LLMs & Generative AI - Page 17 of 55
Latest breakthroughs in large language models and generative AI shaping the future of artificial intelligence and machine learning.
1086 articles
View complete article list
BalCapRL adds length-based reward masking, boosting LLaVA-1.5-7B and Qwen2.5-VL
Dense image captioning hinges on a brutal trade-off: expert-quality annotations are gold, but their scarcity is a bottleneck.
SFT and RL Reweight Pretrained Distributions via Demonstration and Reward Signals
Forget the fighting words. Supervised Fine-Tuning and Reinforcement Learning aren't warring factions. They're the same basic math trick, just with different inputs. Both techniques take a trained language model and adjust the odds.
Spatial priming beats semantic prompting in chart data extraction study
What works better when asking a multimodal model to read a chart: high-level reasoning guidance, or a simple grid drawn over the image? The answer, at least for now, is emphatically the grid.
GraphDC Uses Divide‑and‑Conquer Agents to Scale Graph Reasoning
Graph reasoning is a mess. The bigger the problem gets, the more it breaks the tools we throw at it. A new system called GraphDC treats that sprawl like an engineering problem: you don't solve it, you split it up. GraphDC works like a surgical team.
RateQuant reveals mixed-precision KV cache pitfall: β decay rates span 3.6‑5.3
Mixed-precision quantization tries to save bits where it hurts accuracy the least. A new paper from RateQuant shows a critical flaw in that approach: the math used to decide where to cut those bits is not universal.
Top 10 2026 LLM Papers Highlight Pass@k Efficiency for Reasoning Models
In 2026, the Pass@k benchmark has shifted from a simple measure of accuracy to a key indicator of computational efficiency. It tracks how many attempts a model needs to produce one correct answer.
Generative AI fuels industrial-scale record 2025 data breaches, ITRC reports
Generative AI drove an unprecedented number of data breaches in 2025, according to a new report from the Identity Theft Resource Center. The ITRC tracked more breaches last year than in any previous period.
Strain drives exponential error growth; vorticity only linear impact
You have two ways to make an error in fluid simulation get bigger. One is an explosion. The other is a nudge. Strain and vorticity are both parts of a fluid's movement. They are not, however, equally guilty.
LKV learns head-wise budgets and token selection for LLM KV cache eviction
Getting AI models to pay attention to long conversations eats memory. A lot of it. The problem, known as the KV cache, grows with every word a model processes, and current fixes are blunt guesses about what to remember and what to forget.
LLM Summarizers Omit Identification, Distinguish Observed vs Inferred Claims
A summary that mistakes speculation for fact is not a summary, it’s fiction dressed in bullet points. The distinction between what was said, what was assumed, and what is merely recommended is not a nuance to fudge for readability.
NVIDIA's Star Elastic bundles 30B, 23B, 12B models; 23B hits 85.63 on AIME-2025
Most new AI models are just old ones with bigger numbers. NVIDIA's latest is the opposite: three models stuffed into one, and the middle one wins.
Understanding 'Compute': The Core Power Driving Modern AI Models
"Compute" is the single most expensive line item in artificial intelligence. It's not a metaphor. It's the literal electricity bill for thinking, the vast and physical cost of turning silicon into sense.
Fields Medalist: ChatGPT 5.5 Pro produced PhD-level math proof in under an hour
The clock had barely ticked past an hour when Timothy Gowers received the proof. A Fields Medalist, one of the most decorated mathematicians alive, had watched a machine do what takes a PhD candidate years.
Key Topics for LLM Engineers: Using Instruction Data to Align Models
A pretrained language model can generate coherent text. It doesn't, however, know how to answer a question directly or avoid harmful content without further training.
Semantic memory query retrieves Friday deployment approval for user-123
A simple query can ask an AI what it knows. Not just what it was told five minutes ago, but what it learned, decided, and filed away for later. This is memory, and it’s the difference between a helpful bot and a competent agent.
OpenAI launches Realtime‑Translate for 70+ languages and Realtime‑Whisper transcription
Voice agents have long been costly to run and tricky to orchestrate. The problem isn’t the models’ ability to converse; it’s the context ceilings that force engineers to build session resets, compress state, and reconstruct layers for every...
Google's Chrome 4GB on-device AI model unchanged, but explanation lacking
Four gigabytes. That’s the heft of Google’s on-device AI model, Gemini Nano, quietly sitting inside Chrome. It’s not new, it hasn’t grown, it hasn’t shrunk. Yet the real story isn’t the model itself, but the fog around who actually gets it.
RVPO boosts HealthBench score to 0.261, beating GDPO’s 0.215 at 14B (p < 0.001)
Apple researchers have a new answer for a stubborn AI alignment problem. Their method, called Risk-Sensitive Preference Optimization (RVPO), specifically targets a model’s tendency to ignore tough rules when simpler goals are up for grabs.
Adding Tools and Memory Expands AI Agent Threat Surface, Study Finds
A single-tool agent with no memory and no outbound actions has one vulnerability: the prompt surface. Give that agent persistent memory, now you’ve opened a second attack vector.
Grammar-Constrained Decoding Adds 495 Bash Tasks, Some Regressions Seen
A method for making code-writing AIs more reliable just added 495 working Bash scripts to its total. It also broke a few that used to work. The technique is called grammar-constrained decoding.
Featured Resources & Reviews
No Code MBA Course Review
Learn to build AI-powered apps without coding. Our comprehensive review of No Code MBA's course.
AI Tools & Resources
Curated collection of AI tools, courses, and frameworks to accelerate your AI journey.
Weekly AI Digest
Get the week's most important AI news delivered to your inbox every week.