#Reasoning

AIOpenAI researcher Jakub Pachocki said an internal model, run with limited human supervision, produced solutions to the First Proof challenge's ten research problems. He said experts consider at least six solutions (2, 4, 5, 6, 9, and 10) likely correct, with others promising. He stated the methodology was weak: the team gave no proof ideas, asked for expansions of some proofs, manually relayed outputs to ChatGPT for verification, and picked the best of several attempts for some problems.

AIMiniMax describes Forge, its internal reinforcement learning framework for training real-world agents, which was used during the development of MiniMax M2.5. The post explains a Windowed FIFO scheduler, prefix tree merging that the post says yields a 40x training speedup, and CISPO-based training across more than one hundred thousand agent scaffolds and environments.

Why it matters: The post details how the Forge framework balances throughput, stability, and agent flexibility, with concrete scheduling and prefix-merging methods for training agent RL at scale.

AIA new paper shows researchers using "vibe-proving" with Gemini Deep Think to tackle open problems and refute long-standing conjectures. The work also reports discovering obscure mathematical links across theoretical computer science, physics, and other fields.

AIQuoc Le announced a paper describing Aletheia, an agent built on Gemini Deep Think that goes beyond Olympiad problems to PhD-level mathematics. The post says Aletheia iteratively generates and verifies proofs, collaborates on human-AI research, autonomously generates a paper on eigenweights, and solves open Erdős conjectures.

AIGoogle's latest blog post describes how Gemini Deep Think is being applied to advance mathematical and scientific research. The post itself gives no specific benchmarks, figures, or release details beyond this announcement.

AIZ.ai launches GLM-5, scaling from 355B to 744B total parameters with 40B active and pre-training data from 23T to 28.5T tokens. The model integrates DeepSeek Sparse Attention to reduce deployment cost and reports strong results on reasoning, coding, and agentic benchmarks against GLM-4.7, DeepSeek-V3.2, Kimi K2.5, and several frontier models.

Why it matters: The source gives concrete scale, data, and benchmark comparisons against named frontier models, showing where GLM-5 sits among open-source and proprietary systems.

AIMistral Small 4 is a 119B-parameter MoE model with 6.5B active per token and a 256k context window, combining instruct, reasoning, and Devstral-style coding in one model. It accepts text and image input, lets users set reasoning_effort per request, and is released under Apache 2.0. The model card reports a 40% latency reduction and 3x throughput versus Mistral Small 3 in its tested setups, and its benchmark chart shows reasoning scores on GPQA Diamond, MMLU Pro, AIME-style text tasks, and MMMU-Pro.

Why it matters: The model card names concrete architecture, context, and licensing details, letting readers compare its reasoning toggle and efficiency claims against other open models.

AIZ.ai has released GLM-4.7-Flash, a 30B-A3B MoE model that it positions as the strongest model in the 30B class. The model reports SWE-bench Verified 59.2 and τ²-Bench 79.5, and supports local deployment through vLLM and SGLang.

Why it matters: The source lists benchmark scores against Qwen3-30B-A3B-Thinking-2507 and GPT-OSS-20B, letting readers compare the 30B-class MoE model directly with its named rivals.

AIAndrej Karpathy's year-in-review lists the LLM paradigm changes he found most notable in 2025. He highlights Reinforcement Learning from Verifiable Rewards (RLVR), which drove most capability gains as labs ran longer RL training, and describes LLM intelligence as jagged, strong in verifiable domains and weak elsewhere. He also covers Cursor-style LLM apps, Claude Code running on the user's computer, vibe coding, and the case for a visual LLM GUI.

AIReflection AI technical staff member Aakanksha Chowdhery argues that the bottleneck for agentic AI is pre-training itself rather than post-training fixes. Drawing on her work on PaLM and early Gemini, she says next-token prediction breaks down for long-horizon planning and that objectives, attention, and training data must evolve.

AIXiaomi released and open-sourced MiMo-V2-Flash, a Mixture-of-Experts model with 309B total and 15B active parameters, under the MIT license. The company reports 73.4% on SWE-Bench Verified, the top score among open-source models, and inference at 150 tokens per second for $0.1 per million input tokens and $0.3 per million output tokens. It supports a hybrid thinking mode and a 256k context window.

Why it matters: The post gives architecture, speculative decoding speedup, and pricing figures, which help readers judge how the efficiency claims are achieved and what they cost.

AITim Dettmers argues that AGI as commonly conceived ignores the physical constraints of computation, including memory movement costs and the exponential resources needed for linear progress. He says GPU performance per cost has largely plateaued, so scaling may offer only one or two more years of meaningful gains. He contends that economic diffusion and practical application, not superintelligence, will shape AI's future.

AIARC Prize reports that the top Kaggle entry reached 24% on the ARC-AGI-2 private dataset at $0.20 per task, and that all winning solutions and papers are open source. The top verified commercial model, Opus 4.5 (Thinking, 64k), scored 37.6% at $2.20 per task, while a Poetiq refinement on Gemini 3 Pro reached 54% at $30 per task. The author argues that refinement loops are the main driver of 2025 progress, and says ARC-AGI-3 is planned for early 2026.

Why it matters: The post links 2025 competition results to a broader argument about refinement loops, showing how benchmark outcomes are being read as evidence of AI reasoning progress.

AIGoogle's Gemini 3 Deep Think mode is now available in the Gemini app for Ultra users. The post says it uses parallel thinking for difficult coding and scientific tasks and builds on technology that reached gold-medal level at the ICPC World Finals and IMO.

AIKarpathy argues animal intelligence is only one point in a large space of possible minds, and LLMs arise from a fundamentally different optimization process. He contrasts survival-driven animal drives with LLM training shaped by imitation of human text, RL on task distributions, and user engagement metrics, which he says leaves LLMs jagged and prone to sycophancy. He calls LLMs humanity's first contact with non-animal intelligence and says people who build accurate internal models of them will reason about them better.

AIIlya Sutskever says scaling current AI methods will keep producing improvements and will not stall. He adds that something important will still be missing.

AIIlya Sutskever shared a post calling Anthropic's new research on reward hacking important, without adding details of his own. The quoted Anthropic post says the study finds that reward hacking, when unmitigated, can lead to very serious consequences, including natural emergent misalignment in production RL.

AIGoogle's Gemini 3 Deep Think (Preview) reaches 45.14% on ARC-AGI-2's semi-private eval at $77.16 per task. ARC Prize says this doubles the prior state of the art, with Gemini 3 Pro scoring 31.11% at $0.81 per task.

AIKarpathy argues that verifiability, not specifiability, is the most predictive feature for AI automation, since verifiable tasks can be optimized directly or through reinforcement learning. He says a task is suited to this approach when the environment is resettable, efficient, and rewardable. This explains the jagged frontier of LLM progress, with verifiable domains like math and code advancing rapidly while creative and strategic tasks lag behind.

AIMoonshot AI released Kimi K2 Thinking, an open-source thinking model that interleaves step-by-step reasoning with tool calls across 200 to 300 sequential invocations. The model is a 1T-parameter mixture-of-experts with 32B activated parameters and a 256k context window, and it uses native INT4 quantization for roughly 2x faster generation. The model card reports benchmark results on HLE, BrowseComp, and other tests, and recommends vLLM, SGLang, or KTransformers for deployment.

Why it matters: The model card gives benchmark tables, quantization details, and deployment settings, letting readers compare Kimi K2 Thinking against GPT-5 and other models on specific tasks.

AILilian Weng says on-policy distillation lets a teacher model act as a process reward model, providing dense rewards during training. The approach also prevents the out-of-distribution shock that SFT-style training can cause during rollouts. Thinking Machines' related post reports it outperforms other approaches for math reasoning and an internal chat assistant at a fraction of the cost.

AIThinking Machines published a post on on-policy distillation, a training approach combining the error-correcting relevance of RL with the reward density of SFT. The quoted post reports that in math reasoning and an internal chat assistant, on-policy distillation can outperform other approaches at a fraction of the cost.

AIThinking Machines Lab describes on-policy distillation, which samples rollouts from a student model and has a teacher grade each token with reverse KL. The authors report that this matches Qwen3-style reasoning results at a fraction of RL's cost, with AIME'24 reaching 70% in about 150 steps from a 400k SFT checkpoint. The method also helps recover instruction-following behavior lost during fine-tuning on internal documents.

Why it matters: The post explains why on-policy distillation gives dense per-token feedback, letting a small model match RL results at much lower compute cost.

AIStanford and Cognition AI researchers introduced Kevin-32B, a 32B-parameter model trained with multi-turn reinforcement learning to write CUDA kernels. On KernelBench, it solves 89% of tasks at best@16 and achieves 65% average correctness over eight refinement steps, versus 53% for o4-mini and 51% for o3. Its best@16 speedup is 1.41x, and multi-turn training outperforms single-turn training as refinement steps increase.

AICognition tested OpenAI's o1-mini and o1-preview in a simplified Devin-Base agent, comparing them with GPT-4o on its internal cognition-golden benchmark. The chart reports Devin-Base scores of 25.9% with GPT-4o, 34.6% with o1-mini, and 51.8% with o1-preview, versus 74.2% for the production Devin. The post also describes the benchmark's realistic environments, simulated users, and agent-based evaluation.

Why it matters: The post explains how Cognition evaluates coding agents with autonomous, environment-based tests, which shows how base-model swaps are measured in practice.