#Reasoning

AIKorean AI lab Upstage has released Solar Mini 4, a proprietary reasoning model that scores 24 on the Artificial Analysis Intelligence Index with 35B total and 3B active parameters. It is priced at $0.10/$0.40 per 1M input/output tokens and has a 1M-token context window, but averages 7.1 minutes per task due to heavy output token use. Its weights are not released, and its size cannot be independently verified.

AIArtificial Analysis reports that Google's Gemini 4 Argon scores 53 on its Intelligence Index with high reasoning, matching GPT-6 Astra (max) and one point ahead of GPT-6.1 Sol (max). At the current 50% launch discount, its cost per task is $1.99, about 60% of GPT-6 Astra's $3.26, but the discount's end date is unconfirmed and standard pricing would raise it to $3.98. The model is being rolled out to selected users and is not publicly available.

Why it matters: The benchmark compares Gemini 4 Argon's cost per task and hallucination rate with GPT-6 Astra, showing where its value depends on a temporary 50% discount.

AINumerical differences between a rollout engine and a trainer can make reinforcement learning collapse even when algorithm and data stay identical. In a GLM 5.2 experiment, reward fell from about 0.9 to under 0.2 around step 20 without alignment, while aligned numerics kept reward stable over 25 steps. A Qwen3.5-MoE investigation traced a significant mismatch to how expert outputs were combined, and router replay alone was judged insufficient.

AIFrançois Chollet suggests a meta-benchmark for human-level generality: an AI system should be able to pass ARC-AGI-(n+1) immediately upon its release. A quoted reply from @dylantmoore adds that such a test would presumably not hold up forever.

AIBAAI released AREX-2, a 27B-parameter long-horizon agent model that improves solutions over multiple test-time rounds by proposing, measuring, reflecting, and revising. It was trained on machine-learning and algorithmic-programming tasks with verifiable feedback, and the source reports that this self-improvement transfers to deep research. The model is Apache License 2.0 licensed and has a 262,144-token context length.

Why it matters: The source compares AREX-2 against closed and open models on coding and deep-research benchmarks, showing how test-time self-improvement is measured across task types.

AIInternLM's AutoVerifier, built on Qwen3_5MoeForConditionalGeneration with about 68 GiB of weights across 40 safetensors shards, evaluates natural-language mathematical proofs, explains errors, and identifies the earliest incorrect step. It serves as the automatic grader for AdvancedMathBench's ProverBench, which accepts a proof only when all eight judgments report -1. The model is a learned grader rather than a formal proof checker and can make errors.

AIGoogle says Gemini 3.8 Flash, its most intelligent workhorse model, improves on 3.7 Flash in software engineering, agentic tasks, and multistep reasoning by running extra reasoning steps and calling tools iteratively. The post highlights four community builds, including a model rocket simulation, an animated ink-painting effect, a 3D dinosaur skeleton, and an interactive automatic transmission simulation. Developers can try the model through Google Antigravity and Google AI Studio.

AIAlexander Doria argues that embedding AI models in industrial processes requires easing audits and supporting local languages and professional registers. He says this need is real despite sarcasm about the idea. Background from Aleph Alpha notes that reasoning models think in English even for German prompts, and that small German reasoning data doses can hurt performance while large doses mostly recover it.

AIXiaomi MiMo reports that tool-call repetition in MiMo-V2.6 reached over 0.05% of responses across agent harnesses, causing stalled agents and wasted context. The team traced the cause to an RL flooding penalty set at 32 calls per turn, which missed smaller excess behavior, and replaced the approach with a specialized teacher distilled via MOPD. Repetition rates for both Pro and Flash dropped substantially, at roughly $90,000 versus an estimated $2.31 million for the alternative fix.

Why it matters: The post traces an agent failure to a reward blind spot and compares the costs of two fixes, offering a transferable debugging method for RL-trained tool-calling models.

AIXiaomi released an open-source RL environments repository, MiMo-V2.6-RL-oss, on Hugging Face. Hugging Face CEO Clément Delangue promoted the release, and a commenter estimated that comparable commercially sold tasks cost hundreds to thousands of dollars each.

AIAlexander Doria reports that the released set is a smaller selection of 989 environments for RL training a 9B distilled model, not the full MiMo. Rewards are not self-contained: the general part requires setting up a judge, and webdev relies on its own grader service and VLM. The most important content is in the general/envs directory and Docker setup rather than the Hugging Face dataset, offering a solid mix of real and simulated documents.

AIAnthropic reports that Claude solved a nine-loop calculation in the planar N=4 super-Yang-Mills model, surpassing the previous eight-loop record set by Lance Dixon and collaborators. The quoted post says Claude ran largely unsupervised for days in Claude Science using a single prompt, at a total cost of a few thousand dollars, and Dixon independently verified the result. Kevin Weil's own text praises the achievement and expects AI to advance high energy physics over the coming 12 months.

Why it matters: The quoted Anthropic post gives a concrete benchmark: Claude ran for days to reach nine loops, extending the previous eight-loop record in a physics model.

AIAnthropic reports that Claude solved a nine-loop scattering amplitude problem in planar N=4 super-Yang-Mills, surpassing the previous eight-loop record set by SLAC's Lance Dixon and collaborators. Working largely unsupervised for days from a single prompt, at a total cost of a few thousand dollars, Claude used methods developed by Dixon's group, and Dixon independently verified the result.

Why it matters: The post shows Claude solving a nine-loop physics calculation beyond the previous eight-loop record, verified independently, which bears on AI use in theoretical physics research.

AIGoodfire Research introduces Block-Sparse Featurizers (BSF), which decompose model activations into subspaces rather than single directions. Applied to DINOv3 and Stable Diffusion XL, BSFs find interpretable multidimensional features that better explain activations and enable fine-grained steering. The authors report that most concepts they examined have a stable rank of about two to four dimensions.

AIGoodfire researchers found that Llama 3.1 8B solves month and day arithmetic, such as six months after August, through a single addition module in layer 18. The module represents numbers as Fourier-feature circles and computes modular sums in parallel, and steering those circles changed the model's predicted month.

AIGoodfire Research shows that steering Llama-3.1 8B along the curved representation manifold of weekdays produces output probabilities that follow the model's natural cyclic behavior, shifting probability mass smoothly from Monday to Tuesday to Friday. Linear steering along a straight vector, by contrast, cuts across the behavior manifold and yields noisy off-target tokens, some not days of the week at all. The authors argue that representation geometry and behavior geometry are linked bidirectionally.

AIGoodfire Research examines how language models track emotional dynamics across a story, sentence by sentence. The authors prompt Llama 3.1 8B to rate six emotions after each sentence, then harvest activations from each sentence's last token and fit a manifold to show stories tracing trajectories through it.

AIGoodfire Research examines how sparse autoencoder directions relate to curved manifolds in neural representations, identifying shattering, compact capture, and dilution as three ways lines can represent them. The team trained an autoencoder on synthetic data containing shapes such as donuts, spheres, and Möbius strips, and reports that real features in Llama 3.1 8B show dilution. It also describes an unsupervised pipeline that clusters features by firing patterns to surface manifolds in that model.

AIIn a wide-ranging interview at The Information AI Summit, Koray Kavukcuoglu said Gemini 4 is coming soon and that Google is "certainly at the frontier." He argued against slowing development, saying safety is built in lockstep with progress, and he pointed to Google's continued hardware advantage with TPUs.

AITencent Hunyuan extends classical critical-batch-size theory to online LLM reinforcement learning, where models generate their own training data. Across GRPO and PPO, learning-rate retuning preserves learning per response over a bounded range of batch sizes. On fixed hardware, larger batches raise PPO generation-stage throughput by up to 2.29×, and the best measured GRPO setup reaches the same validation target in 29% less time.

AIRedwood Research argues that latent reasoning architectures such as COCONUT and full-bandwidth transformers could let models reason without putting information into readable chain-of-thought. The authors say this would make AI agent behavior harder for humans to monitor and could raise takeover risk. They argue that developers who adopt such architectures should be transparent about it.

AIAnthropic's Cat Wu points users to ways to try Opus 5.5 and Claude Tag in Slack. The main post carries no further details on access, features, or availability.

AIRedwood Research found that padding prompts with meaningless filler tokens improves GPT-6-Astra's no-reasoning answers on serial reasoning tasks, rising from about 10-20% to about 50% on 4-hop natural facts. Other tested models improved far less, and the authors argue this means Astra can perform cognition it does not verbalize in its chain of thought, making such monitoring harder.

AIFireworks Research released Ember-1, a specialized model built on Kimi K3 that it says delivers the same quality with 40% fewer tokens. Across five industry benchmarks, Ember-1 matched K3 max quality at a fraction of the cost, and in two customer A/B tests it used about 35% fewer tokens per task. It is available as a Research Preview on Serverless, and Fireworks is also launching training support for customized models.

Why it matters: The source gives benchmark and A/B results for cutting reasoning tokens while holding quality, which bears on cost planning for coding and agent workloads.

AIJS Denain of Epoch AI discusses recursive self-improvement, arguing public evidence does not yet show a software intelligence explosion, though OpenAI's reported 2X monthly growth in researchers' Codex spending suggests substantial value. He also addresses the US-China AI gap, distillation, and whether open or closed models are safer. The episode, hosted by Nathan Lambert, expresses significant uncertainty about the trajectory of AI progress.

AIFrançois Chollet described summer 2026 as a crazy period for AI in a brief post with no further specifics. The post is linked to ARC Prize 2026's ARC-AGI-3 Progress Prize, where $37,500 in prizes will go to top open-source solutions on September 30, with Tufa Labs, Lord Han Solo, and NVARC3 currently leading.