Why AI Scientists Don't Do Science: Inside Corral's 25,000-Run Epistemic Audit
A landmark study across 8 scientific domains reveals LLM agents ignore evidence 68% of the time, and scaffolding accounts for only 1.5% of variance over base model reaso…
The day's biggest AI story, explained at length with its sources. Page 1 of 12.
At $0.10 per million tokens and 370+ tokens/sec, Meta's Muse Ember changes the economics of agentic loops, proving frontier planners need cheap, multimodal executors to scale.
A landmark study across 8 scientific domains reveals LLM agents ignore evidence 68% of the time, and scaffolding accounts for only 1.5% of variance over base model reaso…
Researchers from UW and Microsoft discover that RLVR on a single problem unlocks frontier reasoning, revealing the strange physics of post-saturation generalization.
A technical deep dive into retrieval-augmented generation: chunking mechanics, dense and sparse retrieval, cross-encoder reranking, and where production pipelines break.
Samsung SAIL's TRM strips away billions of parameters and verbal Chain-of-Thought, matching frontier models on ARC-AGI using a 2-layer network with recursive latent upda…
By replacing static BPE tokenizers with entropy-driven dynamic patching, Meta's BLT matches Llama 3 at 8B parameters while halving inference compute and curing tokenizer…
Salesforce AI Research treats reasoning as a latent variable, enabling LLMs to self-reward and outperform supervised fine-tuning using only question-answer pairs.
Bologna and CINECA researchers unveil an analytical law for inference energy, proving why reasoning models and agentic workflows trigger super-linear energy blowups.
Tsinghua researchers decouple teacher supervision from token probability distributions, enabling black-box agent distillation across different model families and tokeniz…
By turning black-box game interaction into category-theoretic rules and Lean 4 proofs, AWM proves impossibility, boosts Opus to 99.9 on ARC-AGI-3, and slashes actions by…
Understand the architectural differences between raw LLM function calling and the Model Context Protocol, including token overhead, progressive discovery, and selection…
Microsoft Research's Diff Transformer subtracts attention noise, matching standard LLM performance with 65% of the compute while neutralizing activation outliers.