Max Planck Society

Value-driven optimistic exploration has historically crumbled under deep network approximations, but diagnosing and repairing failure modes in uncertainty propagation allows pure model-free RL to beat complex model-based exploration baselines.

Complex spatial decisions and flocking dynamics naturally fall out of the harmonic modes of neural ring attractors, replacing decades of hand-crafted behavioral heuristics with an exact spectral theory.

Moving beyond passive prompt-and-generate workflows to genuine human-AI collaboration requires solving open technical challenges in mutual agency, real-time latent steering, and dynamic evaluation.

TDA can dramatically streamline the computation of maximum likelihood thresholds in colored Gaussian graphical models, revealing insights previously obscured by traditional methods.

Combining automatic language analysis with clinician judgments can significantly enhance the accuracy of assessing patient experiences in psychiatric interviews.

Bluezz's Reactive Peripheral Modeling uncovers vulnerabilities in BLE firmware that traditional rehosting methods miss, achieving over 2.6 times the basic-block coverage of its predecessors.

LITTLELEARNER reveals that even a well-defined knowledge scope can yield a competent language model, but it won't expand its capabilities beyond its educational boundaries.

Planning success in object-centric world models peaks at high slot quality but plateaus, revealing a nuanced relationship between representation and performance.

AI swarms can now be simulated at scale, revealing how coordinated influence campaigns can manipulate beliefs without detection.

Achieving chemical accuracy with triple-$\zeta$ basis sets in transcorrelated calculations could redefine efficiency benchmarks in computational chemistry.

A clear, open-source implementation of Renormalising Generative Models could revolutionize how we apply active inference in complex environments.

Exploiting a vulnerability in LLM APIs allows attackers to extract proprietary reasoning and sensitive data without directly breaching the more capable models.

LVLMs are misled by superficial cues, with injected signals causing drastic shifts in sarcasm detection accuracy, revealing a fundamental flaw in their reasoning abilities.

Gestures alone can predict referential intent, revealing their critical role in multimodal dialogue even when speech is ambiguous.

ODRA's innovative approach to modeling patient resistance leads to synthetic therapy sessions that are not only more realistic but also preferred by licensed psychologists.

The Push-Forward Transform reveals intrinsic geometric features of shapes while remaining robust to common transformations like rotation and scaling.

Governance mechanisms in DAOs can introduce critical vulnerabilities that are ripe for exploitation, even when implementations are bug-free.

Nearly 1.4 billion state-invariant transactions are clogging Ethereum and Layer-2 blockchains, with speculative MEV driving much of this inefficiency.

Users can now steer avatar interactions with precise control over gaze, head motion, and emotional expression, transforming virtual communication dynamics.

Despite the promise of zero-knowledge proofs, current security tools only catch a fraction of vulnerabilities in real-world applications, exposing critical gaps in their effectiveness.

GNM redefines the landscape of head modeling by integrating intricate anatomical details, achieving superior geometric fidelity that existing models overlook.

Achieving state-of-the-art hair dynamics in head avatars, DynHair allows for realistic and controllable hair movement that adapts to head motion.

Sound probabilistic safety bounds reveal that even state-of-the-art LLMs can be rigorously evaluated for harmful output risks, transforming our approach to model safety.

Rem3Di redefines molecular representation by enabling the differentiation of enantiomers without relying on classical 2D fingerprints, achieving state-of-the-art results in property prediction.

Operators can now dynamically analyze distributed traces across multiple dimensions, enhancing anomaly diagnosis with tailored insights from both visual and natural language interfaces.

LLMs can generate precise structured queries from natural language without any fine-tuning, achieving perfect accuracy in a complex neuroimaging metadata context.

Achieving state-of-the-art accuracy in large-scale quantum chemistry calculations while drastically reducing computational costs could redefine practical limits in many-body physics.

The PanAf-SBR dataset reveals that fine-grained social behaviour recognition in wild great apes can be significantly improved through targeted cross-dataset pre-training.

OT-ICA achieves superior performance in independent component recovery by maximizing the Wasserstein distance, challenging the dominance of traditional proxy-based methods.

Raman scattering can create chiral couplings that amplify coherent light in one direction, a breakthrough for quantum photonics.

MCAO reveals a way to optimize Gaussian basis sets for copper solids, overcoming linear dependence issues while maintaining accuracy in correlated-wavefunction benchmarks.

SLORR achieves substantial model compressibility with under 1% training overhead, outperforming traditional regularization methods in preserving performance.

BRMSD redefines structural comparison by allowing researchers to focus on rigid domains while minimizing the influence of flexible regions, enhancing the accuracy of biomolecular analyses.

Students using text prompts outperformed their peers using voice input, highlighting the need for careful consideration of input modalities in programming education.

A dual-$\boldsymbol k$-mesh strategy transforms BSE calculations, achieving unprecedented accuracy and efficiency in modeling absorption spectra.

Deliberately injecting bugs into GenAI-generated code significantly enhances students' debugging skills and success rates in subsequent attempts.

Novice programmers often overlook crucial details in prompts, leading to a reliance on AI that can hinder their debugging skills and understanding of code generation.

Uncovering novel black hole solutions, this method reveals metrics with trapped interiors that challenge existing paradigms in general relativity.

The Oracle Distance theorem reveals that all noising processes achieve the same optimal negative ELBO, linking diverse loss functions in diffusion models.

Achieving state-of-the-art garment modeling accuracy without the need for physical simulations could revolutionize digital fashion design.

PruneGround cuts through the clutter, achieving state-of-the-art 3D visual grounding by intelligently narrowing the search space based on language cues.

RC-xTC not only surpasses standard xTC in energy accuracy but also outperforms CCSD(T)-F12a in the double-ζ regime, reshaping expectations for small basis set calculations.

Uncovering the hidden architecture of molecular films reveals spatial heterogeneity that could revolutionize the design of functional devices.

A new dataset and transformer model that together nearly double the success rate for 3D hand-object pose estimation in real-world settings.

ReactionAtlas uncovers nearly 47,000 reactions from just eight seed molecules, revolutionizing our understanding of carbohydrate chemistry and its implications for the origins of life.

Stochastic training can transform Looped Transformers from brittle to robust, dramatically reducing OOD variance and improving performance across diverse tasks.

A new attribution method reveals the full spectrum of information flow in ETGNNs, enhancing explainability beyond traditional event-related embeddings.

Sequence probability can signal correctness in LLM outputs, but tweaking decoding methods often fails to enhance accuracy.

SLMs can uncover over 10% of relevant papers in systematic reviews that human reviewers might miss, significantly speeding up the screening process.

Achieving a 9x speedup in gyrokinetic simulations could revolutionize the efficiency of plasma physics research.

Evolving structured concepts with LLMs uncovers new families of quantum error-correcting codes that challenge conventional designs.

Group selection can transform LLM populations from self-interested agents into cooperative entities, revealing a powerful mechanism for promoting prosocial behavior in AI.

Trolls are more effective at polarizing discourse by leveraging moral condemnation rather than promoting foreign narratives, leading to higher engagement rates.

IMAGIN-4D enables unprecedented fine-grained control over human-object interactions by leveraging spatio-temporal image conditioning, outperforming traditional methods that rely on single-token representations.

Action-aligned representations can be achieved without complex training methods, enabling robust planning in dynamic environments.

TD learning can achieve variance bounds comparable to Monte Carlo methods, but with the added advantage of shorter updates leading to more stable estimates.

Expanding the classification space from 35 to 64 classes, DeepForestVisionV2 dramatically enhances the accuracy and utility of camera-trap monitoring in diverse ecological contexts.

The psychological profiles of LLMs are largely illusions created by measurement bias, not genuine traits.

Generative models may not just produce images; they commodify social interactions, embedding ideological biases that reshape visual culture.

Achieving full-stack fidelity in live simulations without sacrificing performance could revolutionize how we evaluate distributed systems before deployment.

d-OPSD enables dLLMs to learn from their own future outputs, drastically improving sample efficiency and performance in reasoning tasks.

The nitrogen site in fluoropyridine reveals a surprising sensitivity to local vibrations that is dramatically enhanced by electronic excitation, while the fluorine site remains largely unaffected.

Fixed-point convergence enables adaptive computation in reasoning tasks, allowing models to efficiently tackle complex challenges without unnecessary resource expenditure.

A novel hybrid framework allows for the recovery of unknown ion channel dynamics directly from voltage recordings, bridging gaps in biophysical neuron modeling.

Forgetting isn't just a loss; it's a chaotic regression that can be vividly visualized through the decay of neural networks.

Rapidly prototype and deploy new rendering methods across diverse hardware platforms with just a few lines of code using XPR.

STRAND revolutionizes the analysis of persistence diagrams by enabling hypothesis testing and vectorization from a unified representation, enhancing both interpretability and performance in machine learning tasks.

Control interventions are often detected by LLMs, with awareness levels varying significantly across models and tasks, revealing vulnerabilities in AI safety protocols.

The first detection of HDO ice in a protoplanetary disk reveals a strikingly high deuterium enrichment, reshaping our understanding of chemical inheritance in planetary formation.

Learning can emerge directly from a system's physical responses without the need for explicit backpropagation or centralized processing.

Elo rankings can accurately reflect model performance, achieving over 90% correlation with ground-truth accuracy, even amidst stylistic biases.

PaperMentor delivers actionable writing feedback that 90.6% of users found useful, setting a new standard for AI-assisted manuscript development.

GraphDETR can detect diverse subgraph patterns in graphs with up to 1000 nodes, achieving an impressive average precision of 91.2% for molecular functional group detection.

Fine-tuning with just 30% of high-fidelity data can achieve quantum-chemical accuracy while slashing computational costs by 60%.

STRIDE reveals that training data influences can be efficiently traced in LLMs using sparse recovery, achieving attribution 13 times faster than traditional methods.

PyFEXScan uncovered over 200 previously unknown malicious packages on PyPI, revealing the vulnerabilities in current analysis tools.

QMT reveals that smartphone video can provide reliable 3D movement analysis, making it a game-changer for monitoring chronic pain in real-world settings.

Speculative tool calls can expose user intent before agents even commit, but new privacy contracts can effectively mitigate this risk.

Low-Pass Flow Matching reduces sampling costs while enhancing image quality by aligning flow matching with the natural frequency decay of data.

Current vision-language models can be surprisingly blind to subtle, context-dependent harms lurking in image-text pairs, but a new reasoning-augmented training framework can help them see the bigger picture.

PARCEL redefines visual tokenization, achieving superior efficiency and performance by dynamically anchoring feature extraction to spatial pool tokens.

Imagine telepresence where your avatar convincingly blends into any environment, relit in real-time based on the scene's actual lighting, all from a single headset.

Safety benchmarks may be measuring a model's knowledge of how evaluations are designed, not genuine safety.

State Space Models can now generate time series with provable universality, outperforming fixed-grid models, especially on irregular data.

PEFT methods aren't just about downstream accuracy; they have distinct "stability-plasticity profiles" that reveal how well they retain general capabilities, and most overshoot the optimal balance anyway.

Uncover the hidden dimensions driving representations in brains, behavior, and AI with a new method that outperforms traditional similarity comparisons.

Unlock olfactory prediction from raw sensor data: SCENT aligns mass spectra with molecular structure, enabling odor prediction without needing explicit chemical formulas.

Video models stumble when the camera angle changes, revealing they're often just memorizing visuals, not grasping physics.

Current AI security benchmarks are fundamentally flawed due to exploitability, staleness, and runtime variability, rendering their results unreliable.

Observed sample displacements can be integrated into optimal transport to carve expressways through the input space, leading to more reliable modeling of distribution shifts.

Condensin's loop extrusion mechanism relies on catch bonds, meaning that, counterintuitively, applying a small amount of force actually *increases* the lifetime of a key intermediate state.

Wi-Fi PIN inference attacks, previously thought to be a major threat, crumble when faced with realistic typing variations, revealing that current performance metrics are misleading.

Granular Mixture-of-Experts can now be efficient: AIR-MoE's two-stage routing slashes routing costs without sacrificing performance.

Deterministic decoding can outperform stochastic self-consistency in constrained domains by systematically exploring high-probability reasoning traces, leading to better performance with less computation.

Stop penalizing your ANN search algorithms for failing to retrieve irrelevant neighbors – Semantic Recall offers a more nuanced and effective way to measure retrieval quality.

Ditch sparse contact cues: LEXIS-Flow uses a learned manifold of interaction signatures to capture dense, continuous proximity between humans and objects, leading to more realistic 3D HOI reconstructions.

Synthetic counseling dialogues can be made significantly more realistic and useful for fine-tuning by grounding them in structured Client Psychological Graphs that capture the interplay of a client's thoughts, emotions, and behaviors.

LVLMs can self-detect and correct object hallucinations by focusing on specific image regions, offering a simple, training-free fix.