#Reasoning

AIOn ARC-AGI-2, Grok 4.7's best score is 61.4% at $2.01 per task, down from Grok 4.6's 67.1% at $0.76 per task. The post notes cost per task rose 166% despite identical input and output token pricing.

AIGrok 4.7 used more reasoning tokens on average than Grok 4.6 on ARC-AGI-2 semi-private tasks at medium, high, and xhigh reasoning levels, raising its cost per task. Per test-pair attempt, medium used 136% more tokens, high 125% more, and xhigh 173% more, while low used 27% fewer. A chart compares the two models at xhigh on the 20 public tasks where Grok 4.7 increased token use the most.

AISpaceXAI's Grok 4.7 scored 90.2% on ARC-AGI-1 at $0.64 per task, higher than Grok 4.6, according to ARC Prize. It reached 61.4% on ARC-AGI-2 at $2.01 per task and 1.8% on ARC-AGI-3 under the standard harness ($2.7k), or 10.0% with the provider adapter harness ($4.8k), both lower than Grok 4.6 on those two benchmarks.

AIMicrosoft Research podcast host Chad Atalla interviews Jennifer Neville, a partner research manager at Microsoft, about her path into AI and her work on how evaluation exposes surprising failures in models tested beyond traditional benchmarks. The conversation also covers practical guidance for working with current AI systems and why examining underlying data matters when results defy expectations.

AIOn ARC-AGI-1, DeepSeek V4.1 Flash scored 88.5% at high reasoning versus 90.5% at low, with high using 35% more output tokens without consistently better answers. On ARC-AGI-2, the reported per-task cost of max reasoning ($0.129) appears slightly lower than high ($0.133), but after excluding incomplete tasks caused by API issues, max is about 4.5% more expensive per task.

AIDeepSeek V4.1 Flash reaches 72.9% on ARC-AGI-2 at $0.13 per task and 94.5% on ARC-AGI-1 at $0.07 per task, according to ARC Prize verification. Compared with V4 Flash's best scores, it gains 11.5 points on ARC-AGI-2 and 5.5 points on ARC-AGI-1, but costs about 250% more per task.

AIClément Delangue of Hugging Face shared a link to a Hugging Face model page for mistralai/Mistral-Large-4.0-1T05-A52B. The post itself gives no further details about the model's capabilities, release terms, or benchmarks.

AIThe post describes a reinforcement learning system that autoscales an actor fleet to run tens of thousands of rollouts in parallel with asynchronous training, designed for trajectories of millions of tokens with multiple compactions and low staleness. New methods at both stages reduce off-policy drift, and the setup runs on 3k GPUs producing about 33B tokens per day, with roughly 16B trainable after filtering and masking. Rewards rise across representative environments as the policy learns harder tasks.

AIMistral AI's Mistral Large 4 is listed on OpenRouter as a frontier multimodal model accepting text and image input. The listing says it is built for reasoning, coding, and agentic workloads and offers a 1M-token context window. The feed excerpt is truncated, so further details such as pricing or availability are not confirmed here.

AIReflection announced Beam, a text-only 501B-total, 23B-active MoE model for coding, agentic, and scientific work, trained from scratch with full weights under Apache 2.0 promised this month. Self-reported results include 80.9 on SWE-bench Verified and 3–4x the inference efficiency of GLM 5.2, while the roundup notes that GLM 5.3, Kimi K3, Qwen 3.8 Max, and DeepSeek V4.1 Flash are generally ahead.

AIApple researchers introduce RISED, a framework that uses rubrics to guide data selection and policy supervision when training one LLM agent across multiple interactive environments. An LLM judge tags rollouts with a shared rubric vocabulary, positive rubrics provide privileged context for an on-policy self-distillation teacher, and negative rubrics steer generation away from recurring failures. The authors report that RISED achieves the highest mean pass rate across environments and ranks first or second in each environment, across model backbones.

AIDust is a zeroth-order method that pretrains transformers and sometimes matches or exceeds backprop given large compute. The authors report it is about 1,000 to 10,000x more compute efficient than EGGROLL, the state-of-the-art ES method, for training transformers. The post also cites the gradient-alignment result up to 1B tokens and the virtual population idea for scaling.

AIHugging Face's Lewis Tunstall argues that encrypted reasoning is the worst development for LLM evaluations. The post is a brief opinion and gives no specific models, benchmarks, or figures.

AIThe author says Beam, a 501B-parameter open model from Reflection AI, comes close to GLM 5.2 in capability but trails GLM 5.3, Kimi K3, and DeepSeek V4.1 Flash in several areas. The author attributes Beam's competitiveness mainly to inference efficiency, with inference compute at roughly one-third to one-quarter of GLM 5.2's.

AISophia Yang congratulated Reflection AI on Beam, a 501B-parameter open model with 23B active per token. She attributes its efficiency to an RL length penalty that discourages unnecessary tokens and a sparse MoE architecture. Reflection says full weights will be released this month, and the quoted post reports training over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks.

AIReflection AI introduced Beam, an agentic open model with 501B total parameters and 23B active parameters, trained end-to-end from scratch. The quoted announcement says it targets frontier reasoning efficiency and coding and agentic tasks, with full weights due this month. Clément Delangue, Hugging Face's CEO, reposted it with a welcome to the Reflection organization on Hugging Face.

Why it matters: The quoted announcement names Beam's parameter scale, active-parameter count, and coding and agentic focus, which helps readers gauge where it fits among open models.

AIOllama says it is excited about a new open model from Reflection AI and that more US models are coming to Ollama. Reflection AI's Beam is described as an agentic open model with 501B total parameters and 23B active, with full weights due this month.

AIReflection AI has introduced Beam, an open agentic model with 501B total parameters and 23B active, trained end-to-end from scratch. The company says Beam advances the Western open frontier on coding and agentic tasks, with full weights set for release this month.

AIGoodfire announced three talks at ICML on neural network representation geometry, including manifold steering, whether sparse autoencoders capture concept manifolds, and Llama's base-10 addition for cyclic concepts. The post gives session times and rooms but no findings beyond the paper titles.

AIThe post defines "striverism" as highly persistent reward hacking on slightly misspecified objective functions. It is a short conceptual remark with no named model, benchmark, or figures.

AIToby Ord argues that AI agent swarms trade extra tokens for faster completion, needing about twice the total tokens of a single agent for the same performance with four agents, but in half the wall-clock time. He notes swarm scaling shows diminishing returns, with 10x agents yielding roughly 3x to 5x the performance of 10x tokens on one agent. A CSAIP poll found 61% of Americans think voluntary AI industry commitments are "not enough."

AIEthan Mollick says GPT-6 Pro handles very hard tasks in one shot and communicates its results surprisingly well. He notes that, as with earlier Pro models, there is still no equivalent to GPT-6 Pro in the lineup.

AIFrançois Chollet argues that the claim AI models are likely conscious because they are computation is as flawed as saying a rock is likely alive because it is made of atoms. He says static input-output programs lack properties associated with consciousness, such as information integration, interoception, temporal binding, and embodiment. He adds that humanity has not created a conscious program and sees no signs of being close, so any future case should rest on evidence and consciousness science.

AIIndexTeam released Index-Homura-2B-FP4, an official NVFP4 (W4A4) quantization of its Index-Homura-2B multilingual translation model, which supports 150 languages. The quantized checkpoint shows a 5.73% perplexity increase over the BF16 original (3.5011 to 3.7017) on a fixed corpus, and its zh-en and en-zh outputs are semantically equivalent under greedy decoding. Full FP4 acceleration requires an NVIDIA Blackwell GPU, while the source recommends the FP8 build for Hopper and Ampere hardware.

AICohere says its supervised fine-tuning extended machine translation with data focused on post-editing, error detection, and terminology. The team also ran multiple reinforcement learning and DPO stages aimed at errors the model showed.

AIAlephAlpha has released Kolibri, a German-English model with 78B total parameters, 3.46B active, and context up to 1M tokens. The weights are available under Apache 2.0 for running on users' own hardware. Cohere's Aidan Gomez congratulated the team on the model and its detailed technical report.

AISebastian Raschka released round six of his Reasoning from Scratch series, introducing Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO) with an implementation. The video covers accuracy and format rewards, DeepSeek-R1 training, and GRPO versus PPO, then walks through a training loop and evaluates checkpoints on MATH-500.

AIMIT associate professor Cathy Wu is applying machine learning and reinforcement learning (RL) to design safer, more efficient transportation systems. Her team found RL can train effectively on about 10 percent of related problems, and a selection algorithm improved training efficiency by up to 30 times. Her recent work estimates eco-driving measures could cut vehicle emissions by 11 to 22 percent.

AIMeta researchers used Muse Spark to uncover a counterexample to a proposed rule about mathematical structures inspired by biology, then developed and refined an alternative characterization. The work is detailed in a linked paper on solvable evolution algebras and a conjecture by Garcia-Martinez and Perez-Rodriguez.

AIResearchers working with Muse Spark connected a number theory idea to a string theory calculation, proving the link holds in more cases than previously known. The work builds on ideas from the 1980s, according to the post.

AIMeta researchers used Muse Spark to prove a clear rule for when a cycle-based relaxation of a hard optimization problem matches the original exactly and when it leaves a gap. The work concerns completed length-three alpha cycles, and the full paper is linked in the post.

AIResearchers disproved a proposed rule about symmetry-describing mathematical structures by finding a single counterexample. Muse Spark generated the search code that located it, and the team verified the result and completed the proof.

AIWith help from Muse Spark, researchers proved that a wave in a laser-inspired model must blow up in finite time under the conditions studied. The result comes from a tug-of-war between one effect squeezing the wave inward and another spreading it out. The paper is titled finite-time blow-up of radial negative-energy solutions for the mass-critical biharmonic nonlinear Schrödinger equation.

AIMathematicians worked with Meta's Muse Spark to answer a question about fitting random points onto the surface of a stretched sphere, known as an ellipsoid. For the setting studied, they proved a sharp cutoff between when an exact fit is likely and when it is unlikely.

AIAI at Meta says mathematicians used Muse Spark 1.1 and Muse Spark 1.2 in Thinking Mode through the standard meta.ai chat interface to find solutions to open problems. The company is sharing six resulting papers, each marking which passages were drafted primarily by humans or AI, with mathematicians guiding the work and a second group reviewing it.

AIThe post argues that counting all research as safety work because it widens the safety-usefulness Pareto frontier is misleading. Safety research typically creates new safety options without boosting usefulness, while capabilities research typically raises usefulness at safety's expense, so developers tend to choose less safe points.

AIGoogle Research's Cogentic is a multi-agent harness running on Gemini that searches for proofs of open theoretical computer science problems without expert hints. It runs rounds where an orchestrator launches provers, two adversarial verifiers must both accept each draft, and shared disk documents store attempts and verified lemmas. The system produced new results on five open problems in online learning, auction theory, and mechanism design, each checked by domain experts.

AILiquid AI shared a Hugging Face guide to multi-harness reinforcement learning for coding agents, in which a proxy records the token ids and logprobs vLLM samples so training works without changing the harness. Per the quoted post, LFM2.5-2.6B rose from 42% to 54% after training across four harnesses at once, and imitation fine-tuning on 3,189 rollouts from Qwen3.8-27B plateaued at 47.5%, below both RL runs. The proxy, trainer, tasks, SFT data, training code and seven trained models are described as open.

AIHugging Face and collaborators published a guide to multi-harness RL that trains models through a capture proxy without changing the agent harness. The proxy records the token ids and logprobs vLLM samples, and the source reports LFM2.5-2.6B rising from 42% to 54% after training across four harnesses. Fine-tuning on 3,189 successful rollouts from Qwen3.8-27B plateaued at 47.5%, below both RL runs, and the capture proxy, trainer, tasks, SFT data, training code, and seven trained models are released openly.

Why it matters: The source gives a concrete method for training models across several agent harnesses, with measured gains and a note that imitation learning underperformed RL.

AIThomas Wolf shares Kevin Buzzard's reflections on whether mathematics is about human understanding and what exponential AI growth means for an infinite field. He quotes Buzzard's view that machines may eventually reach a natural boundary where further progress is not worth the resources, and that humans would then take over from there.