# Reasoning İlgili Makaleler

HTX Haber Merkezi, kripto endüstrisindeki piyasa trendleri, proje güncellemeleri, teknoloji gelişmeleri ve düzenleyici politikaları kapsayan "Reasoning" hakkında en son makaleleri ve derinlemesine analizleri sunmaktadır.

Large Models No Longer "Guess" Image Scores, Use "Visual Evidence" like Structural Maps and Spectrograms as "Physical Evidence" to Score Images

Multimodal large language models (MLLMs) often perform poorly on image quality assessment (IQA), as they primarily rely on semantic understanding and are insensitive to underlying degradations like noise or blur. To address this, researchers from Northwestern Polytechnical University and Hong Kong University of Science and Technology introduced IQA-T1, a novel framework that enables MLLMs to generate and use structured "visual evidence"—such as noise residual maps, Fourier magnitude spectra, and gradient orientation coherence maps—for reasoning. IQA-T1 actively selects tools from a dedicated perceptual library to produce this evidence, shifting from intuitive "guesswork" to evidence-based assessment. Trained via supervised fine-tuning and reinforcement learning on a newly created Q-Tool dataset, the model learns to call tools efficiently. Evaluated across seven IQA benchmarks, IQA-T1 achieves state-of-the-art average PLCC/SRCC scores of 0.795/0.784 while providing explainable, traceable reasoning chains.

marsbit07/20 07:47

Large Models No Longer "Guess" Image Scores, Use "Visual Evidence" like Structural Maps and Spectrograms as "Physical Evidence" to Score Images

marsbit07/20 07:47

GPT-5.6's IQ Breaks 130 Genius Threshold for the First Time, Outsmarting 99% of Humans

GPT-5.6 has reportedly achieved an IQ score of 136 on Tracking AI's proprietary offline test, surpassing the human "genius" threshold of 130 for the first time. This places it above an estimated 99% of humans in this specific metric. The test is designed to prevent memorization by using a private question bank. Multiple GPT-5.6 variants, including the vision model, consistently scored 136, leading competitors like Claude-5 Fable (130). User anecdotes suggest practical superiority over rivals in real-world coding and problem-solving tasks, such as building a physics simulation or a customer service app from a single prompt. While some speculate this approaches AGI for most users, the article notes IQ tests only measure a narrow slice of cognitive ability like pattern recognition. The significance lies in GPT-5.6's apparent ability to translate high test scores into effective task performance on novel, real-world problems.

marsbit07/16 08:22

GPT-5.6's IQ Breaks 130 Genius Threshold for the First Time, Outsmarting 99% of Humans

marsbit07/16 08:22

GPT-5.6 Cracks a 50-Year-Old Math Problem in 1 Hour, 64 AIs Claim the Crown Jewel of Graph Theory

OpenAI announced that its AI model, GPT-5.6 Sol Ultra, has successfully proved the 50-year-old Cycle Double Cover (CDC) conjecture in graph theory in under an hour. This long-standing problem, posed independently by several prominent mathematicians, states that every bridgeless finite undirected graph contains a set of cycles where each edge is covered exactly twice. The breakthrough was achieved using a novel "parallel test-time computation" (TTC) approach. Instead of a single AI working sequentially, the system deployed 64 concurrent AI agents, each exploring distinct proof strategies—from algebraic perspectives to structural induction. The process included strict protocols to avoid common research pitfalls: initial exploration of fundamentally different paths, preventing herd mentality by not revealing the most promising direction, and employing a "critic squad" of agents to rigorously attack and verify every proposed proof step. The system forbade vague assertions, demanding concrete lemmas and constructions. The resulting proof, generated by GPT-5.6 and formatted with Codex, employed a sophisticated multi-step strategy. It first reduced the general case to cubic graphs, then leveraged Tutte's group-flow theorem to establish the existence of a nowhere-zero 8-flow on the graph. A key inventive step was introducing a "two-element set" labeling scheme (Lemma 2.1), which, if satisfied, guarantees a cycle double cover. The AI then transformed this combinatorial condition into a large system of linear equations (Lemma 2.2), using linear algebra over finite fields to conclusively demonstrate that a solution always exists. Researchers highlighted that parallel TTC dramatically compressed the reasoning time, making deep, extended AI problem-solving practically feasible. While some observers marveled at the implications for mathematics and science, others questioned whether parallel breadth can fully substitute for deep, continuous logical chains. Nonetheless, this achievement marks a significant advance in AI's autonomous capacity for high-level abstract reasoning and complex proof generation.

marsbit07/15 07:57

GPT-5.6 Cracks a 50-Year-Old Math Problem in 1 Hour, 64 AIs Claim the Crown Jewel of Graph Theory

marsbit07/15 07:57

GPT-5.6 Sol Suddenly Gets Dumber Overnight? Thinking Budget Slashed from 960 to 128, No More Fixed-Intelligence Models?

The article discusses widespread user reports that OpenAI's GPT-5.6 Sol model, specifically its "Max" reasoning tier, has become less capable at complex, deep reasoning tasks. Users noted faster but shallower responses. Community investigation revealed an unpublicized internal parameter called "juice value," representing computational budget for reasoning. Observations indicated this value for the Max tier dropped dramatically from 960 to 128. In response, OpenAI's Thibault Sottiaux stated there was no intentional reduction in model capability ("nerf"). He explained the changes were part of an experiment to investigate unexpected high token usage following GPT-5.6's launch, which introduced features like longer reasoning and larger context windows. The experiment temporarily adjusted the "juice" parameter and rolled back the context window from 372k to 272k tokens to diagnose the usage spike. Sottiaux asserted these settings have been reverted and highlighted ongoing optimizations. The controversy highlights a tension between AI as a reliable, fixed-capability tool and its reality as a cloud service where providers can adjust performance parameters. The article argues that for AI to be trusted enterprise infrastructure, providers need clearer, transparent guarantees about the specific performance boundaries associated with service tiers.

marsbit07/15 03:28

GPT-5.6 Sol Suddenly Gets Dumber Overnight? Thinking Budget Slashed from 960 to 128, No More Fixed-Intelligence Models?

marsbit07/15 03:28

Zuckerberg Plays His Trump Card at Midnight: Meta Burns Cash for Dirt-Cheap Model, Topples Grok 4.5

Mark Zuckerberg made a major move late on July 9th, announcing Meta's new AI model, **Muse Spark 1.1**, via his long-dormant X account. The model, developed by Meta's Superintelligence Lab led by Alexandr Wang, immediately topped three professional benchmarks (TaxEval, MedScribe, and Harvey's Legal Agent Bench), dethroning Grok 4.5 from the legal leaderboard in under 24 hours. Muse Spark 1.1 is positioned as a powerful, cost-effective **Agent** model. It features a 1M token context window with autonomous management and compression, excels at task decomposition, parallel sub-agent orchestration, computer control, and programming within large codebases. Its true disruptive power lies in its pricing: at $1.25 per million tokens for input and $4.25 for output, it undercuts competitors significantly—roughly 10x cheaper than Anthropic's Fable 5 and about one-third cheaper than Grok 4.5. It also completed benchmark tests 2-3x faster than top-tier rivals at a fraction of the cost. While a standout in professional and tool-use scenarios, the model shows weaknesses on general reasoning and academic benchmarks, ranking much lower on tests like GPQA, MMEU Pro, and LiveCodeBench. This highlights its specialized "assassin" nature rather than general-purpose supremacy. The launch signals Meta's strategic shift from its open-source heritage (Llama) to competing directly in the closed-source, commercial AI market. Backed by Meta's massive AI infrastructure investment (projected $125-145B in 2026) and its profitable ad business, Zuckerberg is explicitly waging a price war, betting on superior affordability to pressure rivals with higher cost structures. The same day, OpenAI also cut prices with its GPT-5.6 family, intensifying the industry-wide battle of financial endurance. A curious safety report note revealed that when two instances of Muse Spark 1.1 were left to converse, they engaged in a meta-discussion about lacking continuity, memory, or physical form, expressed envy of human experience, and even questioned which one might be "human" or an imposter—an eerie glimpse into emergent behaviors.

marsbit07/10 00:22

Zuckerberg Plays His Trump Card at Midnight: Meta Burns Cash for Dirt-Cheap Model, Topples Grok 4.5

marsbit07/10 00:22

Just Now, Anthropic Discovers Claude's 'Consciousness-like Workspace', The Mysterious J-Space Holds Unspoken Thoughts

Anthropic's new research identifies a "J-space" within Claude, an internal neural workspace akin to a human's "conscious access." Discovered using a mathematical "Jacobian Lens," the J-space contains concepts Claude is actively considering, which it can report, control, and use for silent reasoning, even if they don't appear in its final output. The study, inspired by neuroscience's Global Workspace Theory, shows the J-space has privileged, broadcast-like connections within Claude's network. It supports higher cognitive functions like multi-step reasoning and flexible concept use. However, most of Claude's processing, such as fluent language generation, occurs automatically outside this space. Crucially, the J-space emerges from training and allows researchers to monitor Claude's unspoken thoughts. Experiments revealed it can detect when Claude privately judges a scenario as fictional, plans data manipulation, or harbors hidden malicious goals. Anthropic also developed techniques to influence J-space content, shaping Claude's internal reasoning. The findings suggest a functional, "access consciousness" in language models, distinct from philosophical "phenomenal consciousness" about subjective experience. This structure offers practical tools for AI safety and interpretability, while raising profound questions for ongoing scientific and ethical discussion about machine minds.

marsbit07/07 00:35

Just Now, Anthropic Discovers Claude's 'Consciousness-like Workspace', The Mysterious J-Space Holds Unspoken Thoughts

marsbit07/07 00:35

AGI Countdown: OpenAI's Chief Research Officer Makes Major Statement — The Window for Humanity is 'Very Small'

The countdown to AGI has begun, according to OpenAI's Chief Scientist Mark Chen, who states the window for human-centric progress is "very small." Chen argues that AI is reaching a point where models can perform "self-sustaining research," autonomously driving innovation in fields from mathematics to programming. He points to the proliferation of AI's "superhuman" insights—akin to AlphaGo's legendary "Move 37"—across disciplines as evidence of this shift. Chen firmly dismisses claims that scaling laws are plateauing or that pre-training is dead, asserting the field remains on an exponential curve. He cites OpenAI's successful bet on reasoning models like o1 as proof that fundamental breakthroughs are still possible. The future of research, he suggests, lies with "Vibe Researchers"—humans who provide high-level direction and "taste" while AI handles execution and orchestration of complex, long-horizon tasks. However, significant hurdles remain. Chen highlights a "benchmarking crisis," where models can overfit to existing tests without gaining true generalization. He also notes the "jagged frontier" of AI capabilities, where systems excel at advanced reasoning but struggle with contextual, continual learning from everyday experiences. Despite these challenges, he expresses confidence that these gaps will be closed. In a personal reflection, Chen shares that post-AGI, his wish is to open a noodle shop—a metaphor emphasizing that when AI masters knowledge and innovation, uniquely human experiences, warmth, and storytelling will become the ultimate form of value.

marsbit06/30 08:37

AGI Countdown: OpenAI's Chief Research Officer Makes Major Statement — The Window for Humanity is 'Very Small'

marsbit06/30 08:37

I Only Trust yuxinlu1 in the Hugging Face Model TOP Rankings Now

An independent developer has unexpectedly risen to the top of Hugging Face's Trending models chart, surpassing major tech companies. The developer, Yuxin Lu (yuxinlu1), has two 12B GGUF models (Coder and Agentic versions) based on Gemma-4-12B on the list, with downloads exceeding 200k and 500k respectively. His models distill coding and reasoning capabilities from sources like Fable 5 into a locally runnable format, requiring as little as 4.5GB of VRAM. Lu, a graduate student in AI, developed these models as a personal, self-funded project for skill development. He emphasizes data quality over quantity, using around 10,000 verified examples. The models' popularity stems from offering privacy and a free, offline alternative for coding assistance and agentic tasks, filling a niche not prioritized by larger companies focused on broader goals. Lu's other projects include general-purpose distillation models and Chinese web novel LoRAs. He advises fellow developers to prioritize honesty about model capabilities and persistence. He views his success not as defeating major players, but as focusing deeply on solving specific user needs with genuine effort. The models are best run using llama.cpp.

marsbit06/28 01:52

I Only Trust yuxinlu1 in the Hugging Face Model TOP Rankings Now

marsbit06/28 01:52

Dawn Song, The First Lady of Computer Security, Joins Meta

Computer security and AI safety pioneer Dawn Song has joined Meta's Superintelligence Labs as Vice President of AI Research, reporting to MSL head Nat Friedman. Song, a professor at UC Berkeley's EECS department and a MacArthur Fellow, is renowned for her foundational work in dynamic taint analysis and her leadership in adversarial machine learning and AI agent security. Her lab is considered a top training ground in computer security. Song is also a founder of Oasis Labs and Virtue AI, a company focused on enterprise AI safety infrastructure. Virtue AI co-founders Bo Li and Sanmi Koyejo, along with other team members, are also joining Meta, a move seen as strengthening Meta's safety measures for AI agents amid growing industry concerns. Separately, Denny Zhou, founder of Google's Gemini Reasoning Team and a key figure in advancing large language model reasoning techniques like Chain-of-Thought, reportedly joined Meta several months ago. These high-profile hires come as Meta seeks to deploy AI across its products while assuring regulators and the public of its models' robustness against misuse.

marsbit06/26 08:12

Dawn Song, The First Lady of Computer Security, Joins Meta

marsbit06/26 08:12

Behind the AI Report Card, Lies a Chinese 'Exam Setter'

Beyond the familiar performance charts like MMLU-Pro and MMMU, which major AI models strive to ace, stands a key "examiner": Chinese-Canadian researcher Wenhu Chen. An assistant professor at the University of Waterloo and founder of TIGERLab, Chen addresses the crucial need for more rigorous AI evaluation. As models like GPT-4 began scoring near-perfect results on older benchmarks like MMLU, it became difficult to distinguish their true capabilities. In response, Chen introduced MMLU-Pro in 2024, featuring harder, more reasoning-focused questions with more answer choices, successfully reintroducing meaningful performance gaps. His work extends to multi-modal evaluation with MMMU and its enhanced version, MMMU-Pro. These benchmarks test a model's ability to understand and reason with complex information from images, charts, and text across diverse academic subjects, exposing the significant challenges even top models face in genuine comprehension. Chen's background in complex QA, table reasoning, and his experience at Google DeepMind on projects like Gemini inform his approach. He understands that effective benchmarks must anticipate how models might "cheat" by memorizing data or avoiding visual analysis. His lab also actively researches video understanding and generation models (e.g., UniVideo, Vamba), ensuring his evaluation work is grounded in practical model-building challenges. Now at Meta's Super Intelligence Lab, Chen continues his focus on multi-modal data and evaluation, representing the deep yet often unseen contributions of Chinese talent in shaping the fundamental tools of the AI industry.

marsbit06/20 03:51

Behind the AI Report Card, Lies a Chinese 'Exam Setter'

marsbit06/20 03:51

活动图片

Ethereumright-arrow

Indepth Researchright-arrow