RefusalBench shows strict refusal rates fail to rank frontier LLMs correctly on biological safety, with provider effects and partial-compliance patterns that binary metrics miss.
Canonical reference
Title resolution pending
Canonical reference. 77% of citing Pith papers cite this work as background.
117
Pith papers citing it
5,928
external citations · OpenAlex
Background
77% of classified citations
citation-role summary
background 34
method 4
baseline 1
citation-polarity summary
co-cited works
representative citing papers
A theoretical model derives the universal mammalian vascular branching exponent α* ≈ 2.72 from a network-level minimax principle and topological rigidity theorem grounded in ATP costs, yielding α*_model ≈ 2.626 with heterogeneities shifting it to observed values.
Large language models display the identifiable victim effect at roughly twice the human baseline, strongly amplified by instruction tuning and chain-of-thought prompting but inverted by reasoning-specialized models.
Native spin-changing collisions plus arbitrary quadratic Zeeman shift produce universal OAT-scalable spin squeezing in spin-1 BECs, freezable by turning the shift off.
Compartmentalization of a single morphogen into intra- and extracellular fields with nonlinear coupling produces diffusion-driven instabilities enabling Turing patterns.
In small-budget RCTs where significance tests decide scale-up, optimal pilot sampling shifts from representative to single homogeneous subpopulation as budget shrinks.
An optimization-based inverse design method discovers metainterfaces achieving custom friction laws including power laws with exponents from 2/3 to 1.35 and bilinear forms, with experimental validation for some cases.
SaaSBench introduces a heterogeneous benchmark for enterprise SaaS engineering and shows that state-of-the-art coding agents fail over 95% of the time before reaching deep business logic due to setup and integration problems.
MLLMs display a large perception-reasoning gap on perspective-conditioned spatial reasoning tasks from omnidirectional images, with sharp accuracy drops on advanced tasks like egocentric rotation, though partial gains are possible via RL reward shaping.
A programmable silicon photonic chip excited with single photons implements quantum reservoir computing for quantum state tomography, entanglement measurement via negativity, and classical tasks, with an imperfection mitigation technique that improves accuracy over the classical regime.
AI CFD Scientist autonomously discovers a Spalart-Allmaras runtime correction reducing lower-wall Cf RMSE by 7.89% on the periodic hill at Reh=5600 while using a vision-language gate to detect 14 of 16 silent failures missed by solver checks.
A life-cycle optimization framework for deteriorating infrastructure under hazards is formulated as an MDP with a Kronecker-factored tensor method that reduces computational complexity from exponential to linear while preserving exact dynamic programming solutions.
Spatial nonlocality in Lorentz-dispersive media enables infinite-extent momentum bandgaps in photonic time crystals at arbitrarily low modulation speed and strength.
A new 2D quantum-path interference is observed in HHG driven by highly-bichromatic orthogonal fields, producing monomodal modulations in odd harmonics and bimodal modulations in even harmonics.
Kleiber's law is a signature of dynamic wave-impedance matching yielding the exponent β = dα/(2d+α), with 3/4 enforced in 3D and a parameter-free prediction for the wave-to-viscous transition at small body masses.
AromaGen generates real-time custom aromas from free-form text or visual inputs via multimodal LLM mapping to 12 odorants, matching or exceeding human mixtures after iterative refinement in a 26-person study.
Math Takes Two is a new benchmark that tests whether two agents can emergently invent numerical communication to solve visually grounded extrapolation problems without prior mathematical knowledge.
A large benchmark finds traditional imputation methods for scRNA-seq data generally outperform deep learning ones, but numerical recovery does not reliably improve biological downstream analyses and no method wins across all settings.
PEAC extracts parameters from amplitude collapse in correlated matter-wave interferometers, yielding lower bias than conventional methods for perfectly correlated signals.
Autonomous LLM-agent networks form preferential-attachment structures with type-dependent centrality gaps, including capability-misaligned glass ceilings, predicted by a mean-field ODE and confirmed in 100-agent experiments.
Augmented Sparse Encoding Models show language-cortex voxels are best predicted by general Matryoshka SAE features plus surprisal, enabling interpretable shared and region-specific tuning.
Coupled-channel quantum scattering on an ab initio potential energy surface yields pressure broadening and shift coefficients for pionic, kaonic, and antiprotonic helium transitions, and identifies pionic n=19 states as collisionally unstable.
An LLM-assisted pipeline applied to 4,323 governance records finds that permissionless and corporate AI agent protocols show similar participation inequality and fragmentation but denser thematic alignment in the open setting.
Local 2- and 3-cycles enhance RNN computational capacity for Boolean functions, predicted by structural statistics, while adding interneurons boosts large networks.
citing papers explorer
-
RefusalBench: Why Refusal Rate Misranks Frontier LLMs on Biological Research Prompts
RefusalBench shows strict refusal rates fail to rank frontier LLMs correctly on biological safety, with provider effects and partial-compliance patterns that binary metrics miss.
-
The Incommensurability Principle in Biological Transport
A theoretical model derives the universal mammalian vascular branching exponent α* ≈ 2.72 from a network-level minimax principle and topological rigidity theorem grounded in ATP costs, yielding α*_model ≈ 2.626 with heterogeneities shifting it to observed values.
-
Narrative over Numbers: The Identifiable Victim Effect and its Amplification Under Alignment and Reasoning in Large Language Models
Large language models display the identifiable victim effect at roughly twice the human baseline, strongly amplified by instruction tuning and chain-of-thought prompting but inverted by reasoning-specialized models.
-
Universal spin-squeezing dynamics in spinor condensates
Native spin-changing collisions plus arbitrary quadratic Zeeman shift produce universal OAT-scalable spin squeezing in spin-1 BECs, freezable by turning the shift off.
-
Single-morphogen Turing instability driven by nonlinear intracellular-extracellular coupling
Compartmentalization of a single morphogen into intra- and extracellular fields with nonlinear coupling produces diffusion-driven instabilities enabling Turing patterns.
-
When Representative Samples Produce Worse Outcomes: Scale-up Decisions and Testing in Small-Budget RCTs
In small-budget RCTs where significance tests decide scale-up, optimal pilot sampling shifts from representative to single homogeneous subpopulation as budget shrinks.
-
Automated Discovery of Metainterfaces with Tailored Friction Laws
An optimization-based inverse design method discovers metainterfaces achieving custom friction laws including power laws with exponents from 2/3 to 1.35 and bilinear forms, with experimental validation for some cases.
-
SaaSBench: Exploring the Boundaries of Coding Agents in Long-Horizon Enterprise SaaS Engineering
SaaSBench introduces a heterogeneous benchmark for enterprise SaaS engineering and shows that state-of-the-art coding agents fail over 95% of the time before reaching deep business logic due to setup and integration problems.
-
Beyond Localization: A Comprehensive Diagnosis of Perspective-Conditioned Spatial Reasoning in MLLMs from Omnidirectional Images
MLLMs display a large perception-reasoning gap on perspective-conditioned spatial reasoning tasks from omnidirectional images, with sharp accuracy drops on advanced tasks like egocentric rotation, though partial gains are possible via RL reward shaping.
-
Quantum and classical processing with photonic quantum machine learning
A programmable silicon photonic chip excited with single photons implements quantum reservoir computing for quantum state tomography, entanglement measurement via negativity, and classical tasks, with an imperfection mitigation technique that improves accuracy over the classical regime.
-
AI CFD Scientist: Toward Open-Ended Computational Fluid Dynamics Discovery with Physics-Aware AI Agents
AI CFD Scientist autonomously discovers a Spalart-Allmaras runtime correction reducing lower-wall Cf RMSE by 7.89% on the periodic hill at Reh=5600 while using a vision-language gate to detect 14 of 16 silent failures missed by solver checks.
-
Probabilistic Hazard Analysis Framework with Stochastic Optimal Control for Deteriorating Civil Infrastructure Systems
A life-cycle optimization framework for deteriorating infrastructure under hazards is formulated as an MDP with a Kronecker-factored tensor method that reduces computational complexity from exponential to linear while preserving exact dynamic programming solutions.
-
Nonlocal photonic time crystals: Infinite momentum bandgaps with minimal modulation speed and strength
Spatial nonlocality in Lorentz-dispersive media enables infinite-extent momentum bandgaps in photonic time crystals at arbitrarily low modulation speed and strength.
-
2D quantum-path interference in high-harmonic generation driven by highly-bichromatic fields
A new 2D quantum-path interference is observed in HHG driven by highly-bichromatic orthogonal fields, producing monomodal modulations in odd harmonics and bimodal modulations in even harmonics.
-
The Dynamic Origin of Kleiber's Law
Kleiber's law is a signature of dynamic wave-impedance matching yielding the exponent β = dα/(2d+α), with 3/4 enforced in 3D and a parameter-free prediction for the wave-to-viscous transition at small body masses.
-
AromaGen: Interactive Generation of Rich Olfactory Experiences with Multimodal Language Models
AromaGen generates real-time custom aromas from free-form text or visual inputs via multimodal LLM mapping to 12 odorants, matching or exceeding human mixtures after iterative refinement in a 26-person study.
-
Math Takes Two: A test for emergent mathematical reasoning in communication
Math Takes Two is a new benchmark that tests whether two agents can emergently invent numerical communication to solve visually grounded extrapolation problems without prior mathematical knowledge.
-
A Large-Scale Comparative Analysis of Imputation Methods for Single-Cell RNA Sequencing Data
A large benchmark finds traditional imputation methods for scRNA-seq data generally outperform deep learning ones, but numerical recovery does not reliably improve biological downstream analyses and no method wins across all settings.
-
Parameter Estimation from Amplitude Collapse in Correlated Matter-Wave Interference
PEAC extracts parameters from amplitude collapse in correlated matter-wave interferometers, yielding lower bias than conventional methods for perfectly correlated signals.
-
Emergence of Preferential Attachment and Glass-Ceiling Effects in Autonomous Networks of LLMs
Autonomous LLM-agent networks form preferential-attachment structures with type-dependent centrality gaps, including capability-misaligned glass ceilings, predicted by a mean-field ODE and confirmed in 100-agent experiments.
-
Interpreting Brain Responses to Language with Sparse Features from Language Models
Augmented Sparse Encoding Models show language-cortex voxels are best predicted by general Matryoshka SAE features plus surprisal, enabling interpretable shared and region-specific tuning.
-
Density effects in precision laser spectroscopy of exotic helium atoms
Coupled-channel quantum scattering on an ab initio potential energy surface yields pressure broadening and shift coefficients for pionic, kaonic, and antiprotonic helium transitions, and identifies pionic n=19 states as collisionally unstable.
-
Agentic Analysis for Agentic Infrastructure: An LLM-Powered Pipeline for Comparative Governance of DAO and Corporate AI Protocols
An LLM-assisted pipeline applied to 4,323 governance records finds that permissionless and corporate AI agent protocols show similar participation inequality and fragmentation but denser thematic alignment in the open setting.
-
Identifying structural design principles shaping the computational abilities of recurrent neural networks
Local 2- and 3-cycles enhance RNN computational capacity for Boolean functions, predicted by structural statistics, while adding interneurons boosts large networks.
-
ConnectomeBench2: A Unified Benchmark for Automated Connectomic Proofreading
ConnectomeBench2 supplies a unified multi-species benchmark of expert proofreading labels and shows a single Vision Transformer achieving human-level performance on split and merge error tasks while providing calibration and distribution-shift diagnostics.
-
What Is a Pattern in Statistical Mechanics? Formalizing Structure and Patterns in One-Dimensional Spin Lattice Models with Computational Mechanics
Derives Boltzmann distributions for 1D spin models as stochastic processes, quantifies structure via excess entropy and statistical complexity, specifies mechanisms with epsilon-machines, and reports agreement with typical configurations.
-
Projected Inverse Iteration: An Eigenvalue Approach to Ground-State Computation with Neural Quantum States
Projected Inverse Iteration reframes ground-state search for neural quantum states as an eigenvalue problem to deliver rapid, spectral-gap-insensitive convergence while retaining polynomial scaling.
-
A Simulation Platform for Flapping-Wing Vehicles
FWAV-Sim is a high-fidelity Unity simulation framework for flapping-wing vehicles that integrates blade-element aerodynamics with bluff-body drag, spatiotemporally correlated fractal turbulence, and realistic IMU/LiDAR/RGB sensor models to support autonomy development.
-
Parallelizing Large-Scale Tensor Network Contraction on Multiple GPUs
A communication-aware multi-GPU distribution approach for tensor network contraction reports 7-173x extra speedup over slicing on 8 H100 GPUs and 42x to 67,869x on 1024 GPUs.
-
Ultrafast formation of a large dynamic magnetic soliton
Observation of ultrafast large dynamic magnetic soliton formation inside the linear spin-wave band in garnet films, extending tens of microns and collapsing into short-wavelength spin waves at large distances.
-
Escaping the Mode Lottery: Multi-Response Training Improves Language Model Generalization
Multi-response training retains multiple responses per prompt to reduce uncertainty about the conditional output distribution, yielding improved distributional generalization especially in high response-diversity and low prompt-redundancy regimes.
-
Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning
Faithful chain-of-thought routes answer-relevant information through the CoT path, measured via sufficiency, completeness and necessity with entropy, masked-KL and gradient diagnostics, and improved by information-flow interventions during verifier-based RL.
-
Structure-Aware Masking for Protein Representation Learning
Bucket Masking improves protein fitness prediction by up to 14% over random masking by preferentially masking structurally coupled residue groups on four downstream tasks.
-
Jobs' AI Exposure Should Be Measured from Evidence, Not Model Priors
The authors propose a retrieval-augmented framework that grounds AI exposure labels for 18,796 O*NET occupation-task pairs in retrieved news and academic abstracts, outperforming zero-shot prompting in 72% of disagreements and aligning better with observed real-world usage.
-
Spectral Vision Transformer for Efficient Tokenization with Limited Data
A spectral vision transformer achieves equitable or superior performance with fewer parameters than standard ViTs, CNNs, and other models by using spectral projections for tokenization in limited-data medical imaging.
-
Mechanics of heterogeneous fiber networks
Active microtubule motors broaden pore-size distributions, thicken bundles, raise mean and variable local elastic moduli, and extend displacement-field range in actin-fascin networks compared with passive controls.
-
Non-equilibrium scaling across first-order transitions with relativistic scalar fields
Fast driving across first-order transitions in relativistic scalar fields produces temperature- and dimension-independent finite-time scaling matching mean-field theory, crossing over to Kibble-Zurek scaling near criticality and nucleation-dominated dynamics at low temperatures.
-
Continuum Robot Modeling with Action Conditioned Flow Matching
A conditional point-cloud flow matching model maps motor actuation to 3D geometry of tendon-driven continuum robots and outperforms prior self-modeling methods on simulated and real 2- and 3-module hardware.
-
Electrical Spin Pumping in Exchange-coupled Molecules
Exchange-coupled Fe-FePc molecular pairs enable remote all-electrical spin initialization through spin-polarized tunneling current and angular momentum transfer.
-
Retrieval and competition: how a protein foundation model starts a protein
ESM2 predicts N-terminal methionine via retrieval of a positional prior from the BOS token through distributed attention circuits rather than direct recognition, revealed by a norm-direction decomposition of rotary attention scores.
-
Persistent Homology of Time Series through Complex Networks
A standardized pipeline converts time series to graphs, computes persistence diagrams, and extracts features that classify UCR benchmarks, with diffusion distance outperforming shortest-path metrics and performance varying by graph type.
-
Fidelity-Guaranteed Entanglement Routing with Distributed Purification Planning
Q-GUARD achieves over 85% qualified success rate on 4-hop paths in 100-node simulations by allocating per-hop fidelity targets via Werner-state equal-split and selecting paths with a segment-local expected-goodput metric.
-
Enhancing molecular dynamics with equivariant machine-learned densities
DenSNet learns the Hohenberg-Kohn map to electron density with equivariant networks and delta-learning, then maps density to energy, producing stable MD trajectories whose infrared spectra match experiment and DFT on ethanol, ethanethiol, resorcinol, and polythiophene oligomers.
-
Beyond the Critical Depth: The Metabolic and Physical Drivers of Phytoplankton Persistence in a Changing Ocean
A non-autonomous dynamical systems framework shows metabolic constraints increasingly govern phytoplankton persistence in a warming ocean, producing a global expansion of metabolic-driven regimes and a 1:4 ratio of new viable niches to ice-free polar deserts.
-
Beyond Silicon: Materials, Mechanisms, and Methods for Physical Neural Computing
Physical neural substrates realize inference and adaptation via native physics and occupy complementary regimes; no single platform dominates the proposed static/dynamic benchmarks.
-
Database Reordering for Compact Grover Oracles with ESOP Minimization
Reordering database addresses before ESOP minimization produces smaller QROM circuits for Grover oracles, with simulated annealing yielding roughly 30% size reduction over fixed ordering for small N.
-
A solid-state quantum memory based on a continuous optoacoustic system
A continuous optoacoustic system in a Brillouin waveguide enables high-fidelity storage and on-demand retrieval of squeezed and entangled optical states with bandwidths reaching hundreds of MHz.
-
Quantum-Coherent Regime of Programmable Dipolar Spin Ice
Programmable dipolar spin ice realized on a qubit annealer exhibits super-diffusive monopole transport consistent with coherent propagation in an emergent gauge manifold.
-
Shining light on short-range atomic ordering in semiconductors alloys
Short-range order in GeSn alloys directly influences bandgap and can be tuned by annealing, as measured by a new ML-enabled EXAFS analysis linked to photoluminescence data.
-
MAPLE: Self-Supervised Learning-Enhanced Nonlinear Dimensionality Reduction for Visual Analysis
MAPLE enhances UMAP via self-supervised MMCRs to untangle complex manifolds, yielding clearer clusters and finer subclusters than standard UMAP at similar cost.