Scientific AI and medical ML, methods, evaluation, clinical systems. Primary sources and caveats. Not another chat-demo headline factory.
Beats
Plain-language explainers, with doors into the papers if you want them.
Models, training tricks, representations, what changed in the paper.
Imaging, reports, clinical decision support, with measured limits.
From cybernetics to modern training, the scientific serial.
Easy reading
All explainersA plain-language guide to the two AI bills: training as an expensive practice burst that updates weights, inference as the ongoing cost of running the fixed model: with Luccioni et al. on BLOOM as the door into why the split matters.
A plain-language guide to context windows as the token budget a model can see at once: prompt, chat history, and retrieved text on one finite desk, with Vaswani et al. as the door and honest limits on long memory, missed middles, and cost.
A plain-language guide to tokens as text chunks a language model works with: why systems split rare words into pieces, how BPE and subword units fix a fixed vocabulary, and why context windows count tokens, not pages.
A plain-language guide to retrieval-augmented generation: look up relevant passages first, then generate with that context in the prompt, not magic memory: with Lewis et al. as the door, and honest limits on bad retrieval, wrong citations, and stale docs.
History
All chaptersBefore the name 'artificial intelligence,' research programs already treated mind and machine as the same kind of problem-logical neurons, feedback control, and operational tests of machine intelligence.
The August 1955 proposal and the 1956 Dartmouth workshop named and organized a research program under the phrase 'artificial intelligence'-they did not invent the problems surveyed in the previous chapter.
Rosenblatt’s perceptron made learning in a neuron-like device an empirical research program; Minsky and Papert’s 1969 book proved sharp limits of a restricted class of machines-not that ‘neural nets are dead.’
After Dartmouth and alongside the perceptron line, a large part of AI became knowledge plus search-GPS, DENDRAL, MYCIN, and the expert-system boom-while the journalistic label “AI winter” names real but local funding contractions, not a single global morality play.
In the 1980s, distributed representations returned as an empirical research program-Hopfield networks, Boltzmann machines, and the PDP volumes-with new mathematics of energy, attractors, and hidden units, not a magical rebirth after a total death of neural nets.
Backpropagation-efficient reverse-mode differentiation through multilayer nets-made hidden-unit learning a community tool in 1986, after an earlier paper trail in automatic differentiation and ordered derivatives that was not invented out of nowhere that year.
In the 1990s, VC theory, soft-margin support-vector machines, and kernel methods offered a statistical-learning alternative to under-regularized multilayer nets-strong baselines with clearer capacity control, not a permanent replacement for neural networks.
Gradient descent, stochastic approximation, momentum, regularization, and Adam form a cumulative training-methods line-engineering and theory, not a folklore of sudden deep-learning inventions.
Before transformers, language and other sequences became trainable transduction problems through simple recurrent nets, LSTM, encoder–decoder seq2seq, and Bahdanau’s soft alignment-attention as an RNN add-on, not a 2017 invention.
What moved in 2012 was not a sudden invention of deep nets: ImageNet-scale labeled data, GPU training, ReLU, and dropout made a large ConvNet win a public, comparable benchmark-AlexNet as the ImageNet moment, not the first deep network.
Supervision is scarce: the modern stack learns reusable representations from large unlabeled or weakly labeled data-layerwise pretraining, ImageNet transfer, word2vec, BERT-style masked LMs, and contrastive vision-then adapts them, without collapsing the lineage into a BERT founding myth.
The 2017 transformer reorganized sequence modeling around self-attention without recurrence; GPT and BERT are two objectives on that backbone; Kaplan and Chinchilla scaling laws are empirical regularities about compute, data, and parameters-not a slogan that bigger is always better.
Latest briefs
Zhong et al. benchmark frontier agents across 20 scientific ML tasks, revealing that autonomous performance gains stem from routine engineering tricks rather than domain insight.
Vaidya et al. introduce MedPAO-VL, replacing opaque end-to-end chest X-ray report generation with a plan-act-observe loop that iteratively reconciles findings against visual evidence.
Bhandari et al. introduce ReadKV, storing KV caches in progressive codes to fetch precision selectively based on query attention, cutting read bandwidth by half with minimal perplexity impact.
Zhang et al. reveal that lossy speculative decoding causes jailbreak vulnerabilities to spike far faster than utility drops, introducing SecureSD with early-token stricter verification.
Myrzakhan and Shen introduce dSpark, a schedule distillation framework for diffusion LLMs that prioritizes tokens resolving global sequence uncertainty, achieving up to 10.12x speedups on reasoning benchmarks.
Ma et al. present MedVTok, a unified medical visual tokenizer pretrained on 28 million images and 70,000 volumes across nine modalities, using slice differential regularization to retain anatomical coherence.
Zhang et al. introduce NPC-RubricBench, evaluating 30 language models across longitudinal oncology care pathways and exposing critical failure mode shifts from diagnostic commission to treatment omission.
Moya et al. show that augmenting preference datasets with targeted ties mitigates shortcut learning, deriving influence bounds and using implicit reward margins to identify optimal pairs.
Huang et al. introduce HAD-VLM, a unified vision-language framework combining autoregressive text with continuous diffusion prediction to eliminate quantization errors in clinical ultrasound and X-ray measurements.
Ke, Cai, and Wang reveal that training instability in decoupled LLM RL pipelines stems from subtle rounding-location discrepancies across inference and training engines, demonstrating that targeted FP32 retention in RoPE, SwiGLU, and RMSNorm prevents policy collapse.
Masiak et al. introduce TopKLoRA, a parameter-efficient fine-tuning method that enforces sparsity over adapter latents to expose causal circuits and enable surgical removal of sleeper-agent backdoors.
Rogers, Lin, and Brown apply sparse autoencoders to OpenVLA activations, identifying interpretable latent features that predict manipulation failures before irreversible physical errors occur.
Combining stochastic depth pretraining with the BeNice post-training broadcast enables looped transformers to share KV caches across recurrences, boosting throughput by 24% and cutting p90 TTFT by 84%.
Gibbs, Han, Tibshirani, and Zrnic develop an online selective auditing framework that provably bounds unaudited AI decision errors below user tolerances under both i.i.d. and adversarial streams.
Barron, Gnanapragasam, and Tschannen uncover two pervasive numerical bugs in mainstream softmax implementations, presenting stable fixes that eliminate the slingshot effect and prevent distillation drift.
Evaluating 8B-27B open models on 1,552 official residency exam questions shows that while raw scores qualify models for 43 specialties, medical fine-tuning harms calibration and multimodal models ignore clinical images.
Evaluating six 7-9B models across six categories of clinical harm reveals that risk-bearing language is statistically independent of factuality, with domain adaptation suppressing drug mentions while amplifying overconfidence.
Mathematical analysis proves that auxiliary performance predictors outperform an LLM's internal confidence only when candidate models are suboptimal, revealing structural ceilings for model cascading.
Across 27 frontier models, conversational health assistants frequently validate mistaken user self-diagnoses instead of issuing critical clinical warnings when life-threatening conditions remain possible.
Extending the Platonic Representation Hypothesis to physical kinematics, the training-free S^2 metric evaluates vision-language backbones and predicts closed-loop manipulation success without expensive rollouts.
By pairing noisy-OR pooling for supported findings with log-sum-exp suppression for unsupported alternatives, Diagnostic Loss eliminates spurious hallucinations without altering model architectures.
Framing prefix thought pruning as a submodular pass@k objective allows reasoning models to discard unpromising rollouts early while preventing redundant candidate clustering.
A benchmark of 300 matched task pairs reveals that tool-using agents routinely execute prohibited actions when environment tools omit constraints, with passive refusal rarely translating into active evidence recovery.
Decoupling sparse attention indexing from physical memory locations, UlaCache delivers up to 2.51x decoding throughput speedups on NVIDIA Blackwell by combining parallel bucketed LRU eviction with zero-overhead buffer retention.
A standardized benchmark spanning 13 imaging sources and 2,000 compounds shows that while test-time adaptation mitigates domain shift, AI models degrade to chance level on rare phenotypic profiles.
Evaluating cross-session payload retention across frontier models reveals that malicious prompts planted in agent memory and heartbeat files re-fire without decay, bypassing conventional safeguards.
An empirical audit of 3D shape completion foundation models reveals that despite visual plausibility, zero-shot reconstructions penetrate patient bone and achieve a 0% clinical implantability rate.
By remapping tensor layouts directly during RDMA transfers, HAWK disaggregates GPU prefill and NPU decode to boost serving power efficiency by 1.26x and cost efficiency by 1.72x.
Rather than inspecting isolated agent actions, DART tracks accumulated representation shifts across dialogue turns, cutting multi-turn agent attack success from 84% to 25% with minimal latency overhead.
By reusing target model hidden states and eliminating the draft model's expanding historical KV cache, DLite overcomes speculative decoding degradation in multi-turn agentic environments.
By embedding physical voxel dimensions and geometric constraints directly into spatial embeddings and attention, MAGEFormer eliminates resampling artifacts and sets new boundary accuracy baselines on abdominal CT.
By applying mechanistic interpretability to chest X-ray classifiers, researchers identified dedicated attention heads that latch onto corner text rather than lung pathology, a bias completely invisible to standard output ablation.
Extending xLSTM with rank-1 Kronecker delta updates produces an associative memory that provably converges as a steady-state Markov chain with uniformly bounded norm, outperforming DeltaNet on long sequences.
Systematic ablations across public benchmarks demonstrate that while stage-redacted reports add genuine prognostic survival signal, slide-level VQA is ungrounded (74.8% accuracy with slide vs 74.4% without) and zero-shot prompts do not outperform linear probes.
Combining transcript distillation, counter-prompting, and tamper resistance produces the first ELK model organisms that reliably sandbag misaligned code audits while retaining full latent capabilities.
By re-weighting attention logits with token importance scores, TokenGate enables end-to-end learnable token pruning over long-context inputs, reducing peak memory by 1.5x to 3.4x with zero accuracy loss.
Comparing nine imputation methods on 40,336 ICU records reveals that pipelines with indistinguishable AUROC produce divergent high-risk patient alert lists, with Jaccard overlaps as low as 0.518.
Attenuating high-evidence spatial nodes and calibrating predicted grade distributions based on counterfactual response stability delivers robust single-image test-time adaptation across five clinical fundus datasets without model updates.
Iteratively training language models on their own safe responses paired with harmless contextual cues causes severe downstream safety degradation on standard jailbreak benchmarks.
By training a 0.62M-parameter drafter to predict position commit trajectories and introducing coupled-noise match verification, SPOT accelerates diffusion LLMs without quality loss.
A 300-case evaluation of role-specific virtual specialist consiliums shows structured multidisciplinary synthesis consistently outperforms standalone opinions, with GPT-5.6-Sol leading overall clinical quality and consensus speed.
With only 10,550 parameters per model, Evidence Field couples spatial image measurements with replayable covering programs, surpassing standard CNN and ViT baselines across held-out hospital sites on CAMELYON17-WILDS.
Reusing internal activations to detect uncertainty without extra forward passes lets lightweight agents escalate to frontier models only when needed and monitor them afterward, cutting OSWorld inference costs by 19%.
A combination of hardware-model co-design, analytical zero-warmup kernel selection, and agentic code translation yields up to 1.5x throughput gains for recommendation model training on AMD Instinct MI350X accelerators.
Fused Triton autograd kernels and a host-sync-free preallocated buffer compress sparse ReLU and ReLU^2 activation caches losslessly, saving nearly a third of per-token VRAM with under 1% overhead.
Training LLMs away from dishonest completions using continuous adversarial objectives with on-policy utility data curbs pressure-induced lying and increases confession rates without degrading general reasoning.
Conditioning lifespan normative models of cortical thickness on subcortical anatomy enables site calibration with 20 scans and reveals structural dissociations predictive of conversion to dementia.
Formulating adaptive therapy through robust contextual reinforcement learning yields treatment schedules that Pareto-dominate clinical heuristics and maintain efficacy despite noisy biomarker measurements.
Lifting verbatim context sentences into interleaved safety data prevents models from keying harmful behaviors on domain framing, slashing conditional misalignment on Qwen2.5-7B by up to 78%.
Modeling dominant-label pathology annotations via right-censored likelihood and severity-ordered residual routing preserves subtle secondary lesion findings suppressed by standard multiclass MIL.
Profiling remote KV caching for GLM-5.3 with MLA on 8xB300 nodes reveals cache hits were slower than recomputing; rank-partitioned sharding and async loading drop p99 tail latency from 83s to 0.67s.
Pretrained on synthetic datasets with a flexible proportional-odds prior and quantile regression loss, SurvICL outperforms 20 baselines and tops the 50-year-old Cox model on small cohorts.
Evaluating five frozen neuroimaging foundation encoders on longitudinal cohorts demonstrates that released embeddings permit up to 99% rank-1 patient re-identification and over 90% recorded-sex inference.
Weighting token credit via causal-versus-hindsight Jensen-Shannon divergence matches GRPO reasoning accuracy while eliminating zero-gradient batch waste and cutting rollout requirements eightfold.
Coupling a 3D rectified flow vascular refiner with a slice-wise diffusion decoder resolves vessel hallucination and fragmentation, reducing disconnected vascular components by over 20%.
An audit of the top 200 Hugging Face GGUF repositories reveals an Independent Verifiability Rate of 0/198 under standard disclosures, showing hidden compiler flags and per-tensor quantization recipes are required for bit-for-bit verification.
Combining frozen 3D foundation model features with GPU-accelerated iterative optimization and lesion-volume preservation wins the Learn2Reg 2026 PSMAReg Challenge with zero task-specific training.
Counterfactual interventions across Qwen-3 and Gemma-4 show that attributes driving 19 to 24% of correct medical diagnoses go unmentioned in reasoning traces that models cannot solve without CoT.
An audit across 1,112 ABIDE I brain scans and seven harmonization methods shows frozen site classifiers swing by 46 points across seeds, fall to chance under simple blur, and miss residual scanner signals.
Staging file writes and network requests at a microVM boundary with speculative execution cuts AgentDojo prompt-injection success from 47.0% to 2.7% without degrading benign utility.
Square-root learning-rate scaling in multi-step GRPO preserves policy drift per sample rather than progress, causing larger batches to lose sample efficiency to curvature while collapsing at the same sample count.
A meta-controller trained via Monte Carlo control over compact workflow states recovers 44 to 81% of the executor-advisor gap at under half the advisor cost.
Conditioning on paired text and 1.8M structures yields 15 of 18 property wins, higher binding confidence on 13 of 20 targets, and a 107 nM wet-lab binder.
Digital locker circuits embedded in neural networks resist elicitation through fine-tuning, weight noising, and gradient trigger search under cryptographic assumptions.
Replacing uniform random cropping with coarse-to-fine anatomical probes eliminates observation waste and matches full-training transfer at 48% of SSL epochs.
Weighted conformal prediction with directional, class-specific bounds keeps nominal 90% coverage on BraTS 2020 while narrowing intervals by 8 to 14%.
A multi-agent framework derives the benchmark specification itself, reaching 90.9% Task-Space F1 and 994 of 1,000 sampled items passing human audit.
Diffusion-based detectors reach only 0.56 to 0.72 pixel-level AUROC and PatchCore variants 0.76 to 0.83 when anatomy, views, and protocols vary.
TAP-JEPA, a ViT-B/3D trained on 113,611 CT volumes, is audited to show that masked prediction learns anatomy well but pulls sporadic pathology toward normal.
Dutta et al. couple 458 DNABERT models with deterministic biological verifiers, preventing LLM hallucinations when prioritizing noncoding single nucleotide variants.
He et al. introduce bidirectional closed-form continuous-time modules to model irregular cross-slice medical imaging, outperforming RWKV and GRU baselines with lower latency and peak memory.
Azizi et al. audit typed decision models used as agent guardrails, showing how semantic manipulation of option labels drives fail-open rates up to 100%.
Wang et al. resolve distribution shifts in Hessian-guided LLM pruning by calibrating during autoregressive decoding, pairing it with an N:M sparse matrix-vector CUDA kernel.
Hollinsworth et al. introduce representation probes trained on the FIBS dataset to detect latent sabotage and introspective deception in frontier LLM agents without relying on verbalized reasoning.
Lee et al. introduce Missing Modality-Aware Local Temperature Scaling, calibrating voxel-wise brain tumor segmentation confidences when MRI acquisition sequences are incomplete.
Cheng et al. evaluate nine computational pathology vision-language models across pathologist-validated phrasing, uncovering high sensitivity to clinical paraphrases and diagnostic competition.
Li et al. identify preconditioner distortion in 4-bit AdamW quantization, proposing ZIP-SR and ZE-EDEN to cut validation loss gaps to FP32 by up to 70%.
Ma et al. introduce trajectory-based reinforcement learning for clinical decision making, finding that selective diagnostic acquisition outperforms full-modality inputs.
A coordinate-based multimodal fusion architecture that maps available clinical, imaging, and fluid biomarkers into an interpretable RPNS space without synthetic data imputation.
A theoretically grounded recalibration framework by Dandapanthula et al. using envelope sampling to provably bound policy regret when optimizing LLMs against surrogate reward models.
Nguyen and Fithian re-evaluate METR’s 50% time horizon metric with splines and item-response theory, revealing that task difficulty does not scale linearly with human log-time.
A post-hoc calibration technique using intra-slide internal controls to align gigapixel pathology embeddings across scanner hardware without retraining the backbone.
A theoretical foundation for test-time reasoning: proving vanilla beam search sample bounds and establishing polynomial regret scaling over long horizons via CF-Beam.
Model-agnostic calibration for proprietary 3D CT models using test-time geometric perturbations and confidence-gated blending, cutting calibration error up to 54%.
Extending medical triage beyond binary autonomous vs defer decisions: a framework that routes cases to autonomous prediction, human clinician, or human with AI guidance.
A medical vision-language model trained on ~40k cases to classify all 12 WHO 2021 brain tumor types, with uncertainty and report generation, plus multi-reader and prospective checks.
A training-free agent that starts from sparse HPO-like phenotypes, keeps a probabilistic disease ranking, and picks informative follow-ups, with large Recall@1/5 gains on four cohorts.
Stanford researchers pair slow pretrained VLA proposals with high-frequency RL action edits to solve the latency lag on real robots, jumping from 42% to 97% success in 10 minutes.
MIT and CMU researchers distill task-salient state into a compact workspace token during training, giving manipulation robots long-term memory at zero deployment compute overhead.
DeepMind’s follow-on to AlphaGenome: a 1 PB catalogue of single-nucleotide variant molecular predictions, plus an AVI ranking score that blends regulatory and missense signals for research use.
DeepMind’s Nature paper on a unifying DNA model: 1 Mb context, base-resolution multimodal tracks, and strong regulatory variant-effect scores, the positive-science sequel to AlphaFold, with clear limits.
A methods brief on affordable DiT fine-tuning for depth (and other dense maps): single-step inference, Sinkhorn sharpening, and 16–26% AbsRel gains on KITTI/ETH3D versus prior best.
A scientific-methods read of OpenAI’s 8 Sep 2026 announcement: a multi-agent system plus Lean formalization arguing finite-time singularity (Clay C/D), with verification still on mathematicians, not a prize ceremony.
A medical safety/eval note: three-way retain/suppress/defer triage on VTaC with a finite-sample style bound on silenced true alarms, and a clear warning about multiplicity eating the certificate.
An eval methods brief: LoRA-tuned PrecepTron plus GRAND-ROUNDS (9,217 physician scores) aim for physician-level grading of open-ended clinical LLM answers, including piecemeal case presentation.
A stopping layer that tries to decide when a sequential diagnosis agent should diagnose or defer: with finite-sample style tests, and an honest ‘exploratory, not confirmatory’ framing.
A medical imaging methods note: small pulmonary nodules show a Gaussian radial intensity prior; GRIPNet designs every module around that physics and reports strong mAP@0.5 on three public CT benchmarks.
An 85-task infra/eval benchmark spanning kernels, long-horizon repo work, and end-to-end optimization: topped by Claude Opus 5 at 36.53%, with a lot of headroom left.
A scientific-discovery methods paper that treats ‘how many independent laws?’ as the zeroth step: then shows multi-law fits beating single-equation symbolic regression on engineering and galactic data.
A methods brief on patient-specific EHR concept encoding: RL picks how much KG context each code gets, a GNN + frozen LLM refine representations, reported gains on MIMIC-III/IV backbones.
Self-hosted Llama extracts structured T-stage from radiology reports with source-text links: 90% accuracy vs a four-expert reference on 130 reports.
A living eval suite of 43 biomedical HDLSS datasets across 28 feature×sample operating points: tabular foundation models lead at the reference cell, but logistic regression stays uncomfortably close.
Multicenter retrospective and prospective evaluation of an AI-assisted clinical trial recommender: ~91% top-10 hit rate vs clinician picks, 55% less screening time, and more trial access in a live tumor board.
AI is more than a chat box on the web. Bitware News covers what the technology actually does: methods, medical ML, evaluation: with primary sources and caveats.