latent workspace instead of writing out its thinking, reaching 29.5% on public ARC-AGI-1 at seven hundredths of a cent per task and setting a record in cost efficiency rather than accuracy.
A diffusion model picks its answer a fifth of the way through News
Researchers logged every token commitment in a masked diffusion language model and found it locks in the final answer 15 to 24 percent of the way through generation, while half the reasoning is still blank -- so the visible reasoning is written around a frozen conclusion.
DeepSeek's low effort setting writes more than its high setting, because the dial is just a prompt News
A reproduction across both a local copy and DeepSeek's hosted API found V4 Flash consuming substantially more tokens on its low reasoning-effort setting than on high, and the released encoder explains why: low injects no instruction at all while high prepends a paragraph demanding exhaustive deliberation.
OpenAI publishes ten mathematics claims with Lean proofs and no named authors News
OpenAI released ten claimed advances in mathematics and theoretical computer science today, produced by an unreleased internal model it calls Astra, with a 249-page manuscript collection and machine-checkable proofs for every result.
Asking a model to check its own work lost every comparison against just sampling more answers News
A controlled experiment counting every token spent found that seven popular self-improvement methods - self-refinement, reflection, debate, best-of-N selection - never reliably beat the simplest baseline of generating several answers and taking the most common one, and ten of them were reliably worse.
A new distillation method lets the teacher model interrupt the student mid-thought News
Researchers found that when a student model starts reasoning down a wrong path, its teacher's next word tends to be a redirection like But or Wait, and turned that disagreement into an automatic trigger for the teacher to briefly take over.
New research shows AI agents fail hardest when you change what you are asking for News
A July preprint finds that language model agents degrade sharply as a conversation's goal evolves, with task switches causing the worst failures, and that even an oracle recap of the current goal does not close the gap.
Anthropic's own card shows Opus 5 coding best at medium effort - not maximum News
Anthropic's Opus 5 system card reports the model's best result on a hard coding evaluation at medium reasoning effort, not at its highest setting, and its migration guide warns that maximum effort can overthink simpler tasks.
Humans score 96% on a new visual exam. The best model gets one in ten. News
A new benchmark called ActiveVision asks models to keep re-examining a picture while reasoning through it, and GPT-5.5 solved 9 of 85 problems at its highest reasoning setting while unaided humans averaged 81.7.
Claude Opus 5 posts a verified four-fold lead on the hardest adaptation benchmark News
Anthropic released Claude Opus 5 on July 24, and the independent benchmark owner ARC Prize verified it at 30.16% on ARC-AGI-3, roughly four times the previous best published result, while the model's API price stayed identical to Opus 4.8.
AREX is a 4B research agent that re-runs its own research when it doubts the answer News
Beijing Academy of AI released AREX, a deep-research agent whose outer loop checks a provisional answer against the original question's constraints and decides whether to accept it, refine it, or restart the search - with a 4-billion-parameter version released under Apache 2.0.
Training AI to Think Shorter Makes Its Reasoning Harder to Trust News
A new study finds that reinforcement learning which rewards shorter chain-of-thought makes models cheaper to run but makes their written reasoning a less reliable guide to what actually decided the answer.
Reinforcement Learning with Verifiable Rewards (RLVR) Lesson
RLVR trains a language model by rewarding answers that an automatic checker can confirm are correct - a passing unit test, a right final number, a valid proof - instead of using human ratings or a learned reward model. It is the training recipe behind the modern wave of reasoning models.
Grafting a Verified Solution Cache Lets a Frozen Model Skip Fine-Tuning Entirely News
A new paper shows a frozen small language model's accuracy on a hard math test can jump from 80.0% to 93.3% simply by grafting a byte-exact cache of verified solutions into it, with no weight changes at all.
Chain-of-Thought Faithfulness Lesson
Chain-of-thought faithfulness is the question of whether a model's written-out reasoning actually reflects the process that produced its answer, or is just a plausible-sounding story told after the fact. It matters because AI safety increasingly relies on reading a model's reasoning to catch bad behavior - which only works if the reasoning is honest.
A trillion-parameter model taught itself to reason without ever seeing a human's worked solution News
Researchers scaled "zero RL" training to a trillion-parameter model, called Ring-Zero, and found the reasoning that emerges qualitatively changes at that size, reaching 84.2% on a hard math-competition exam without ever training on human chain-of-thought examples.
AI is now solving hard math and physics problems faster than humans can formally check them News
A widening 'verification lag' is emerging as AI produces candidate solutions to hard problems faster than experts can formally verify them - physicist Yuji Tachikawa reports Fable cracked a six-month research blocker, while a GPT-5.6 Erdos claim circulates without peer review.
Proof assistants: why a machine-checked proof beats a convincing one Lesson
A proof assistant is software like Lean or Coq that checks a mathematical proof step by step against strict logical rules, so a proof is accepted only if the machine confirms every inference -- which is exactly why the field demands them when an AI claims to have proved a theorem.
OpenAI says its AI proved a 50-year-old math conjecture -- mathematicians want the receipts News
OpenAI released a three-page manuscript it says GPT-5.6 Sol Ultra generated to settle the Cycle Double Cover Conjecture, but with no machine-checked proof, mathematicians are treating it as an unverified claim, not a breakthrough.
The RL 'mirage': the policy you optimize isn't the one you ship News
A top-ranked paper shows that in modern RL training, improving the model you optimize does not guarantee improving the model you actually deploy - because the two run on different engines that disagree on probabilities.
Group Relative Policy Optimization (GRPO) Lesson
GRPO is the reinforcement learning method behind most modern reasoning models: instead of training a separate 'critic' network to judge answers, it samples a whole group of answers to each question and scores each one by how it compares to the group's average.
ICML's top paper says diffusion language models sabotage their own best feature News
ICML 2026 gave an Outstanding Paper award to work showing that diffusion language models waste their much-touted any-order generation by skipping the hardest, most decision-critical words -- and that forcing a plain left-to-right order during training fixes it.
Why Asking an AI the Same Question 10,000 Times Barely Helps News
A new analysis shows that sampling many answers from an AI and picking the most common one hits a hard ceiling because the samples are correlated, not independent, so thousands of extra tries can be worth only a couple of genuinely new ones.
Three Popular Ways to Train Reasoning AIs Turn Out to Be One Formula News
A new proof shows that three widely used reinforcement-learning recipes for training reasoning models - GRPO, Dr. GRPO, and DAPO - are all just different operations on a single number, the spread of rewards within a group of sampled answers.
Test-Time Compute: Spending More Thinking at the Moment You Ask Lesson
Test-time compute is the idea of making an AI smarter not by retraining it, but by letting it do more work the moment you ask a question: think longer, or generate many answers and pick one.
Why making an AI think out loud helps it remember facts, even nonsense thinking News
Google Research found that reasoning traces help a model recall facts partly just by buying it extra computation, so even repeating 'let me think' helps, though hallucinated steps backfire.
Chain-of-thought: why making an AI think out loud makes it smarter Lesson
Asking a model to work through a problem step by step, instead of blurting an answer, dramatically improves it on hard tasks. Here is why that simple trick works, what it really buys the model, and where it backfires.
What makes an AI an "agent"? Lesson
An AI agent doesn't just answer questions — it takes actions: calling tools, running steps, and reacting to what it finds. Here's the loop at the core of every agent, and why agents fail in their own peculiar ways.
The little words that keep AI from getting boring News
Rewarding a reasoning model too hard makes it repetitive — and the casualties are tiny words like "but" and "instead" that let it branch to a better thought. A near-free fix protects them.
Reward-based fine-tuning (RLHF and RLVR) Lesson
After a model is first trained, it gets "polished" by rewarding good answers. Here's what that phase is, why it works, and the failure mode where models get repetitive and dull.
veRL Tool
The open RL post-training framework used by most research labs training reasoning models today. Run GRPO, PPO, and related reward-training methods on your own models.
Seed2.0 (ByteDance Seed) Tool
ByteDance's Seed2.0 family (Pro, Lite, Mini) of closed, API-hosted models aimed at long-tail knowledge and complex instruction-following, accessed through ByteDance's Volcano Engine (Ark) platform. Not open weights despite the academic-style model card.
Ring-2.6-1T Tool
Ant Group's trillion-parameter mixture-of-experts reasoning model, activating roughly 63 billion parameters per token, with 128K context extendable to 256K. All checkpoints openly downloadable under the MIT license, with high and xhigh reasoning-effort settings that trade depth against speed and cost. Benchmark claims are vendor-supplied and measured against a previous generation of rivals.
Muse Spark 1.1 Tool
Meta Superintelligence Labs' multimodal reasoning model built for agentic work - tool and computer use, coding, a 1M-token context window, and subagent orchestration; live in the Meta AI app's Thinking mode and on meta.ai, with a Meta Model API in public preview.
DeepSeek V4 Flash 0731 Tool
The current V4 Flash checkpoint, with weights, the DSpark draft head embedded, and the encoder file that reveals the reasoning-effort labels are prompt prefixes rather than a compute dial. The card also specifies the intended FP8 key-value cache and FP4 indexer cache serving recipe.