All published pieces, newest first. Filter by topic.
Two H100s on different clouds gave bit-identical Evo 2 scores for 3,893 BRCA1 variants. An L40S running the same container reshuffled two thirds of the ranking. Running the model in FP32 brought the two GPUs back into agreement, at three to six times the compute.
Labelling cells in a new single-cell study often means training a method on cells that already carry labels. A recorded benchmark, not peer reviewed, trained six methods on one reference, ran two released models, and tested them on four blood studies kept out of training (32 donors). The benchmark could not establish a clear difference among four learned methods in agreement with author labels. At the threshold chosen on other studies to label 90% of cells, no method labelled 90% of the pooled test cells.
You have a spectrum, 250 possible molecules and a budget to check five. Trained alignment models found the true structure about six times in ten in this benchmark, but even a high score could not reliably tell when the answer was missing.
A model can put proteins in a useful order while barely improving its predictions. Six rhodopsin models show why error in nanometres matters alongside rank correlation.
If you can test 100 variants in a splicing reporter, which score should rank them? On 8,297 held-out MFASS variants, a simple trained baseline put 61 splice-disrupting variants in its first 100, and SpliceAI and Pangolin put 63 to 66. The difference was not established. Pangolin ranked the whole list better in exploratory analyses.
A small CNN cut prediction error by 28% against cheap features on unseen sequence families. For choosing the top 1% of sequences, the two were indistinguishable, and frozen UTR-LM embeddings added nothing.
In this exploratory study, with annotation, reference genome and variants matched, SpliceAI and Pangolin put 63 to 66 assay-confirmed splice disruptions in their top 100 of the same 8,297 held-out MFASS variants. Masking raised average precision slightly for both tools; no top-100 difference was established.
Genomics MCP is a local MCP server that gives AI agents bounded, provenance-tracked access to real genomic data: discover public files, then read bounded regions from bigWig, BAM, CRAM and VCF. Includes a case study where Claude used only its tools to compare BCL11A enhancer signal across 26 public ENCODE files.
A recall reversal and two constant-prediction controls show why benchmark comparisons need populations, selection rules and reference scores.
The baseline, SpliceAI and Pangolin found 61, 64 and 65 disruptions in their top 100 MFASS variants. The differences do not establish a reliable winner.
RNA models share an alphabet, not an estimand: choose them by molecular object, output, information inputs, evidence class, and artifact contract.
A reported genomic-model comparison is auditable only when its tokenisation, strand rule, biological target, evaluation split, and artifact terms are explicit.
Protein generators propose different biological objects. A defensible design starts with the assay, records the complete computational stack, and preserves every experimental denominator.
A practical guide to choosing, extracting, adapting and validating protein language-model representations without mistaking a system score for biological generalisation.
A task-first guide to antibody representations, structure prediction, CDR design, humanisation and developability, with the evidence and reproducibility checks that model scores leave out.
Suppose somebody hands you a FASTA file and asks for “the structure.” First decide whether the target is one chain, a protein assembly or a mixed complex containing ligands or nucleic acids.
One pretrained genomic model can look dominant on a leaderboard and lose to a linear baseline on the next task. Two evidence ledgers explain why.
Carbon-3B matches Evo2-7B on sequence recovery, variant-effect prediction, and motif-perturbation discrimination while generating DNA over 150 times faster.
A 1-billion-parameter model conditioned on RNA chemical-probing reactivity folds a held-out transcript to an F1 of 0.987 against an experimentally guided reference.
ESMFold predicts atomic-level structure from a single sequence. No multiple sequence alignment, no database search, no Evoformer churning over homologs.
Reinforcement learning works when you can check the answer. A chess engine wins or loses. A code-generation model passes or fails the test suite.
Vision-language models can describe a scene in paragraph-length detail and still fail to tell you whether a red cube is in front of or behind a blue cylinder.
How do you verify scientific knowledge when there are 2.9 million papers on arXiv alone, with thousands more added every day? A new paper extracts nearly two million claims from 16,087 manuscripts and compares machine evaluation to human peer review, with 81% agreement.
Over 80% of developers now use AI assistants for code generation, yet at least 62% of AI-generated code contains vulnerabilities. As AI writes code faster than humans can review it, the engineer's primary job shifts from writing code to verifying it through formal methods.
Structure-based and ligand-based drug design evolved as separate fields solving the same problem. ConGLUDe, a contrastive geometric learning model, unifies both approaches and outperforms specialist methods on realistic benchmarks without requiring pre-defined binding pockets.
Google DeepMind's AlphaGenome reads 1 million base pairs of DNA and predicts thousands of regulatory functions at single-nucleotide resolution, beating 25 of 26 specialized models.
AlphaGenome processes 1 million DNA base pairs to predict variant effects across 7,000+ genomic tracks in one second, outperforming specialized models on 25 of 26 VEP benchmarks.
A technical look at AlphaGenome's architecture, its 2D pairwise embeddings for splicing prediction, and what the model means for clinical variant interpretation.
Basecamp Research's EDEN model trains on proprietary environmental metagenomics to design gene-insertion enzymes, antimicrobial peptides, and synthetic microbiomes -- all validated in the wet lab.
A practical guide to selecting genomic foundation models for bioinformatics tasks. Covers ESM-2, DNABERT-2, HyenaDNA, Nucleotide Transformer, scGPT, and Evo with scoped comparisons for DNA, proteins and single-cell analysis, with corrected references and clear distinctions between frozen representations and trained predictors.
How TTT-E2E achieves constant inference latency regardless of context length by treating long context as a learning problem rather than an architecture problem.
A technical deep dive into DeepSeek's Engram architecture, which introduces conditional memory as a new axis of sparsity for large language models.
How the same transformer architecture powering GPT learned to predict molecular properties by treating chemistry as a language problem
How researchers adapted BERT for molecular property prediction, turning SMILES strings into drug discovery insights
How generative diffusion models can serve as fast surrogates for expensive biological simulations, achieving 22x speedup while preserving the stochastic diversity that makes these models scientifically useful.
Machine learning models now outperform FDA-approved biomarkers in predicting treatment response, but the best-performing models often resist explanation. Here's how precision oncology is navigating the trade-off between performance and interpretability.
Recent research suggests RL training optimizes search efficiency over existing capabilities rather than expanding reasoning capacity. Here's what the pass@k evidence actually shows.
VL-JEPA changes the prediction target and separates visual inference from text decoding. Here is how its objective differs from token prediction and latent diffusion, and what its efficiency results establish.
How Meta's Code World Model applies the "dreaming car" insight from robotics to software engineering, achieving 65.8% on SWE-bench Verified by training on execution traces rather than static code.
Why computational biologists should stop building embeddings and start building simulators, with three tractable project ideas you can implement today using flow matching, Neural ODEs, and cell fate trajectory modeling.
Understanding when you're working with an environment versus a benchmark changes how you design experiments, interpret results, and communicate findings. This guide covers the practical differences every RL practitioner should know.
Recent research shows 1024-layer networks achieve 2x to 50x improvements in goal-conditioned RL. Here's why extreme depth works now, and when you should consider it for your own agents.
A comparative analysis of approaches to catastrophic forgetting in language models, from parameter regularization to sparse memory architectures that reduce forgetting from 89% to just 11%.
DeepMind's DiscoRL discovers reinforcement learning algorithms that outperform hand-designed methods like PPO and DQN. By treating algorithm design as a meta-learning problem, it found alternatives to value functions and bootstrapping through optimization alone.
Are large language models merely stochastic parrots, or do they develop genuine internal representations of the world? This investigation examines evidence from Othello-GPT, spatial encoding in LLMs, and the symbol grounding problem to explore what cognitive science reveals about AI understanding.
Mixtral uses 46.7B parameters but only activates 13B per token. This architectural trick called Mixture of Experts powers Gemini 1.5, DeepSeek V3, and more. Learn how MoE works, its hidden costs, and when to use it.
Pedro Domingos proposes that neural networks and symbolic AI are the same mathematical operation - a logical rule can be equivalently written as a tensor equation in Einstein summation notation. If true, we've been building separate tools for problems that share identical structure.
GPT-4's 128K context window? It only uses about 10% effectively. Google's TITANS architecture introduces test-time memory learning that outperforms GPT-4 on long-context tasks with 70x fewer parameters.
Why biological systems offer the ideal training ground for reinforcement learning: automated verification through physics, not human judgment. From protein design with AlphaFold to RNA folding with ViennaRNA, biology provides the verifiable inverse problems that RL needs at scale.
Machine learning is illuminating biology's hidden half: intrinsically disordered proteins and RNA structures that traditional methods could never capture. But can we trust what we're seeing?
World models enable AI agents to imagine futures and plan actions, achieving 10-100x better sample efficiency than traditional reinforcement learning. From DreamerV3 collecting diamonds in Minecraft to foundation models like Sora and Genie, world models represent AI's shift from pattern matching to simulating reality itself.
MIT researchers used generative AI to create novel antibiotics from scratch, not finding them, but designing them. The breakthrough matters less for what it creates than for what it proves.
Learn how DSPy transforms brittle prompt engineering into systematic, testable code with automatic optimization
Nested Learning: The Illusion of Deep Learning Architectures - A comprehensive guide to the arXiv paper revealing how neural networks learn at multiple timescales through hierarchical optimization.
Neural networks catastrophically forget previous knowledge when learning new tasks—not due to capacity limits but fundamental constraints in distributed learning systems.
Google DeepMind's AlphaEvolve broke a 56-year-old matrix multiplication record and matched or beat human solutions on 95% of 67 mathematical problems.
Are LLMs truly exhibiting emergent capabilities, or are we mistaking measurement artifacts for genuine phase transitions?
How gene length, trait specificity, and luck systematically distort which genes we think are important for disease, and what researchers should do about it
An AI research system maintains coherent reasoning across 200+ agent steps over 12 hours, generating 42,000 lines of code while reviewing 1,500 papers. Kosmos achieves 79.4% accuracy through structured world models and parallel agents, but verification remains humanity's bottleneck.
New research reveals a disturbing truth: reasoning models maintain high performance until they hit a complexity threshold, then collapse entirely. They don't degrade gracefully - they fall off a cliff.
Why does it feel like our tools weren't designed by pathologists? Billions poured into AI models that compress whole slide images into tiny vectors, ignoring how pathologists actually examine tissue. The evidence reveals why scaling won't fix this disconnect.
The definitive guide to building AI agents that monitor their own thinking through dual-loop architecture, Reflexion patterns, and production-ready LangGraph implementations.
An AI system at Google DeepMind discovered how bacteria share genes across species barriers using 7 days of computational reasoning. When tested, it matched unpublished experimental observations exactly.
Traditional readability formulas miss the mark. Modern embedding models can capture semantic nuance and syntactic structure, but do they actually predict complexity better?
Discover why monolithic embeddings fail for RAG systems and learn how chunking strategies can transform your retrieval performance.
Learn the RATS pattern: Retrieval-Augmented Tool Selection for building scalable AI agents that dynamically select the right tools for each task.
Language models can write poetry but struggle with basic spelling. Discover why current tokenization breaks language, and how morphology-aware approaches fix it.
