reasoning

"The model can actively extend the thought process when in need of deep reasoning, and return results quickly in simple questions."

"Phi-4-Reasoning-Vision-15B is trained with Supervised Fine-Tuning (SFT) on a carefully curated mixture of reasoning and non-reasoning data."

"GPT-5.4 updates (via TheInformation)"

"OpenAI is preparing to launch GPT-5.4, which will feature an "extreme" reasoning mode and a context window of 1M tokens, up from 400K in GPT-5.2"

"The core finding is that even if a user writes 'must be followed', it is treated as one of the 'strong context' internally."

"This work develops a dynamical theory of sequential reasoning in Hopfield networks."

"ActMem transforms unstructured dialogue history into a structured causal and semantic graph."

"This dual-verification method enables the scalable assessment of "true" reasoning capabilities."

"The paper introduces a unified framework integrating: Inference, Action, and Planning."

"To answer this question, we first evaluate pretrained models with and without Chain-of-Thought (CoT) reasoning, and show that CoT reasoning has the potential to generate at least…"

"84.0% on ARC-AGI2 (840/1000) using LLM program synthesis + deterministic verification — no fine-tuning, no neural search"

"If a model is given a 'way of seeing'—that is, an axiomatic system for structurally grasping problems—then the quality of the Inference itself changes."

"The most interesting point of this experiment is the fact that the 'accuracy rate without thought' and the 'number of model parameters (brain size)' draw an extremely clean logarithmic linear graph (log-linear)."

"Now I have released WFGY 3.0, which is no longer “just RAG”. It is a TXT-based tension reasoning engine designed to stress-test strong LLMs on problems that look a lot closer to real world fracture lines."

"Yao et al.'s paper (NeurIPS 2023) shows that applying ToT to GPT-4 improved the success rate on the Game of 24 task from 4% (with Chain of Thought) to 74%."

"India needs to own the scoreboard, not just the model."

"Gemini 3.1 Pro's greatest strength lies in the combination of a vast context window and advanced abstract thinking."

"Mercury 2 doesn't decode sequentially. It generates responses through parallel refinement, producing multiple tokens simultaneously and converging over a small number of steps."

"Evaluation across diverse workloads shows that ConfSpec achieves up to 2.24$ imes$ end-to-end speedups while matching target-model accuracy."

"In this work, we perform an in-depth analysis of CoT traces originating from competition-level mathematics questions, with the aim of better understanding how, and which parts of CoT actually contribute to the final answer."

"The correct answer is “NO”"

"We evaluate 10 frontier models on TTG, and closely match the ranking from existing benchmarks such as Humanity's Last Exam, without involving any human effort in creating puzzles."

"For example scoring 77.1% on the ARC-AGI-2 benchmark - more than 2x the performance of 3 Pro."

"In early Azimov stories about robots, the robots who could not speak were less advanced than the later versions who could speak."

"真の問いは「AIが単体で何をできるか」ではなく「Human-AI結合系が何を実現するか」だ"

"In some cases, responses appear to adopt constraint-based decomposition (e.g., outcome modeling through component interaction, optimization under evaluative metrics), even when such structure is not explicitly requested by the prompt."

"Reasoning level changes primarily adjust the allocation of computational resources during inference (Test-time Compute)."

"Vision-Language Models achieve ~84% F1 reading binary grids rendered as text characters (. and #) but collapse to 29-39% F1 when the exact same grids are rendered as filled squares, despite both being images through the same visual encoder."