Scientific Discovery & Drug Design - Weekly Roundup

This work uses the Headroom-Closed Index (HCI) to reveal the problems of existing LLMs, and introduces the RSI concept and its development roadmap: from improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement.

T1, a Mixture-of-Experts model of 122B total trained with reinforcement learning, operating a real shell in a cloud sandbox for up to 300+ tool-call turns per task, rewarded by executing each task's own verifier by executing each task's own verifier.

Textual descriptions of reaction conditions are encoded by a fine-tuned language model trained jointly with Gaussian process surrogates, yielding task-adaptive representations within a multi-objective Bayesian optimisation loop.

An amortized neural sampler that combines operator learning with flow methods for sampling, enabling efficient sampling across families of stochastic differential equations, and theoretically establishes the expressivity and resolution invariance of the framework.

T1, a Mixture-of-Experts model of 122B total trained with reinforcement learning, operating a real shell in a cloud sandbox for up to 300+ tool-call turns per task, rewarded by executing each task's own verifier by executing each task's own verifier.

Textual descriptions of reaction conditions are encoded by a fine-tuned language model trained jointly with Gaussian process surrogates, yielding task-adaptive representations within a multi-objective Bayesian optimisation loop.

An amortized neural sampler that combines operator learning with flow methods for sampling, enabling efficient sampling across families of stochastic differential equations, and theoretically establishes the expressivity and resolution invariance of the framework.

A quantum approach to spectral feature extraction from the density of states (DOS) of a problem-dependent Hamiltonian is developed, and applied to machine learning on signed graphs, and opens routes towards quantum applications in social network balance analysis, spin-glass studies, correlation clustering, and protein-interaction networks.

A modular agentic-AI platform that converts a heterogeneous corpus of process-development documents into a queryable, dual-layer knowledge graph is presented and it is anticipated graph-based architectures will see broader adoption in pharma as a means of transforming fragmented document repositories into structured process intelligence.

CausalArena is introduced, a unified and evolvable benchmark for causal discovery under a common protocol, and substantial ranking shifts across SCM families and protocols are revealed, showing that strong performance in one benchmark regime does not reliably transfer to others.

A Mobius layer is proposed, with rational functions replacing matrix operations, and the recently discovered activation function eml(x,1/x), which allows exp(x) and ln(x) to be recovered separately, and hence all elementary functions to be evaluated within a rational generalization of the neural network.

This work uses the Headroom-Closed Index (HCI) to reveal the problems of existing LLMs, and introduces the RSI concept and its development roadmap: from improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement.

Predicting blood-brain barrier permeability does not require complex graph neural networks: fifteen standard physicochemical descriptors paired with an optimized XGBoost model achieve a 0.928 ROC-AUC and 97.8% recall with full SHAP-derived mechanistic transparency.

Quantum Hilbert space embeddings beat Kernel PCA by 5.5 F1 points on linear credit scoring, yet provide zero performance edge over standard classical tree ensembles.

Black-box simulator data is all a normalizing flow needs to automatically rediscover optimal, nuisance-invariant test statistics like the $t$-test—surpassing standard profile likelihoods in both statistical power and inference speed.

Combinatorial physics discovery spaces that previously required biased human heuristics can now be navigated autonomously, with RL agents outperforming manual intuition to find valid loop-level particle physics operators.

Entangling $t$ quantum copies during measurement improves state tomography efficiency by at most a factor of $\sqrt{t}$, showing that reaching optimal collective sample rates strictly requires entangling $\Omega(r^2)$ samples regardless of classical adaptivity.

Even "leak-free" protein interaction benchmarks are riddled with non-biological shortcuts that models readily exploit—and standard negative-sampling heuristics often make the bias worse.

Across 82 molecular generative models spanning diffusion to Transformers, high in silico benchmark scores still fail to reliably translate to wet-lab therapeutic hits without explicit modeling of receptor flexibility and 3D interaction physics.

Soft physics constraints enforced through automatic differentiation eliminate unphysical density inversions in sparse observational regimes while maintaining—and in some configurations improving—multi-task surrogate accuracy.

Scaling up quantum memory does not inevitably demand more request history: optimal entanglement allocation can expand across larger qubit graphs without increasing sample complexity.

Decoupling multi-center biomedical AI from centralized aggregators, directed trust propagation paired with generative replay reveals coordinated higher-order molecular subnetworks of aging directly across private, fragmented clinical datasets.

Replacing costly aggregate-resolved impact simulations with graph neural operators yields a 4,000x acceleration in predicting dynamic 6-component stress tensors, sidestepping the severe erosion artifacts that compromise standard scalar penetration metrics.

Spacecraft can now flag critical telemetry faults just 30 days post-launch without a single historical anomaly label or manual threshold adjustment.

Extrapolating neural networks backward toward sharp mathematical singularities—historically a fatal failure mode for scientific ML—becomes reliable when physics-guided coordinate transforms are coupled with recursive prediction-extension loops.

Predicting targeted protein degradation on underexplored E3 ligases no longer requires dense labeled datasets: support-conditioned prototype estimation enables zero-fine-tuning transfer across ligases with as few as two samples per class.

Standard self-attention falls short on noisy medical claims: explicitly cross-attending between diagnosis and treatment histories outperforms gradient-boosted trees and established cardiovascular risk models with a 0.772 ROC-AUC.

Generative crystal structure prediction no longer requires thousands of sequential sampling steps: learning average probability flows yields state-of-the-art structural and space-group fidelity in just 1 to 5 evaluations, generating 10,000 candidates in under a minute.

LLM research agents consistently lose to a naive moving-average baseline at forecasting scientific trends—not due to poor reasoning, but because asking models to predict the future systematically biases their search policies away from the most recent literature.

Fragmented satellite data no longer means blind spots: learning over incomplete multi-sensor constellations nearly triples usable methane plume events while slashing false positives by 8.19 points.

Domain-structured bipartite GNNs beat fully digital beamforming under noisy channel conditions while solving severe wideband beam squinting without expensive hardware modifications.

Sampling rollouts via GRPO actively degrades tool-selection training as policies concentrate, whereas exact policy optimization over enumerable combinatorial tool spaces eliminates vanishing gradients and boosts accuracy by up to 14.2 points.

Petabyte-scale brain reconstruction no longer requires massive compute clusters—prioritizing sparse topological skeletons over dense voxel sweeps enables complete neuronal recovery on a single GPU in just one week.

The longstanding performance gap between CNNs and transformers in cell microscopy collapses to under 0.5 macro-F1 once pretraining is controlled, paving the way for distilled tiny students that outperform massive frontier backbones.

Accurate void fraction estimation no longer requires flow-disrupting physical probes: multi-view video coupled with spatio-temporal modeling recovers complex multiphase fluid parameters and directly transfers from synthetic CFD to real experimental flows.

Joint neural estimators went astray on Gaia data by predicting zero dark matter due to an overlooked Jeans' theorem degeneracy—a failure resolved here by decoupling the distribution function and using symbolic regression to recover the closed-form galactic potential.

Polynomial-time sampling in spin glasses and sparse Bayesian regression is tractable deep into low-temperature regimes, slashing the measurement barrier for spike-and-slab posteriors from $k^3$ to $k^{3/2}$ while approaching the Almeida–Thouless phase boundary.

Jointly modeling distinct phenological tasks in a shared multi-task RNN overcomes severe temporal label scarcity, decisively outperforming established domain-specific scientific process models.

Treating natural state trajectories as self-supervised "actions" turns static scientific emulators into interactive, counterfactual world models without requiring manually labeled intervention data.

Calibrated uncertainty quantification in deep scientific surrogates no longer requires expensive outer-loop sweeps: embedding empirical Bayes directly into heteroscedastic neural Gaussian processes delivers calibrated predictions at dramatically reduced compute.

Differentiable cardiovascular physics can filter out ICU sensor artifacts that fool standard deep nets, yielding a 5-point SOTA leap on VT alarm reduction alongside a 2x boost in label efficiency.

Autoregressive graph surrogates fail to generalize over long horizons without explicit level-set geometric encodings, while divergence-free physics regularization proves largely negligible in laminar flow but non-negotiable under turbulence.

Instead of relying on crude, handcrafted perturbations for radiation therapy safety margins, generative latent velocity fields can now simulate realistic, continuous 3D patient anatomical deformations at full clinical CT scale.

Equivariant-Free Transformer-Autoencoded Latent Flow Matching, a two-stage generative framework that relies entirely on a single fixed-dimensional molecule-level latent representation to generate variable-size molecules, is introduced.

Deep ensembles pretrained solely on carbon scattering data can extrapolate to unseen nuclear isotopes and unmeasured kinematic domains through target-dependent, layer-selective fine-tuning.

Shifting industrial process monitoring from brittle bespoke models to a pre-trained foundation model cuts soft-sensing error by up to 14.6% while guaranteeing calibrated uncertainty across seven years of operational drift.

Stiff boundary conditions don't always need full-fidelity simulation: learned error gates determine exactly when a solver can safely drop to cheaper Dirichlet limits without violating numerical tolerances.

Deterministic weather models can match dedicated probabilistic ensembles within 0.04–0.13 CRPSS at a 10-day lead time simply by perturbing raw network weights at inference without any retraining.

A history-aware offline RL policy which predicts iterative cost weights in these dense regimes to improve convergence across placement densities by utilizing readily available features from the router and which reduces design rule violations (DRVs) by an average of 92% over the top public baseline while simultaneously reducing runtime by 10%.

Sheaf theory can unify fragmented temporal data paradigms under a single algebraic framework, making it possible to formally quantify information loss across representation shifts and tractably decompose dynamic systems.

Off-the-shelf LLMs outperform specialized supervised baselines at predicting hypoglycemia across multi-step horizons, but their success hinges entirely on prompt representation rather than feeding the model richer clinical context.

Neural operators can now stably emulate full 3D global ocean and sea-ice dynamics out to 60 days at 0.25° resolution, breaking the compute bottleneck of numerical ocean circulation models.

Standard deep segmentation models shatter when noise swamps signal, but enforcing explicit open-curve geometric topology cuts structural connectivity breaks by over 60% even at -20 dB SNR.

Static textbook circuit problems can now be autonomously transformed into causally grounded, interactive learning environments by coupling multimodal diagram parsing directly with deterministic physics engines.

Histopathology can substitute for missing clinical tabular features in cancer subtyping, but multimodal survival prediction remains acutely fragile to tabular data dropouts and cohort shifts regardless of model architecture.

Natural language instructions can now drive zero-shot 3D organoid phenotyping by fusing Cellpose geometric priors with SAM3 inside an autonomous multi-agent framework.

Field of view matters far more than raw pixel resolution when downsampling biological microscopy imagery, where modern vision transformers decisively outperform traditional quantitative image analysis pipelines.

Energy convergence in higher-order cluster methods confirms the utility of cluster mean field theory for strongly correlated systems, unlocking new computational efficiencies.

These results establish a concrete quantum circuit foundation for simulating polymerization kinetics on fault-tolerant quantum hardware, and a first step toward exploiting exponential state-space compression for high-dimensional polymer reaction networks.

Clinical histories are fundamentally irregular and multimodal, yet foundation models have treated them as static or uniformly discrete—until a generative variational transformer trained on 559 million events unified continuous-time forecasting, multi-modal health records, and counterfactual intervention simulation.

Optimal boomerang uniformity of one in odd characteristic is finally achievable across infinite families, delivering the first general construction of maximally boomerang-resistant APN functions.

Mid-circuit measurements break standard flat quantum compilers; restructuring qubit mapping around hierarchical control flow slashes routing SWAP counts by up to 52% and error rates by up to 40% across monolithic and chiplet QPUs.

Tripling the previous state-of-the-art on full-lifecycle scientific discovery benchmarks, Qiushi Engine satisfies over 82% of end-to-end research rubrics for just $15 per project.

This work outlines a pre-registered prospective test under a shared loss and logging standard, requiring clarified measurements and records, a defined process decision and relevant outputs, credible economic inputs, and validation at the intended scale.

Coupling wet-lab robotics directly to an exascale supercomputer via a unified cloud substrate collapses half a day of manual scientific analysis down to sub-minute interactive queries.

This work extracts the time-invariant topological skeleton of the fluctuating host matrix of the fluctuating host matrix to construct robust, gas-specific physical priors that account for individual molecular dimensions, and decodes the hidden microscopic mechanisms of confined diffusion.

This work proposes BackFlip-2: a fast SE(3)-equivariant graph neural network trained to directly predict dynamical descriptors, such as directional backbone flexibility and pairwise dynamic correlations, from an equilibrium structure, and shows that the proposed equivariant architecture is especially well-suited for capturing anisotropic motions in proteins.

The results suggest that geometric stability and inference-time property guidance should be considered as coupled design objectives in protein-pocket-conditioned 3D molecule generation.

This work demonstrates quantum-accurate treatment of all-atom, complete enzymes in explicit solvent comprising up to 54k atoms and 1 microsecond of total simulation time using the machine-learned interatomic potential (MLIP) eSEN-omol, establishing MLIPs as a practical route to QM-accurate simulations of enzyme catalysis.

Cluster methods can significantly enhance the accuracy of fermionic system modeling, outperforming conventional CAS-SCF techniques in critical applications.

Ring opening—not Hofmann elimination—is the universal kinetic bottleneck driving the degradation of oxidized TEMPO flow-battery catholytes, though functional group tailoring can significantly widen the protective energy barrier.

TSBench is introduced, a benchmark in which an LLM agent uses structure-editing tools to construct three-dimensional transition-state (TS) guesses verified by an automated quantum-chemical pipeline, yielding a physics-grounded pass/fail verdict, establishing a mechanism-level yardstick for LLM agents in mechanism-sensitive tasks such as synthesis planning and autonomous experimentation.

It is shown that physics-informed search can complement benchmark-based evaluation by uncovering hidden failure modes by uncovering hidden failure modes in MLIP Detective, an agentic framework for active failure mode discovery.

Synchronized burst pulses in dolphin communication reveal complex social dynamics that standard analysis methods overlook.

Automating peer review to survive the flood of AI-generated papers creates an exploitable adversarial loop, incentivizing authors to game predictable evaluators and ultimately corrupting the downstream corpus used to train future frontier models.