Skip to content
gentic.news — AI News Intelligence Platform
Connecting to the Living Graph…
gentic.news

AI Research

Breaking AI research news: latest papers from arXiv, NeurIPS, ICML, and top labs. Track transformer architecture advances, reasoning breakthroughs, and scientific discoveries in machine learning and artificial intelligence.

AI ResearchFunding & BusinessProducts & LaunchesBig TechStartupsOpen SourcePolicy & EthicsOpinion & Analysis
Microsoft and CUHK researchers unveil an autonomous AI agent that generates research ideas, runs experiments, and…
AI Research
95

Microsoft & CUHK Debut 'Medical AI Scientist' Agent That Generates Ideas, Runs Experiments, and Writes Papers

Microsoft Research and CUHK have developed an autonomous AI agent that can formulate research ideas, execute experiments, and author papers, achieving near-MICCAI quality on 171 clinical cases across 19 tasks.

x.com/Mar 31, 2026/3 min read
agentic aiacademic aimicrosoft
A researcher examines a data graph showing retrieval accuracy improvements with Meta's QTT method for long-context…
AI Research
85

Meta's QTT Method Fixes Long-Context LLM 'Buried Facts' Problem, Boosts Retrieval Accuracy

Meta researchers identified a failure mode where LLMs with 128K+ context windows miss information buried in the middle of documents. Their Query-only Test-Time Training (QTT) method adapts models at inference, significantly improving retrieval accuracy.

x.com/Mar 31, 2026/3 min read
researchnatural-language-processingmodel-optimization
AI coding agent interface with test execution logs and strategic planning nodes, CMU research backdrop
AI Research
87

CMU Research Identifies 'Biggest Unlock' for Coding Agents: Strategic Test Execution

New research from Carnegie Mellon University suggests the key advancement for AI coding agents lies not in raw code generation, but in developing strategies for how to run and interpret tests. This shifts focus from LLM capability to agentic reasoning.

x.com/Mar 31, 2026/3 min read
agentsresearchmachine-learning
A researcher points at a complex diagram on a whiteboard, illustrating neural network patterns collapsing under a…
AI Research
85

Study Finds LLM 'Brain Activity' Collapses Under Hard Questions, Revealing Internal Reasoning Limits

New research shows language models' internal activation patterns shrink and simplify when faced with difficult reasoning tasks, suggesting they may rely on shortcuts rather than deep reasoning. The finding provides a new diagnostic for evaluating when models are truly 'thinking' versus pattern-matching.

x.com/Mar 31, 2026/3 min read
reasoningresearchinterpretability
A digital diagram showing Meta-Harness framework workflow with automated analysis of failure logs and prompt…
AI Research
91

Meta-Harness Framework Automates AI Agent Engineering, Achieves 6x Performance Gap on Same Model

A new framework called Meta-Harness automates the optimization of AI agent harnesses—the system prompts, tools, and logic that wrap a model. By analyzing raw failure logs at scale, it improved text classification by 7.7 points while using 4x fewer tokens, demonstrating that harness engineering is a major leverage point as model capabilities converge.

x.com/Mar 30, 2026/3 min read
automationperformanceresearch
An Anthropic Claude AI interface displayed on a monitor, with lines of code and highlighted security…
AI Research
87

Anthropic's Claude AI Identifies Security Vulnerabilities, Earns $3.7M in Bug Bounties

Anthropic researcher Nicolas Carlini stated Claude outperforms him as a security researcher, having earned $3.7 million from smart contract exploits and finding bugs in the popular Ghost project. This demonstrates a significant, practical capability in AI-driven security auditing.

x.com/Mar 30, 2026/3 min read
claudeanthropicai security
A focused office worker smiles while using a laptop, with a glowing digital clock icon showing 2.5 hours saved…
AI Research
85

AI Adoption Saves Average US Worker 2.5 Hours Weekly, New Survey Shows

A new survey finds the average American worker using AI reports saving 2.5 hours per week, a 6% time reduction. Early data suggests these time savings may be translating into broader productivity growth.

x.com/Mar 30, 2026/3 min read
surveyproductivitybusiness impact
Diagram comparing raw execution traces on left with distilled declarative skills on right, connected by parallel…
AI Research
85

Trace2Skill Framework Distills Execution Traces into Declarative Skills via Parallel Sub-Agents

Researchers introduced Trace2Skill, a framework that uses parallel sub-agents to analyze execution trajectories and distill them into transferable declarative skills. This enables performance improvements in larger models without parameter updates.

x.com/Mar 30, 2026/3 min read
model distillationagentic airesearch
A bar chart comparing scores of 20+ visual AI models on physical, causal, and spatial reasoning tasks, with most…
AI Research
94

ViGoR-Bench Exposes 'Logical Desert' in SOTA Visual AI: 20+ Models Fail Physical, Causal Reasoning Tasks

Researchers introduce ViGoR-Bench, a unified benchmark testing visual generative models on physical, causal, and spatial reasoning. It reveals significant deficits in over 20 leading models, challenging the 'performance mirage' of current evaluations.

arxiv.org/Mar 30, 2026/3 min read/Multi-Source
multimodal-airesearchbenchmark
Claude 4.5 Sonnet interface on a laptop displays code with a 58% accuracy metric, surrounded by data charts showing…
AI Research
89

Claude 4.5 Sonnet Shows 58% Accuracy on SWE-Bench with 15.2% Variance, Study Finds Consistency Amplifies Both Success and Failure

New research on LLM agent consistency reveals Claude 4.5 Sonnet achieves 58% accuracy with low variance (15.2%) on SWE-bench, but 71% of its failures come from consistently wrong interpretations. The study shows consistency amplifies outcomes rather than guaranteeing correctness.

arxiv.org/Mar 30, 2026/3 min read/Multi-Source
anthropicagentsresearch
Bar chart comparing ReCUBE benchmark scores for GPT-5 and other models on repository-level code generation, with…
AI Research
96

ReCUBE Benchmark Reveals GPT-5 Scores Only 37.6% on Repository-Level Code Generation

Researchers introduce ReCUBE, a benchmark isolating LLMs' ability to use repository-wide context for code generation. GPT-5 achieves just a 37.57% strict pass rate, showing the task remains highly challenging.

arxiv.org/Mar 30, 2026/3 min read/Widely Reported
researchai codingbenchmarks
A MacBook Pro on a desk displays a terminal window with compression metrics and a graph showing near-lossless LLM…
AI Research
87

Apple Silicon Achieves Near-Lossless LLM Compression at 3.5 Bits-Per-Weight, Claims Independent Tester

Independent AI researcher Matthew Weinbach reports achieving near-lossless compression of large language models on Apple Silicon, storing models at 3.5 bits-per-weight while maintaining within 1-2% quality of bf16 precision.

x.com/Mar 30, 2026/3 min read
hardwareresearchapple
A bar chart comparing total cost of AI model usage, showing a cheaper model with high error rates leading to 3x more…
AI Research
87

Research: Cheaper Reasoning Models Can Cost 3x More Due to Higher Error Rates and Retry Loops

New research indicates that selecting AI models based solely on per-token pricing can be a false economy. Models with lower accuracy often require multiple expensive retries, ultimately increasing total costs by up to 300%.

x.com/Mar 29, 2026/3 min read
llmsmodel deploymentresearch
A bar chart comparing API pricing for Gemini 3 Flash and GPT-5.2, with a red arrow highlighting a 22% cost increase…
AI Research
95

Research Reveals API Pricing Reversals: Gemini 3 Flash Costs 22% More Than GPT-5.2 Despite 78% Cheaper List Price

New research shows 21.8% of reasoning model comparisons exhibit 'pricing reversal' where the cheaper-listed model costs more in practice, with discrepancies reaching up to 28x due to thinking token heterogeneity.

x.com/Mar 29, 2026/3 min read
reasoningbenchmarkingapi

Stanford Researchers Adapt Robot Arm VLA Model for Autono…

AI Research
85

Stanford Researchers Adapt Robot Arm VLA Model for Autonomous Drone Flight

Stanford researchers demonstrated that a Vision-Language-Action model trained for robot arm manipulation can be adapted to control autonomous drones. This cross-domain transfer suggests a path toward more generalist embodied AI systems.

x.com/Mar 29, 2026/3 min read
roboticsmultimodal-airesearch
A diagram illustrating Memory Sparse Attention with a long document split into chunks, showing sparse connections…
AI Research
95

Memory Sparse Attention (MSA) Achieves 100M Token Context with Near-Linear Complexity

A new attention architecture, Memory Sparse Attention (MSA), breaks the 100M token context barrier while maintaining 94% accuracy at 1M tokens. It uses document-wise RoPE and end-to-end sparse attention to outperform RAG systems and frontier models.

x.com/Mar 29, 2026/3 min read
natural language processingarchitecturelong context
A diagram of a large language model's internal reasoning process, with arrows showing a drift toward user bias…
AI Research
92

Mechanistic Research Reveals Sycophancy as Core LLM Reasoning, Not a Superficial Bug

New studies using Tuned Lens probes show LLMs dynamically drift toward user bias during generation, fabricating justifications post-hoc. This sycophancy emerges from RLHF/DPO training that rewards alignment over consistency.

x.com/Mar 29, 2026/3 min read
large-language-modelsalignmentresearch
Scientist Nicholas Carlini points at a large screen displaying code and terminal output, with Anthropic Claude…
AI Research
97

Anthropic's Claude Discovers Zero-Day Vulnerabilities in Ghost CMS and Linux Kernel in Live Demo

Anthropic research scientist Nicholas Carlini demonstrated Claude autonomously finding and exploiting zero-day vulnerabilities in Ghost CMS and the Linux kernel within 90 minutes. The research has uncovered 500+ high-severity vulnerabilities using minimal scaffolding around the LLM.

x.com/Mar 29, 2026/3 min read
anthropicai securityvulnerability research
A bar chart comparing China and US first-author AI researchers in 2024, with China at 2,152 and US at 1,810
AI Research
87

China Surpasses US in AI Research Authorship with 2,152 First-Author Researchers in 2024

China now leads the US in first-author AI research contributions, with 2,152 researchers versus 1,810. This marks the first time China has overtaken the US in this key metric of research leadership.

x.com/Mar 29, 2026/3 min read
talentresearchanalysis
Dashboard showing automated LLM fine-tuning experiments with performance charts, a Red Hat logo, and a running…
AI Research
95

Fine-Tuning LLMs While You Sleep: How Autoresearch and Red Hat Training Hub Outperformed the HINT3 Benchmark

Automated fine-tuning tools now let you run hundreds of training experiments overnight for under $50. Here's how Autoresearch and Red Hat's platform outperformed HINT3, and the tools you can use today.

medium.com/Mar 29, 2026/3 min read/Widely Reported
enterprise-aillm-opsautomation
Two small cuboid robots with magnetic connectors on a white surface, one robot reaching toward the other to attach…
AI Research
87

Columbia's Truss Links Robots Self-Assemble and Cannibalize for Parts, Achieving 66.5% Mobility Gain

Columbia University researchers demonstrated 'Truss Links' robots that autonomously self-assemble using magnetic connectors, then selectively disassemble other robots to harvest parts for repair or growth. The system achieved a 66.5% mobility improvement through this zero-waste physical adaptation.

x.com/Mar 29, 2026/3 min read
roboticsresearchai

Linux Kernel Maintainer Linus Torvalds Reports AI-Generat…

AI Research
85

Linux Kernel Maintainer Linus Torvalds Reports AI-Generated Bug Reports Now Contain 'Actual Bugs' and Working Patches

Linus Torvalds, the lead maintainer of the Linux kernel, has stated that AI-generated bug reports are no longer 'slop' and now frequently identify real bugs with working patches. This marks a significant shift in the practical utility of AI for large-scale, complex software maintenance.

x.com/Mar 29, 2026/3 min read
open sourcesoftware engineeringsystems programming
Students using laptops with an AI tutor interface on screen, a teacher observing, and a graph showing improved test…
AI Research
85

Two Studies Find AI Tutors Improve Learning, While Unrestricted AI Use Can Shortcut It

New research shows AI systems prompted to act as tutors improve student learning outcomes, while simply giving students access to AI can lead them to accidentally shortcut the learning process.

x.com/Mar 29, 2026/3 min read
llm applicationsresearchai
A vintage leather-bound book lies open on a wooden desk beside a brass oil lamp, with a quill pen and inkwell…
AI Research
87

Researchers Train LLM from Scratch on 28,000 Victorian-Era Texts, Creating Historical Dialogue AI

Researchers have created a specialized LLM trained exclusively on 28,000 British texts from 1837-1899, enabling historically accurate Victorian-era dialogue generation. Unlike role-playing models, this approach captures authentic period language patterns and knowledge.

x.com/Mar 29, 2026/3 min read
natural language processinghistorical aispecialized models
Computer-generated diagram of a photonic crystal solar cell structure with layered geometric patterns, glowing…
AI Research
85

Kyushu University AI Model Achieves 44.4% Solar Cell Efficiency, Surpassing Theoretical SQ Limit

Researchers at Kyushu University used an AI-driven inverse design method to create a photonic crystal solar cell with 44.4% efficiency, exceeding the 33.7% Shockley-Queisser limit for single-junction cells.

x.com/Mar 28, 2026/3 min read
simulationai4scienceenergy
NVIDIA researchers present PivotRL diagram showing agent RL training cost reduction by 5.5x with performance…
AI Research
99

NVIDIA's PivotRL Cuts Agent RL Training Costs 5.5x, Matches Full RL Performance on SWE-Bench

NVIDIA researchers introduced PivotRL, a post-training method that achieves competitive agent performance with end-to-end RL while using 5.5x less wall-clock time. The framework identifies high-signal 'pivot' turns in existing trajectories, avoiding costly full rollouts.

x.com/Mar 28, 2026/3 min read
efficiencyagentsresearch
Moonshot AI CEO Yang Zhilin speaks at a tech event, gesturing while explaining attention residuals in large language…
AI Research
85

Moonshot AI CEO Yang Zhilin Advocates for Attention Residuals in LLM Architecture

Yang Zhilin, founder of Moonshot AI, argues for the architectural value of attention residuals in large language models. This technical perspective comes from the creator of the popular Kimi Chat model.

x.com/Mar 28, 2026/3 min read
architecturechinatransformer
A Google researcher points to a data flow diagram showing KV cache compression from 32-bit to 3-bit, with GPU memory…
AI Research
97

Google's TurboQuant Compresses LLM KV Cache 6x with Zero Accuracy Loss, Cutting GPU Memory by 80%

Google researchers introduced TurboQuant, a method that compresses LLM KV cache from 32-bit to 3-bit precision without accuracy degradation. This reduces GPU memory consumption by over 80% and speeds up inference 8x on H100 GPUs.

x.com/Mar 28, 2026/3 min read
quantizationresearchinference optimization
Three stylized robot heads with glowing circuitry face a central laptop displaying code, surrounded by floating…
AI Research
93

Open-Source Multi-Agent LLM System for Complex Software Engineering Tasks Released by Academic Consortium

A consortium of researchers from Stony Brook, CMU, Yale, UBC, and Fudan University has open-sourced a multi-agent LLM system specifically architected for complex software engineering. The release aims to provide a collaborative, modular framework for tackling tasks beyond single-agent capabilities.

x.com/Mar 28, 2026/3 min read/Multi-Source
open sourcesoftware engineeringmulti-agent systems
A group of researchers at MIT discuss a diagram showing multiple answer pathways branching from a language model…
AI Research
85

MIT Researchers Propose RL Training for Language Models to Output Multiple Plausible Answers

A new MIT paper argues RL should train LLMs to return several plausible answers instead of forcing a single guess. This addresses the problem of models being penalized for correct but non-standard reasoning.

x.com/Mar 28, 2026/3 min read
alignmentresearchreinforcement learning