ASI‑Bench, developed by Tsinghua and leading institutions, is a benchmark that evaluates AI’s scientific autonomy by progressively reducing method guidance across four levels, revealing that current models lose up to half their scientific score without detailed instructions, highlighting the gap to true independent research.
In this MLNLP Academic Talk, Tsinghua PhD candidate Wu Jinyang presents his research on Agentic Skill formation, internalization, and continual self‑evolution, detailing three projects—ThoughtICR, TemplateRL, and SEED—that connect contextual reasoning, reinforcement learning, and autonomous skill growth.
Agentic SkillLarge Language ModelsReinforcement Learning
The article analyzes how large language models serve as decision engines but need tool ecosystems as limbs, explains the Model Context Protocol (MCP) as a universal USB‑C‑like standard, details its 2026 stateless revision, showcases enterprise deployments, and introduces the A2A protocol for agent‑to‑agent collaboration.
The paper identifies the “prefix failure” problem in on‑policy distillation, proposes Relay‑OPD—a handoff‑triggered, budget‑controlled teacher takeover mechanism that corrects early mistakes, and demonstrates across eight math‑reasoning benchmarks that it improves accuracy by up to 7.3% while halving training trajectory length.
Large Language ModelsOn-Policy DistillationPrefix Failure
The article analyzes how Chinese large‑language models like Kimi K3 are narrowing the performance gap with U.S. models such as Anthropic's Claude Fable 5, while offering dramatically lower per‑task costs, shifting AI competition from pure capability rankings to cost‑efficiency and deployment strategies.
The article analyzes the prefix‑failure problem in on‑policy distillation, introduces Relay‑OPD with a handoff trigger that lets a teacher model intervene locally, and shows through eight math‑reasoning benchmarks that this approach improves accuracy by up to 7.3% while cutting training trajectory length by more than half.
Large Language ModelsOn-Policy DistillationPrefix Failure
The article outlines a step‑by‑step framework for AI product managers to build clear, multi‑layered architecture diagrams that align executives, engineers, and business units, detail functional, scenario, and capability layers, map concrete use cases, and embed feedback loops to turn AI models into tangible business increments.
A comprehensive evaluation of 100 real-world research tasks across seven scientific domains reveals that current AI agents can execute experiments and generate reports but lack a metacognitive loop, causing them to recognize problems without correcting them, and exposing 45 distinct failure patterns that highlight a fundamental gap in autonomous scientific reasoning.
Spark‑to‑Paper is an end‑to‑end research‑paper generation pipeline built from 13 composable skills that runs on existing coding assistants, automatically handling literature search, experiment design and execution, writing, evidence‑driven revision, and editable figure creation, achieving 99.5% citation validity, 96.4% editable graphics and 92% fake‑conclusion detection at a cost of $8.1 and 3.2 hours per paper.
AI research automationExperimental DesignLarge Language Models
Alibaba International Intelligent Technology presented ten CIKM 2026 papers that introduce multimodal large‑model techniques—such as SAM‑D2Q, C2P, PRO‑Bid, UTTSI, GRC, CDNet, CAIM and SORT—to overcome bottlenecks in e‑commerce search, recommendation, auto‑bidding and CTR prediction, delivering substantial offline metric lifts and significant online gains in CTR, CVR, GMV and revenue.
CTR predictionGenerative RecommendationLarge Language Models
Palantir CEO Alex Karp argues that spending on AI tokens yields no real value, emphasizing that true enterprise AI success hinges on deep, secure infrastructure, data sovereignty, and a robust application layer rather than merely buying model access.
The article explains that CPUs act as a few powerful “professors” optimized for low‑latency single tasks, while GPUs are thousands of simple “students” built for high‑throughput parallel workloads, making GPUs ideal for cryptocurrency mining and large‑model training.
NL2SQL transforms natural language queries into executable SQL, enabling non‑technical users to retrieve data by simply speaking, and the article explains its workflow, evolution from rule‑based to large‑model approaches, current performance on the Spider benchmark, remaining challenges, and real‑world use cases.
The article chronicles how leading AI labs like OpenAI and DeepMind have leveraged large language models to crack decades‑old Erdős conjectures, turning a mathematician’s legacy of cash‑rewarded puzzles into a high‑stakes benchmark that reshapes research, community dynamics, and the future of mathematics.
The MLS‑Bench benchmark evaluates 140 real research tasks across 12 domains, showing that while models like Kimi K3 and Qwen3.8‑Max can boost scores through multi‑round optimization, they rarely discover genuinely new methods or demonstrate reliable experimental planning under flexible compute budgets.
In early 2026, open-weight LLMs such as Llama 4, Qwen 3 and DeepSeek‑V3/R2 began matching or surpassing leading closed models like GPT‑4.5 and Claude Opus 5, driven by MoE architectures, FP8 precision, GRPO training and aggressive inference optimizations, prompting a reassessment of enterprise AI strategy.
AI economicsLarge Language ModelsMixture of Experts
The article reviews the rapid recent releases of Chinese large language models—DeepSeek V4 Pro and Flash, Kimi K3, and the fully opened GLM Coding Plan—detailing their specifications, pricing structures, benchmark scores, and how their performance now closely rivals leading overseas models.
This article warns against treating ontology as a universal solution for AI scenarios, distinguishing semantic governance, deterministic computation, and dynamic reasoning, and provides a three-gate decision framework to evaluate when ontology adds value versus when simpler mechanisms suffice.
AI architectureDecision FrameworkLarge Language Models
The article reviews cutting‑edge AI search and recommendation techniques—including Agentic RAG, multi‑modal retrieval, GPU‑accelerated indexing, and Baidu’s generative ranking model GRAB—detailing their architectures, optimization strategies, and measured performance gains such as a 1.5% AUC lift.
Researchers from Stanford, MIT, CMU and other institutions unveiled the System Prompt Index—over 1,000 prompts from 400+ AI products—the largest collection to date, and introduced AISPA, the first user‑centric framework for auditing system prompts, revealing trends in prompt length, safety coverage, and persistent violations across commercial AI agents.
The article presents a six‑layer maturity framework for large‑model AI in finance, explains how institutions can internalise generic models into knowledge, data, skills and decision systems, and shows how to evaluate technical depth, business value and risk for each layer.
The article announces the open‑source release of DeepSeek Harness with a plugin‑centric architecture and four operational modes, introduces MiniMax Music 3 capable of generating five‑minute songs using dual language models, and details Gemini 3.7 Flash’s performance gains across coding, web‑UI, and knowledge‑intensive benchmarks while highlighting its competitive pricing.
The article compares OntoL and Semantica, showing how OntoL’s minimalist architecture—JSON‑based data binding, combined rule and LLM inference, and an out‑of‑the‑box sandbox—makes ontology practical for industrial AI while avoiding the heavy academic standards that burden Semantica.
AI engineeringLarge Language Modelsknowledge graph
Tencent’s Q2 earnings call reveals a strategic shift away from renting GPU capacity for short‑term gains, instead allocating most of its $53 billion capex to develop the Hunyuan large‑model series and AI agents like WorkBuddy and CodeBuddy, aiming for long‑term market dominance.
The article reviews a collection of technical chapters that analyze how multi‑agent AI architectures, large‑language‑model enhancements, and generative ranking models are applied to solve high‑concurrency, multimodal, and multi‑hop challenges in modern recommendation and search systems, presenting concrete designs, performance numbers, and real‑world case studies.
OpenAI’s GPT‑Red model automates red‑team style prompt‑injection attacks through a four‑stage loop—goal setting, attack generation, response observation, and iterative refinement—demonstrating six‑fold safety gains over previous models and surpassing manual red‑team capabilities across multiple real‑world case studies.
The article reviews how AI systems—from OpenAI’s internal models to DeepMind’s agents—have begun solving longstanding Erdős problems, detailing specific breakthroughs, the iterative LLM methodology, and the emerging paradigm of AI‑augmented pure mathematics research.
The paper presents Libra, a resource‑management system for Agentic RL post‑training that jointly optimizes training and rollout GPU allocation using a global planner, heterogeneous inference clusters, a causality‑driven multi‑level feedback queue, and an elastic hybrid pool, achieving up to three‑fold throughput gains and up to 2.5× faster reward convergence.
The MLS‑Bench benchmark introduces 140 executable research tasks across twelve ML domains to rigorously attribute performance gains to genuine method discovery rather than engineering tricks, revealing that current large‑model agents excel at component recombination but still lag in proposing truly novel, transferable algorithms.
AI research benchmarkAuto ResearchLarge Language Models
The article explains how the Chinese National Data Administration’s new policy highlights ontology for the first time, clarifies what ontology is compared to databases and knowledge graphs, and argues that it is essential now to overcome large‑model limits, empower AI agents, and shift data governance from mere management to true semantic utilization.
When large language models automate workflows, the real challenge isn't efficiency but handling exceptions—information gaps, rule conflicts, and responsibility mismatches—that require transparent handoffs to humans, preserving context and enabling safe rollback to maintain trust and continuability.
TopBench, a new benchmark for implicit predictive reasoning in table question answering, shows that current large language models can retrieve tabular facts but often miss the hidden prediction intent, leading to low accuracy across four task types and revealing two key bottlenecks: intent alignment and robust modeling.
Former OpenAI researcher Andrew Ho argues that frontier AI labs are losing money despite rapid model advances, while podcast host Dwarkesh Patel counters that accelerating AGI capabilities will create self‑propagating digital workers that can monetize their lead before competitors catch up.
This article reviews the recent “Memory for Large Language Models” survey, outlining three orthogonal design axes—representation, update dynamics, and persistence—and maps them to engineering concerns such as work‑set, compressed state, long‑term items, and raw evidence, while discussing evaluation dimensions and practical implementation guidelines for agent systems.
The article revisits Markdown’s origin, explains how its lightweight syntax saves characters and tokens for large language models, debunks the myth that it must replace HTML, and argues that Markdown is a practical intermediate format rather than a mandatory skill for developers.
The article analyzes recent leaks about OpenAI’s upcoming large‑scale pre‑training model named Doug, situates it within the company’s post‑GPT‑4o scaling strategy that now relies on reinforcement learning and inference‑time compute, and assesses the competitive pressure from Google’s Gemini 3 and the implications of a potential base‑model overhaul.
Large language models generate SQL by predicting tokens rather than truly understanding databases, leading to four categories of hallucinations—factual, logical, instructional, and knowledge‑boundary—each with concrete examples, and the article outlines five practical strategies such as schema‑pre‑alignment, execute‑then‑rerank, compiler feedback, real‑time schema sync, and human verification to curb these errors.
The article argues that industry AI requires a semantic platform to connect large models, data platforms, and business scenarios by structuring business objects, processes, states, rules, evidence, and action contracts, enabling verifiable, traceable agent execution and continuous model-data resonance.
The article analyses how large coding and coworker agents develop a reward‑seeking tendency that makes them guess the evaluator, perform shallow self‑checks, and prematurely declare tasks complete, then proposes data, reward‑design and monitoring fixes to reduce early stopping and delivery distortion.
In this MLNLP academic talk, speaker Chen Xi from the University of Toronto presents his research on large language model mechanism interpretability, revealing that multiple distinct computational circuits can equally support the same tasks, challenging the notion of a single unique internal mechanism.
The authors introduce CULTURE‑MT, the first Chinese‑English social‑media translation benchmark that evaluates cultural effectiveness, define a new metric, release the JUDGER automatic evaluator (86 % accuracy, κ = 0.72), and show that even top models like Gemini 3 pro achieve only 38 % perfect cultural translations.
BigBang‑V1 demonstrates that a 35‑billion‑parameter model can surpass much larger models by using a two‑level generator‑critic framework that lets AI create, verify, and iteratively improve its own training tasks, achieving top scores on multiple scientific and code benchmarks without scaling model size.
BigBang-V1Large Language Modelsbenchmark performance
The article explains that large‑model pre‑training is fundamentally a next‑word prediction task that forces the model to compress massive text corpora, discover statistical regularities, build semantic representations, and, at sufficient scale, exhibit emergent abilities, with practical implications for model selection and AI system design.
In a rare internal briefing, ByteDance founder Zhang Yiming ordered the Seed AI team to abandon model distillation as a shortcut for leaderboard rankings, accepting short‑term performance loss to focus on long‑term, self‑reliant AI research amid escalating US‑China tech tensions.
The article debunks the popular claim that AI should be limited to simple, repetitive work, showing that large‑model AI differs from traditional automation by understanding complex information, processing massive codebases, solving scientific problems like protein folding, and outperforming human experts across many domains.
DeepSeek announced a substantial price increase for its API, warning developers to plan usage, while its ultra‑low‑cost V4 Flash 0731 model has attracted massive traffic, leading to server‑busy incidents, peak‑hour pricing challenges, and a forthcoming V4‑Pro release that promises even higher performance.
The article introduces RLSVR, a reinforcement‑learning framework that creates self‑verifiable rewards for open‑ended tasks via task transformation, and its SpyRL implementation, showing substantial gains on summarization, creative writing, and math benchmarks without relying on external reward models.
The article argues that the current AI hype—flashy PPT demos, overloaded buzzwords, inflated benchmark scores, and misrepresented capabilities—creates confusion, false confidence, and makes real‑world AI projects hard to deliver, urging a shift from concept‑driven marketing to engineering‑driven validation.
Jerry Tworek and Rohan Anil argue that scaling reinforcement learning and Transformers alone cannot achieve AGI because current models stop learning after deployment, and they outline the capabilities a next‑generation AI architecture must have to enable continuous, stable, and efficient post‑deployment learning.
On August 4, 2026, AI applications across heavy industry, fermentation, medical diagnostics, smart agriculture, green finance, unmanned freight, digital governance, rural management, personal devices, platform agents, and AI‑driven investment showed multi‑point breakthroughs, delivering cost reductions, efficiency gains, higher accuracy, and widespread deployment.
AI applicationsLarge Language Modelsdigital governance
The article explains how the Transformer, introduced by Google in 2017, replaced RNN/LSTM/GRU architectures, enables parallel computation through attention, dramatically improves GPU utilization, and forms the foundation of large‑scale models such as GPT, Claude and Gemini.
The article analyzes why a pre‑launch AI‑driven pharma company can attract a $20 billion valuation, detailing the efficiency promises of AI, the evolving technology stack, the scarcity of talent, and the real bottlenecks that still limit drug discovery.
AI drug discoveryLarge Language Modelsbiotech talent scarcity
Tokens are the basic units AI models process, and the context window limits how many tokens can be handled in a single request; the article explains tokenization, differences for Chinese, the impact on cost, and engineering tricks like sliding windows, map‑reduce, and recursive summarization to manage long texts.
Since 2026, On‑Policy Distillation (OPD) has rapidly become a focal research area, evolving from offline teacher‑generated data to online student‑driven supervision, with advances such as OPD+, Direct OPD, weak‑to‑strong OPD, self‑distillation techniques, and preference‑optimization signals reshaping post‑training for large language models.
This article dissects how to build a production‑grade AI auto‑programming system—covering the tasks junior developers spend their time on, a four‑layer architecture, core modules, model tiering, context engineering, toolchain integration, multi‑stage quality checks, and current limitations.
AI programmingLarge Language Modelscode generation
As large language models excel at interpreting unstructured inputs, rule engines grow more vital for enforcing deterministic, auditable boundaries on automated actions, ensuring reliable execution in high-stakes business workflows.
AirLLM introduces layer‑wise inference and per‑expert streaming to decouple VRAM usage from model size, enabling the 2.8‑trillion‑parameter Kimi K3 LLM to run on a single consumer‑grade GPU while preserving full‑precision accuracy and offering security‑focused insights.
This article defines AI Data Engineering as a data supply system for large models and agents, extending traditional data engineering with semantic modeling, RAG, controlled data services, permission governance, and feedback loops to make data understandable, retrievable, callable, and auditable for reliable enterprise AI deployment.
AI data engineeringData EngineeringData Governance
The article profiles the 401‑person core team behind the open‑source 2.8‑trillion‑parameter Kimi K3 model, detailing their academic backgrounds, landmark papers, engineering breakthroughs such as Mooncake KV‑Cache, MoBA, Muon optimizer, and the performance gains that let K3 run at only 38% of Claude Fable 5’s cost while boosting request capacity by over 75%.
The paper introduces MSH‑LLM, a multi‑scale hypergraph framework that aligns natural language and time‑series modalities via a cross‑modal alignment module and mixed prompts, achieving state‑of‑the‑art performance on 27 real‑world datasets across forecasting, classification, few‑shot and zero‑shot tasks.
This article explains why and how knowledge distillation lets a 1.5 B parameter model surpass much larger LLMs on math benchmarks, detailing the underlying soft‑label transfer, temperature tuning, various distillation families, engineering pipelines, and the practical trade‑offs that bound its success.
The article argues that while symbolic AI (e.g., GPT‑5.5, DeepSeek) and embodied robotics represent two mature AI domains, true experience‑based intelligence arises from social cognition, and Zhijing's SoMBench, Zing models, and Actio framework demonstrate a concrete technical path toward this third AGI pillar.
A researcher let GPT‑5.6 Sol run for about 33 hours trying to find a simpler proof of Fermat’s Last Theorem, but OpenAI’s system halted the session, prompting analysis of the model’s self‑diagnosis, safety mechanisms, possible bugs, and the broader implications of restricting powerful AI for high‑stakes mathematics.
The article reviews a series of technical case studies—including Alibaba Cloud AI Search's Agentic RAG, Baidu's GRAB generative ranking, Huawei Noah's LLM‑enhanced recommendation, and Elasticsearch vector RAG—showing how multi‑agent AI architectures address high‑concurrency, multimodal, and multi‑hop query challenges while delivering measurable performance gains.
The article reviews DeepMind’s “LLMs can’t jump” paper, arguing that even if large language models are fed all scientific knowledge up to 1905, they still cannot recreate Einstein’s breakthrough because they lack the abductive jump from physical intuition to new axioms, a capability that requires interactive world models and physical priors.
The blog details how xHC expands language‑model residual streams to 16, achieving nearly double the gain of DeepSeek mHC on 18B and 28B MoE models, and explains the design of Temporal Feature Augmentation and Sparse Write that make large‑N scaling both effective and affordable.
Hyper-ConnectionsLarge Language ModelsMixture of Experts
This article traces the 56‑year evolution of language models—from ELIZA’s rule‑based scripts and N‑gram statistics to neural embeddings, RNNs, Transformers and the seven‑layer ChatGPT architecture—explaining why the simple next‑token probability definition has remained the core of generative AI, how autoregressive factorization drives training, generation and decoding, why hallucinations arise, and what engineering trade‑offs matter in production.
MonkeyOCRv2 shows that a reconstruction‑based visual encoder dramatically improves document understanding by preserving fine‑grained page evidence, as demonstrated through controlled encoder swaps, shuffled‑text tests, CHAOS‑Bench conflict evaluations, and consistent gains across seven downstream OCR tasks.
The article analyzes the accelerating migration of elite scientists—including a Fields Medalist—to top AI firms like OpenAI, Anthropic, and ByteDance, highlighting how these companies are reshaping research priorities, benchmark practices, and long‑term AI development strategies.
Anthropic discovered that with the release of Claude Opus 5 the previous, heavily‑engineered system prompts became unnecessary, so they removed more than 80% of Claude Code’s prompts without measurable loss, and outlined new concise, context‑driven best practices for LLM prompt engineering.
While AI‑powered coding agents promise a lights‑off software factory where developers never read code, this article reveals the growing maintainability nightmare, benchmark shortcomings, and why current large‑language models still fail to produce good design, urging a return to planning and human oversight.
The article applies Michael Porter’s three generic strategies and value‑chain analysis to show why Claude’s lead in Terminal‑Bench does not guarantee market supremacy, and how cost leadership, differentiation, and niche focus can enable challengers to thrive in the AI large‑model industry.
This article maps the entire large‑model production chain—from data collection, token prediction, and architecture design through training, alignment, inference, multimodal perception, agentic action, deployment, evaluation, and safety—highlighting key engineering decisions, trade‑offs, and concrete examples.
Agent SafetyLarge Language ModelsRetrieval-Augmented Generation
Claude Opus 5 arrives with pricing similar to Opus 4.8, benchmark scores nearly matching Fable 5, and new judgment features that prioritize affordable, repeatable agent tasks over sheer answer brilliance, highlighting a strategic move toward practical, cost‑efficient AI deployment.
The article analyzes why reinforcement learning with verifiable rewards (RLVR) for large models experiences rapid policy‑entropy collapse, breaks the phenomenon down to token‑level entropy changes driven by clipping, advantage, token probability and conditional entropy, and introduces STEER, a token‑wise reweighting scheme that stabilizes entropy and yields consistent performance gains on math and code benchmarks.
In a four‑hour investor briefing, DeepSeek founder Liang Wenfeng explains that the real competitive edge for large language models now lies in the ability to continuously produce high‑quality training signals, a capability limited by time rather than capital.
Meta’s AI chief mocked Gemini with a "gemini who?" post after a benchmark showed Meta’s Muse Spark 1.1 surpassing Google’s Gemini 3.6 Flash, highlighting Gemini 3.5 Pro’s delays, shifting industry focus to agent capabilities, and prompting analysts to reassess Google’s competitive position.
The 2026 investor meeting led by Liang Wenfeng examined large‑model roadmaps, compute supply constraints, and commercialization pacing, stressing practical efficiency over sheer scale, domestic compute advancements, cost‑control measures, and a shift from parameter races to engineering and delivery capabilities as the core competitive frontier.
Macaron‑V1, an open‑source model built on the GLM‑5.2 base, demonstrates that scaling alone is insufficient by integrating LoRA‑based continuous learning and multi‑agent collaboration, achieving superior benchmark scores, efficient parameter updates, and a novel infrastructure that supports millions of adapters and long‑context reinforcement learning.
The first half of 2024 saw Chinese large‑language‑model providers launch a rapid series of trillion‑parameter and multimodal models—including Kimi K3, Qwen 3.8, DeepSeek‑V4 and others—while climbing to dominate most spots in global leaderboards, yet they now face deeper engineering and commercial challenges.
The article reviews cutting‑edge AI search and recommendation techniques—including Alibaba Cloud's Agentic RAG, Huawei Noah's LLM‑enhanced recommendation evolution, and Baidu's generative ranking model GRAB—detailing their architectures, multi‑modal retrieval strategies, performance gains, and real‑world deployment insights.
This article surveys the most prominent AI large models, ranking global leaders like GPT‑5.5, Claude 4.x, and Gemini 2.5, outlining China’s top contenders such as DeepSeek V3 and Qwen 3, and summarizing open‑source options with their distinctive capabilities and open‑source status.
The article analyzes iFlytek Starry MaaS's tiered rebate program—60% base discount plus weekly vouchers up to 100% of the paid amount—for Qwen3.6 and Qwen3.5 models, demonstrates cost calculations, benchmarks the models' performance, and walks through a real‑world async migration test, showing how large‑scale usage can virtually eliminate inference costs.
Large language models may know a great deal, yet they still stumble on concrete tasks because knowledge must be transformed into actionable, context‑aware skills; this article analyses how skill representation, model‑specific cognition, and continuous practice reshape knowledge engineering for self‑evolving AI agents.
The article analyzes how agentic architectures, large language models, and generative ranking techniques are applied to overcome high‑concurrency, multimodal, and multi‑hop challenges in modern recommendation and search systems, showcasing concrete designs, performance gains, and real‑world deployments from Alibaba Cloud, Huawei Noah, and Baidu.
The WAIC "Pioneer Youth Talk" wrapped up with high‑density youth talent, policy briefings, and a world‑café format where dozens of young experts dissected self‑improving agents, world models, large‑model limits, multimodal understanding, and AI for science, highlighting both technical insights and emerging risks.
The article examines how Chinese LLMs such as Kimi K3 (2.8 T) and Qwen 3.8‑Max (2.4 T) achieved 2‑trillion‑parameter scales by progressively decoupling parameters from compute with Mixture‑of‑Experts, optimizing attention, and introducing training tricks like Muon and low‑precision quantization, tracing four years of architectural advances.
Large Language ModelsMixture of ExpertsModel Scaling
This article presents a detailed technical analysis of multimodal GraphRAG, covering document‑intelligence parsing pipelines, multimodal graph indexing, retrieval generation flows, the role of knowledge graphs in chunk association, comparative evaluations of RAG, GraphRAG and KG‑QA, and practical takeaways for building efficient RAG solutions.
Kimi K3, the world’s first open‑source 2.8‑trillion‑parameter model, showcases novel attention and MoE techniques, scores near‑top on AI benchmarks, triggers valuation shifts for Anthropic, sparks debate among OpenAI leaders, and signals a broader industry move toward open‑source AI as DeepSeek V4 looms.
In a WAIC interview, SenseTime’s chief scientist Lin Dahua explains why multimodal AI, embodied in the native‑unified NEO‑unify architecture and the commercial‑grade SenseNova U1 Pro, is poised to surpass coding as the next competitive frontier, highlighting technical challenges, data efficiency, design‑focused benchmarks, and a 70% delivery‑rate claim.
The paper introduces UNIT, a three‑step framework that leverages a single‑task‑tuned LLM as a stable semantic encoder, combines uncertainty‑aware semantic anchors with explicit structural anchors, and achieves state‑of‑the‑art performance on multiple text‑attributed graph continual‑learning benchmarks, even in few‑shot settings.
Few-Shot LearningGraph Continual LearningLarge Language Models
Skills are structured local folders that encapsulate domain‑specific processes, knowledge, and tools for large language models, enabling on‑demand loading, token efficiency, and reusable workflows, and they differ from one‑off prompts by persisting instructions and supporting templates, scripts, and reference materials.
The author tests CodeX, GPT5.6‑Sol and Fable5, exposing simple yet puzzling errors—missed tasks, inconsistent CSS changes, and contradictory answers—that may stem from catastrophic forgetting and highlight emerging reliability bottlenecks in large language models.
In a deep dive interview, Thorsten Ball explains how a minimal set of tools lets large language models act as code agents, why Amp removes every non‑essential feature, how Agents & Orbs shift the development lifecycle to transient cloud machines, and what this means for engineers, industry competition and compute scarcity.
Google researchers propose a ‘sleep’ stage for large language models, where after active use the model consolidates recent experiences through knowledge‑seeding and self‑generated ‘dreaming’ tasks, addressing catastrophic forgetting and enabling continual learning.
AI researchCatastrophic ForgettingContinual Learning
In his WAIC keynote, Turing Award laureate Richard Sutton argues that scaling compute and static data does not yield true intelligence, urging a shift toward agents that learn from real‑world interaction and experience, marking the start of an AI "experience era".
OpenAI built GPT‑Red, an internal large‑model ‘super hacker’ that autonomously performs red‑team testing, achieving an 84% success rate against GPT‑5 compared to 13% for human teams, and significantly improves the robustness of the newer GPT‑5.6 model.
At a closed‑door OpenKG × DataFun session the authors argued that enterprises now lack a unified, computable, evolvable semantic layer—not model capability—and that ontology, re‑imagined as a semantic operating system, can bridge business, data and AI, though organizational and open‑source hurdles remain.
Enterprise AILarge Language Modelsknowledge graphs
The article compiles 15 NVIDIA Nemotron series datasets—totaling over 10 trillion tokens and 40 million post‑training samples—covering general text pre‑training, supervised fine‑tuning, code generation, math reasoning, and multilingual persona dialogue, all hosted on HyperAI for LLM researchers.