RIKEN

Second-order Frank–Wolfe optimization is no longer handcuffed to polytopes: damped Newton subproblems coupled with residual backtracking unlock global linear and local quadratic convergence across general compact convex sets.

Multi-agent collaboration bottlenecks when agent personas stay rigid, but dynamic, test-time strategy transplantation across system prompts enables heterogeneous LLMs to collectively align on the most effective reasoning path.

Disjointed cropping destroys visual context—redistributing a fixed pixel budget through continuous foveation around an MLLM's self-selected focus points beats task-trained baselines without any fine-tuning.

Conventional evaluation methods can mislead researchers, as models that excel in single-output quality may not provide the most useful diverse outputs in practice.

Flow matching can revolutionize PET image reconstruction by achieving better bias-variance trade-offs with fewer sampling steps than traditional methods.

Bridging fragmented research, this survey offers a unified taxonomy that redefines human-centric intelligence in the age of foundation models.

ICBQ not only cuts perplexity in quantized models but also salvages performance where traditional methods falter, redefining efficiency in model compression.

Proprietary MLLMs may achieve high diagnostic accuracy, but they still struggle with reliable clinical reasoning, revealing significant gaps in their practical utility.

ZF2ST achieves strong testing power for detecting distributional differences while maintaining rigorous statistical validity, outperforming traditional methods.

AI-generated feedback may not meet learner needs, as it often misaligns with expert evaluations, revealing a critical gap in language education tools.

Conditioning on informative local measurements transforms a memorization-prone estimator into a high-fidelity quantum state tomography tool, achieving 0.95 fidelity with fewer measurements.

NAIS can autonomously conduct complex biomedical research while maintaining rigorous governance and human oversight, achieving results on par with traditional expert-led studies.

A hybrid quantum-classical approach reveals that a coupling cutoff can dramatically simplify Hamiltonian representation while preserving essential dynamics in reaction-center chemistry.

Entangled quantum circuits can significantly hinder generalization, leading to worse performance than non-entangled circuits with the same number of parameters.

High-CFG diffusion inversion can fail dramatically depending on the prompt-latent pairing, revealing a nuanced landscape of reconstruction success that challenges existing assumptions.

Even Japanese-specific LLMs fail to grasp kanji readings, revealing a significant shortcoming in their linguistic understanding.

Calibration, not compilation, is the key to ensuring statistical accuracy in probabilistic programs generated by language models, with detection rates soaring to 97% when using Bayesian workflows.

Achieving data-dependent regret bounds in policy optimization with unknown transitions could redefine our understanding of adaptive learning in MDPs.

Visualizing counterfactuals can unlock reasoning capabilities in LLMs that text alone cannot achieve.

Uncovering how macroeconomic conditions can shift portfolio strategies reveals critical insights that challenge traditional optimization approaches.

Trajectory-guidance solvers can lead to up to 800 times larger errors compared to simple source reweighting, fundamentally reshaping our understanding of flow-based inverse problem solving.

A variational autoencoder significantly boosts the robustness of neural network-based model reduction for turbulent flow predictions, ensuring high accuracy even in complex vehicle geometries.

Merging past knowledge with fast adaptation in the CoVON optimizer leads to superior performance in continual learning tasks, outperforming traditional methods.

EVON transforms structured weight uncertainty into a practical tool, yielding superior performance in language model pretraining without the overhead of complex implementations.

Ultra-peripheral collisions can be transformed into a powerful tool for nuclear imaging, revealing critical insights into nuclear structure through deep learning.

Data-similarity can reliably guide output tracing, but its accuracy skyrockets when refined with data-influence, revealing a powerful synergy between the two measures.

Adversarial perturbations can be effectively managed in bandit optimization, revealing how budget constraints shape regret outcomes in complex loss landscapes.

Manipulating the effective dimension of quantum kernels can enhance generalization and accuracy in quantum vision models, revealing a surprising benefit of noise injection.

VLMs can only partially resolve structural ambiguity using visual cues, revealing significant gaps in their understanding capabilities.

PuDGhost reveals that interference from neighboring DRAM cells can compromise computation accuracy in Processing-using-DRAM operations by nearly 50%.

Discrete-time IGO can achieve global convergence in continuous spaces, even with fixed learning rates, challenging previous assumptions about optimization dynamics.

Abstaining from uncertain claims is wasteful when additional visual evidence can be efficiently acquired, and our BCEA approach proves it can enhance both reliability and coverage in LVLMs.

Cross-lingual evidence significantly hampers the performance of deep research agents, revealing critical integration challenges that go beyond mere retrieval failures.

Training DNNs on synthetic ECG data can yield up to a 33.2% improvement in classification accuracy for rare cardiac conditions, even with limited real-world samples.

Capped evaluation reveals that many high scores from coding agents are just clever shortcuts, not true problem-solving.

LLMs may completely bypass the neglect-zero effect, challenging assumptions about their alignment with human cognitive biases.

A critical temperature \( T_c \) reveals a phase transition-like behavior in LLM outputs, reshaping our understanding of how temperature scaling affects semantic coherence.

Adding just one spatial word can lead MLLMs to consistently choose the wrong answer, revealing a critical vulnerability in their reasoning processes.

Complexity minimization reveals that as pre-training data scales, few-shot adaptation error rates significantly decrease, challenging existing theories on sample complexity.

Clinically-focused NER for prion diseases is now possible with PrionNER, a new dataset that exposes the limitations of existing models in extracting fine-grained, complex information from biomedical literature.

Materials synthesis just got a reasoning upgrade: a new benchmark reveals the limitations of current methods, while a novel provenance-grounded framework leaps ahead in out-of-distribution performance.

Sub-percent accuracy in emulating the nonlinear matter power spectrum is now achievable across a wide cosmological parameter space, thanks to a novel neural network architecture trained on multi-resolution simulations.

Solving nearest neighbor search just got a whole lot faster by turning it into a ray-tracing problem for GPUs.

You can now accurately track hand movements in a room using only low-resolution cameras placed in the corners, opening the door for unobtrusive activity monitoring.

Despite its simplicity, mean pooling works surprisingly well because modern text encoders concentrate token embeddings, preserving crucial information about their distribution.

LLMs reliably capture emotions with explicit lexical markers, but systematically fail on pragmatically complex emotions requiring contextual inference, revealing a critical limitation in their ability to understand nuanced human emotion.

Turns out, how a drug binds to SARS-CoV-2 RNA depends heavily on both the RNA's shape (threaded vs. unthreaded) and the drug's protonation state.

Uncertainty estimates from LLMs can crumble under distribution shift, but the right probe design – think middle layers and token aggregation – can make them surprisingly resilient.

VLMs already contain a rich latent space of aesthetic features that can be unlocked for personalized image ranking with just a linear readout, no fine-tuning needed.

AI agents, guided by a carefully engineered harness, can build and verify complex software libraries at a scale and pace previously unattainable for tasks like polynomial-time problem reductions.

Identifiability in weakly supervised learning doesn't require irreducibility: conditional independence can save the day.

Forget slow bandits: this new algorithm slashes per-round computation to O(1) while staying robust against adversarial corruption and heavy-tailed noise.

NeuronMoE slashes multilingual LLM parameter counts by 40% without sacrificing performance, by cleverly allocating experts based on neuron-level language specialization rather than blunt layer-level assignments.

Automating software repository build and testing across languages and platforms is now possible, unlocking scalable benchmarking and training for coding agents.

Takeuchi's Information Criterion (TIC) accurately predicts DNN generalization gaps, but only when models operate near the Neural Tangent Kernel (NTK) regime.

MLLMs can slash image annotation costs by 1000x while TagLLM closes the performance gap with human annotations by 60-80%.

Save up to 20% on checkpoint storage in HPC apps by surgically excluding unused data elements, identified via automatic differentiation.

Achieve over 12x better compression of X-ray CT data by focusing only on regions of interest, slashing storage and processing costs without sacrificing critical information.

Lipschitz-constrained Transformers, built from gradient flows, can provably approximate any Lipschitz-continuous function, offering a path to more robust and stable architectures.

OpenMP debugging gets a major speed boost: new distributed recording slashes record-and-replay overhead by 2-5x, finally making it practical for large-scale HPC apps.

Even with noisy human preferences, symmetric losses can guarantee rank-preserving rewards, unlocking robust policy optimization for aligning language models.

ChatChemTS lets chemists design molecules with AI through simple chat, no AI expertise required.

Resolve previously undetectable trace elements with hard x-ray spectroscopy using transition edge sensors, opening new analytical possibilities.